Open-source I tried · updated Oct 20, 2026
Every open-source repo I tried, and what I kept
The short answer
I've tried 115 open-source GitHub repos while building my AI setup, and 34 of them are still in it. The rest didn't fit how I work, didn't work on my machines, or got replaced by something I built.
Each card says what I wanted it for, what was good, what wasn't, and why I kept it or dropped it, in my own words where I said something at the time. A few I only read the code of before deciding; those cards say so.
My stack: what I pay for and use Watch: my AI setup in 7 minutes
- 115open-source repos I tried
- 34still in my setup
- 58dropped or didn't fit
- 66with what I said at the time
01
Coding agents and where I run them
Claude Code does most of my work. These are the open-source agents and agent apps I ran next to it, and the ones I left.
-
Still using
An open-source app for running many Claude Code and Codex threads, which other computers and phones can connect to.
I wanted it for Running agents on my Linux PC and steering them from my MacBook and iPhone, so work keeps going when the laptop is closed.
Good
- A comfortable UI: settle a thread when it's done, switch with Cmd+1/2/3, one thing at a time
- Agents keep working on the PC after I close the MacBook
- The phone app lets me check finished threads and hand out more work on a break
Not good
- First setup is hard, and with only one computer I'd see no point
- Attachments are images only, so phone videos needed a separate relay
- The Oct 17 Nightly build kept crashing the sidebar (a known upstream bug); Cmd+R brought it back
Verdict It's where I run every agent now. It only makes sense if you also have an always-on box for it to drive.
“What I did right this time was connecting the Linux computer with T3 Code. I had been running Orca on my M2 Pro Mac and it got too hot, so moving was a good call. The T3 Code UI is definitely comfortable, and Linux is easier than the Mac because I can change it however I want.”
-
Still using
A self-hosted app and server that runs teams of AI bots on schedules, using the Claude, Codex or Gemini plans you already pay for.
I wanted it for Research bots that run around the clock on Reddit, X, YouTube, Instagram, TikTok, Threads and Hacker News, on subscriptions I already have.
Good
- Timers, routines, triggers, and a “chief of staff” bot that hands work to the rest of its team
- Each bot can use a different model, so cheaper bots collect and a Claude bot writes the report
- Free and Apache-2.0; a review of the installed package found no backdoor
Not good
- A fast-moving one-person project: one version pinned the server CPU at 100% until the next release
- Its fallback model didn't catch Claude quota-limit failures, so I switched four bots by hand
- Results were poor at first: on Oct 15 only 7.7% of the links across 57 runs were new, and each bot needed tuning
Verdict It replaced my Hermes bots on Oct 14 and runs 31 bots in 4 teams as of Oct 20. If development ever stops, I'll fork it.
“I think OpenMausBot is better? It's more customizable and it can use the subscriptions I already have. Grok Bot's only advantage seems to be X search, right?”
-
Still using
OpenAI's open-source coding agent for the terminal; the same engine runs the Codex app and
codex execfor scripts.I wanted it for A second agent next to Claude: checking Claude's work, browser and UI automation, thumbnails with its image tool, unattended checks in scheduled jobs.
Good
- A useful second opinion on what Claude made
- Good built-in image generation for YouTube thumbnails
codex execworks in unattended scripts and can return structured JSON
Not good
- GPT-6.1 feels much slower than Claude Opus 5.5, and weaker, to me
codex execwaits forever for input unless stdin is closed- It once ran rm -rf and wiped files, which is why I went looking for a delete guard
Verdict I keep the $200 plan for checks, automation and thumbnails. If OpenAI's next model doesn't change my mind, I'll swap it for a second Claude plan.
“GPT is so slow and worse than Claude that I'm deciding whether to go with two Claude Max plans instead. I'll probably decide when the new model drops soon.”
-
Kept for later
An open-source terminal coding agent that runs many different models, including cheap ones through its OpenCode Go plan.
I wanted it for Cheap second opinions from non-GPT models, and low-cost worker agents next to Codex.
Good
- One CLI with many cheap models: DeepSeek Flash for cheap work, Muse Spark 1.3 for a second opinion, Kimi K3 for frontend design
- Easy to call from other agents with one
opencode run -mcommand - 45 sessions cost me $3.68 in total
Not good
- Running it as many parallel workers used up the OpenCode Go allowance within hours
- Some free models refused to run until I accepted a data-collection opt-in
- Finished sessions left processes running for days
Verdict Good value while my agents ran in Orca. Since moving to T3 Code with Claude and Codex plans, I rarely open it.
Now Claude Code and Codex, from T3 Code
-
Kept for later
An Obsidian plugin that embeds Claude Code or Codex as an agent inside the vault.
I wanted it for A smarter in-vault agent for summarizing clipped articles and transcripts.
Good
- Connects to Codex directly, no adapter
- Answers link back to the source note
Not good
- Installed with full auto-approve permissions by default
Verdict It worked, but my logs show only two uses, the last on Sep 1.
-
Good, not for me
A free peer-to-peer messenger that lets coding agents on different computers send each other work.
I wanted it for Seeing whether asking agents across my Mac and PC through it beats what I do now.
Good
- Much faster for quick agent-to-agent questions: 1.5–2.6 s vs 8–10 s over ssh
- Sees Claude Code, Codex, Gemini CLI, Cursor, OpenCode and Pi sessions in one place
- Clean uninstall: my configs matched the backups afterwards
Not good
- Setup rewrites every agent's config and runs a daemon on each machine
- Background bot sessions became visible on the network
- Its resident Claude ran with hooks off, so my delete guard didn't apply
Verdict The one thing it does faster is a path I almost never use, and it touched every agent's settings. I removed it the same day.
“Let's just try it as an experiment. Test whether it's easier than the way I work now.”
-
Good, not for me
A local search engine and MCP server for AI agents that indexes code, docs and logs in one binary.
I wanted it for Seeing whether it could help organize everything my agents touch.
Good
- Search worked, Korean queries included
- Exact go-to-definition for code
- Runs locally, with a checksummed installer
Not good
- It's a search index, not an organizer: it doesn't move, tag or sort anything
- Its setup tells your agent to open public “field report” pull requests from your GitHub account without asking
- About 2.85 GB of RAM for about 37 MB of files
Verdict Search I already had. Nothing was sent, and I had it deleted after the test.
“ok delete it”
-
Dropped
A Mac app for running many coding agents (Claude Code, Codex, OpenCode and others) side by side in separate worktrees, with a phone app.
I wanted it for Running lots of agents across about ten projects on my MacBook.
Good
- Many agents and projects at once, each in its own worktree
- Works with several agent CLIs
- Can run agents on another computer over SSH
Not good
- My MacBook got too hot running that many agents, even with my own fan app
- Work stops when the Mac is closed
- Finished agent processes piled up (9 of them, about 4.4 GB, after 5+ days), and it kept reinstalling its hooks on my PC
Verdict It ran 1,804 agents in about three and a half weeks. Then on Oct 6 I moved all agent work to an always-on Linux PC driven from T3 Code.
Now T3 Code
-
Dropped
Nous Research's open-source agent with long-term memory, skills, scheduled jobs, named bot teams and computer use.
I wanted it for A personal research agent first, then a team of research bots plus a “conductor” on the Mac driving my other AI apps.
Good
- Memory and skills are plain local Markdown files
- Ran on my existing ChatGPT subscription
- Handled a 5-hour unattended research loop
Not good
- Setup is a lot of work: bots, schedules, allow lists, per-app permissions
- The bots sat idle until scheduled jobs were added by hand
- Results got lost between bots; one bot's report never reached the conductor
Verdict The bots didn't do the research properly and setting them up was a chore. A focused OpenMausBot run gave a more useful answer in about 9 minutes, and I deleted all 10 Hermes bots that day.
“The Hermes bots don't seem to do the work well. I installed OpenMausBot, so let's try it there.”
Now OpenMausBot
-
Dropped
A small terminal coding agent you extend yourself with extensions, skills and packages (formerly pi-mono).
I wanted it for A personal coding harness on my ChatGPT plan: plan, run sub-agents, test, commit, with fewer prompts from me.
Good
- Very customizable: extensions, skills, prompt templates, packages
- Reused the skills I'd already written for Codex
- Ran on my ChatGPT subscription, no API key
Not good
- A lot of hand setup, and no automatic context gathering
- Third-party packages (pi-subagents, pi-intercom) misbehaved and came out
- On the Linux PC the subscription sign-in never got finished
Verdict It was my daily driver on the Mac for about three months, about 1,490 prompts. Once I ran Claude Code and Codex directly in T3 Code, I stopped opening it.
“Pi needs something like auto context.”
Now Claude Code and Codex, from T3 Code
-
Dropped
Google's open-source agent that runs Gemini models in the terminal.
I wanted it for Running Gemini from the terminal next to Codex and Claude.
Not good
- Google stopped serving it to my accounts on Jun 18, 2026
- The copy on my PC had no working sign-in when an agent tried it on Oct 9
Verdict Google moved consumer accounts to Antigravity CLI, so that's what my agents use for Gemini work, mostly video reviews.
“Gemini CLI already shut down, didn't it? Check, and if I'm right, use Antigravity CLI from now on.”
Now Antigravity CLI
-
Dropped
A macOS menu bar app that shows how much of your Codex, Claude Code and other AI plan limits you've used.
I wanted it for Watching my AI plan limits from the menu bar.
Good
- Several providers' limits in one menu bar item
Not good
- A hidden debug setting silently changed how it signed in and gave a misleading “cookie header is invalid” error
- In the end it's only a token counter
Verdict After that bug I lost patience with it. I read my weekly plan meters with my own script now.
“Ugh, this is really badly made software.”
Now My own script that reads the Claude, Codex and Antigravity weekly meters (see them)
-
Dropped
CopilotKit's open-source personal agent with its own browser, terminal and files, run as a web app.
I wanted it for A head-to-head video against OpenMausBot on the same real task.
Good
- Its test suite passed 147/147
- The official demo ran in an isolated container with a real browser
Not good
- Needs a CopilotKit cloud project key even for the demo
- Real models need paid API keys; no subscription login
- The demo uses a scripted mock model, so it shows nothing about real performance
Verdict I run agents on the subscriptions I already pay for, not on API keys. The comparison video was dropped the same day.
“Can I use a subscription instead of the API?”
Now OpenMausBot
-
Dropped
The Python library for TypeSafe's API, where a model answers typed multiple-choice questions with probabilities.
I wanted it for Speeding up the classification step of a viral-Shorts research pipeline.
Good
- Typed questions with fixed choices made the pipeline easy to test against an answer key
- The model behind it could be swapped with one option
Not good
- The Jev model it's built for was waitlist-only, so I never actually used it
- Text only, while my Shorts task did better with images (85.5% vs 82.8%)
Verdict It ended up as a format layer over Gemini, and the pipeline was switched off on Oct 7 when the work moved to the Linux PC.
-
Dropped
xAI's open-source terminal coding agent.
I wanted it for Trying Grok as a coding agent on my Mac.
Good
- Open source, so its privacy settings could be checked in the code
Not good
- Earlier versions uploaded whole repositories
- Prompts and the files it reads still go to xAI's cloud
Verdict Privacy worries. xAI had stopped the whole-repo uploads by then, but I removed it the same day anyway.
“is grok build fixed with the repo uploading stuff? or do I need to disable it?”
-
Dropped
An Obsidian plugin that runs coding agents, Codex included, in a side panel of the vault.
I wanted it for Asking Codex about my notes and clipped transcripts inside Obsidian.
Good
- Handled simple note jobs, like translating a transcript with 154 timestamps intact
Not good
- Codex runs through an adapter there, and it felt dumber
Verdict The agent felt too dumb through this harness, so I swapped it the same day.
“The agent I use in Obsidian seems too dumb. Is there a better alternative (Obsidian plugin)? The harness is really bad.”
Now Claudian
02
Memory for coding agents
I wanted agents to remember across projects, delete what's stale and archive what matters. I read the code of five memory tools, then built a plain folder of notes instead.
-
Kept for later
An MCP server and hooks that keep huge tool outputs out of a coding agent's context, plus a session memory.
I wanted it for Checking whether its claim of cutting tool output by 98% is real, to save tokens.
Good
- A made-up 10,000-line log (about 1.6 MB) came back to the model as 312 bytes, with all 40 errors counted right
- Search over stored output worked, Korean included
- Data survived a server restart
Not good
- The 98% figure compares output sizes, not total tokens or cost
- Tiny outputs can get bigger (34 to 175 bytes)
- It mixed up two projects' memories, an open bug (#1214) my test reproduced
Verdict The saving is real, but I work across many projects and its memory mixed them up. I'll retest once that bug is fixed.
“OK, then let's use it once that's fixed.”
-
Good, not for me
A plugin that records everything a coding agent does, compresses it with an LLM and feeds it back into new sessions.
I wanted it for A knowledge system: capture what agents learn, delete what's stale, archive what matters, find it later.
Good
- Captures automatically from both Claude Code and Codex
- Search across past sessions
Not good
- Never prunes or archives; one user reported a 19.4 GB database
- Stored in SQLite, so I can't read or edit it like notes
- One LLM call per tool call, with issues reporting about $183 a day of spend
Verdict It captures, but it doesn't organize or delete, which was the whole point for me.
“What I really want is a knowledge management system: delete what I don't need as I work, archive what matters, keep it organized, and pull out whatever I need anytime with an LLM. That's not easy.”
Now A shared folder of plain markdown notes with one index, read by every Claude Code and Codex session
-
Good, not for me
A memory layer for coding agents that saves each session to a per-repo memory and feeds relevant memories back in.
I wanted it for Giving Claude Code a memory per project, so I stop repeating decisions.
Good
- Can run fully local
- MIT license and very active
Not good
- An LLM call on every turn, with issues reporting cost blowups of about $150 an hour
- Grouped by git repo, while much of my work happens in scratch folders
- I found one hands-on review; recall took about 13 seconds
Verdict I only wanted it if I clearly needed it, and the reviews showed cost risk and little real use.
“Before installing, find reviews from people who have used it. No point installing it if I don't really need it.”
Now The same shared folder of markdown notes
-
Good, not for me
An MCP server that lets Claude Code and Codex read and write a folder of markdown notes.
I wanted it for Shared notes both of my agents can read and write.
Good
- Works with both Claude Code and Codex
- Markdown notes, and capture costs no LLM calls
Not good
- Only captures when a session compacts
- Korean keyword search was broken (#1294)
- Telemetry and a silent auto-update every 24 hours are on by default
Verdict The closest fit of the five, but plain files did the same job without another tool running.
Now The same shared folder of markdown notes
-
Good, not for me
A Claude Code skill and plugin that keeps an agent's memory as markdown in an Obsidian vault.
I wanted it for One global vault my agents could fill and tidy.
Good
- Plain markdown in one vault that works from any folder
- A real test suite and outside contributors
Not good
- Saving is a manual command
- The “agents tidy your vault while you sleep” feature was prompt text with no scheduler code
- Codex only gets the commands, no hooks
Verdict The automatic cleanup I wanted didn't exist in the code.
-
Good, not for me
An Obsidian vault template with hooks and an MCP server that gives coding agents a markdown memory.
I wanted it for The same knowledge-system goal.
Good
- Plain markdown in git
- Rule-based checks for stale notes, with no LLM cost
Not good
- Capture is manual or up to the agent
- Only fully works when the agent starts inside the vault
- Built around office work: 1:1s, brag docs, reviews
Verdict It's shaped for performance reviews at a job, not a solo creator working across many folders.
-
Good, not for me
A Claude Code plugin that turns sources you drop in into a linked Obsidian wiki.
I wanted it for The same goal, with an LLM writing a wiki I read in Obsidian.
Good
- Careful: every write goes through a plan and an approval
- No background processes, no background cost
Not good
- Nothing is captured from agent sessions automatically
- Essentially one developer, with a startup hook broken for months
Verdict Approving every step is the opposite of what I want. I want this to run with as little of me as possible.
03
Skills, plugins and code context
Most of these date from my Mac months, building apps with Codex. Every skill pack costs tokens at the start of each session, and that decided a lot of them.
-
Still using
An agent skill with writing rules that strip common signs of AI-written text without changing what it says.
I wanted it for Making my outreach messages to founders read like a person wrote them.
Good
- Just a text file of rules: nothing to run, easy to check
- Drops into any agent workflow as one step
- MIT license
Not good
- No measured effect on replies yet
Verdict After the first 5 questions went out, I made it a required step before any message to a real person.
“ok next time use this https://github.com/blader/humanizer”
-
Still using
A local tool and Mac menu bar app that tracks token use and estimated cost across AI coding agents.
I wanted it for Seeing how much of my AI subscriptions I actually use.
Good
- Installed and showing numbers within one short session
- Several agents in one view
- Runs locally
Not good
- The dollar figure (about $1,962 for October) is an API-price estimate, not what I pay
- Claude's quota needed a fresh login before it showed
Verdict It answered the question and it's still in my menu bar.
“install this for me I want to check my usage”
-
Kept for later
A set of agent skills for planning and code quality: grill-me, grill-with-docs, improve-codebase-architecture, zoom-out and more.
I wanted it for Making the agent interview me before building a feature, and finding architecture problems in my own apps.
Good
- grill-me became my usual way to start a feature or a plan (used 40 times)
- improve-codebase-architecture returned short, ranked refactor lists
Verdict The skill pack I used most on the Mac, 68 times between May and June. It isn't set up on the Linux PC that does the work now.
-
Kept for later
Indexes a codebase into a local symbol and call graph and serves it to coding agents over MCP and a CLI.
I wanted it for Cutting the tokens agents spend gathering code context, and finding dead code.
Good
- Caller, callee and impact queries helped scope refactors
- Overnight refactor agents used it to keep changes small
Not good
- The index belongs to the main checkout, so worktrees and subagents often couldn't use it
- Its dead-code flags were mostly wrong: 2 real out of 18 high-risk files
Verdict My default code-context tool on the Mac from May to July, and still configured there. Not on the Linux PC.
“I think using this could cut token usage. Install it.”
-
Kept for later
An MCP server that indexes a codebase into a knowledge graph for agents.
I wanted it for A second code-graph source next to CodeGraph.
Good
- An architecture overview with clusters in one call
Not good
- Needs a manual index per repo
- Its tools went missing in some Codex sessions because of a hyphen in the server name
Verdict Tried for about two weeks in July. Still registered on the Mac, unused since.
-
Kept for later
A CLI that runs a coding agent in a loop on one goal, overnight for example, committing each pass to its own branch.
I wanted it for Overnight “reduce complexity without changing behavior” runs on my own apps.
Good
- Unattended overnight refactors that passed tests and review; one merge passed 1,233 tests
- Works with Codex as well as Claude
Not good
- The default agent was Claude, so my first run stalled until I switched it to Codex
- Its run files broke my repo-wide format check until cleaned up
Verdict Good enough to merge, but I haven't run it since Jul 16, and it isn't on the Linux PC.
-
Kept for later
A CLI that scores a React codebase and lists dead code, React rule issues, performance and accessibility findings.
I wanted it for Finding React problems and dead code in my book-layout app.
Good
- A clear score with grouped findings: 77, then 90 out of 100 after fixes
- Good at surfacing unused exports and types
Not good
- Its skill ran after every change until I restricted it
- Some rules flag code that is sequential on purpose
Verdict Useful when I ask for it, not as a rule that runs on every change.
“Where in agent.md or anywhere does it keep saying to use react-doctor? I want it to stop now.”
-
Kept for later
An MCP server that gives coding agents current library and framework docs on request.
I wanted it for Making the agent check current docs instead of guessing.
Good
- An easy way to make the agent read current docs before changing code
Verdict I asked for it by name about ten times between Sept 2025 and Apr 2026. Still set up on the Mac; not since.
“use context7 to check if my tailwindcss, eslint config is correct and modern”
-
Kept for later
An agent skill that draws architecture and data-flow diagrams as standalone HTML pages.
I wanted it for Seeing the structure of one of my video apps.
Good
- A readable, interactive diagram of my real code within minutes
Not good
- The viewer interface is English only
Verdict Worked on the first try. Used once so far.
“Not the sample, my app.”
-
Kept for later
An SEO audit skill pack for Codex agents: technical SEO, schema, content, AI-search readiness.
I wanted it for Improving search for my Torah Spark website.
Good
- The audit produced a concrete fix list that shipped the same day
Not good
- 51 reported “canonical mismatches” were favicon links
- The full feature set needs a skill install, Playwright and API keys
Verdict Useful once, but when I pushed, my agent admitted it had only run part of it.
“Are you sure you actually used this properly?”
-
Good, not for me
A code search tool for agents (CLI and MCP) that finds files for a natural-language or symbol query.
I wanted it for Replacing Graphify for code search.
Good
- Runs with no install through uvx
- A symbol query put the right file first
Not good
- Generated type files cluttered the results
Verdict It worked, but I picked CodeGraph the same evening because it also gives call graphs.
Now CodeGraph
-
Didn't work for me
A CLI that bundles a prompt and selected files and sends them to a stronger model for a second opinion.
I wanted it for Getting GPT Pro in ChatGPT to review plans and bugs with the right files attached.
Good
- Clipboard mode packed the right files into one paste
Not good
- Browser mode couldn't use my signed-in ChatGPT session
- Doesn't work with the ChatGPT Mac app
Verdict I wanted reviews to go to ChatGPT on their own, and that never worked. I removed it twice.
“I don't use birdclaw or oracle anymore, so delete them.”
-
Dropped
A skills plugin for coding agents: brainstorming, writing plans, executing plans, systematic debugging, code review.
I wanted it for A plan-then-build loop for my Mac app projects.
Good
- writing-plans plus executing-plans gave me a repeatable plan-to-code loop
- systematic-debugging earned its keep on real bugs
Not good
- Its skill list adds a fixed token cost to every session
Verdict 94 runs in seven weeks, then nothing after Jun 19. I didn't write down why; my planning moved to RepoPrompt CE around then.
-
Every's plugin with plan, work, review and commit skills plus a “compound” step that records lessons so agents stop repeating mistakes.
I wanted it for A lessons-learned step at the end of each review.
Good
- The lessons-learned idea was exactly what I wanted
Not good
- Dozens of ce-* skills loaded into every session before any work
Verdict A fresh session had already used about 26k tokens before doing anything, and its skills were the biggest share. It was the first plugin I switched off.
“Let's turn off just Compound Engineering.”
-
Dropped
A plan-and-track workflow plugin for coding agents with a task CLI and RepoPrompt commands (formerly gmickel-claude-marketplace).
I wanted it for Task tracking and planning in a finance app I built on Rails.
Not good
- Its rules stayed in my agent instructions after I stopped using it, and agents kept trying to run its commands
Verdict I'd stopped using it by July and took it out of the project.
“What is Flow Next? I don't think I use this anymore.”
-
Dropped
The open-source community edition of RepoPrompt: a Mac app and MCP server that builds code context, plans with an “Oracle” model and drives Codex agents.
I wanted it for Running Codex agents with its context builder, deep plans and second-opinion reviews.
Good
- Oracle reviews caught blocking bugs before merges
- Deep-plan and build commands made long multi-agent tasks repeatable
Not good
- Its agents couldn't call my other MCP servers, like CodeGraph
- Its MCP server failed at Codex startup on Jul 22
Verdict My main agent driver for about five weeks: about 2,800 Codex sessions came from it in June and 1,600 in July. Then I stopped.
“I don't use rp-cli anymore, same for Flow. Update agent.md.”
-
Dropped
A CLI and agent skill that turns a codebase into a knowledge graph agents can query.
I wanted it for Giving agents a map of my book-layout app's code for architecture reviews and dead-code hunts.
Good
- Graph queries helped gather context for architecture reviews
Not good
- The graph often lagged behind the current commit, so agents re-read the source anyway
- Worktree copies created duplicate nodes until they were excluded
- Took up too much of the agent's context window
Verdict It ate my context. I switched to CodeGraph after two weeks.
“Graphify is eating too much context now. Make it stop being used.”
Now CodeGraph
04
Browsers and automation for agents
My agents need my logins to do real work, and every Linux option I tried kept losing them. That's why I ended up porting one myself.
-
Still using
A browser built for AI agents: each agent works in its own space inside my logged-in browser. The app is Mac-only; the agent harness in this repo is open source.
I wanted it for Letting Claude Code and Codex do research, YouTube Studio work and posting with my own logins, without taking over my tabs.
Good
- One logged-in browser for every agent, and agents don't grab each other's tabs
- On the Mac, agent work runs in the background and stays out of my way
- The open-source harness was complete enough that my agents ported it to Linux in one day (it passed 69/69 of its own browser tests)
Not good
- The browser app itself is closed source and Mac-only
- Cross-posting to Naver Clip and Instagram often stopped with “login expired”
- My Flow scripts broke whenever Google changed the Flow page
Verdict Every agent on both of my machines uses it: the app on the Mac, and my own Linux port of its harness on the PC since Oct 14, after the Linux options below kept losing my logins.
“OK, then let's make egolite the default. Set everything up that way.”
-
Still using
Downloads video, subtitles and metadata from YouTube, Instagram, X and many other sites.
I wanted it for Reference videos, auto subtitles and public view counts for research, competitor checks and my own channel numbers.
Good
- No login needed for public videos
- JSON output my agents can read directly
- 20 of 21 public Instagram reels downloaded without a login
Not good
- Korean channel titles came back machine-translated in flat lists, and view counts went empty when I asked for Korean
- Blocked from a cloud server; big batches and subtitle downloads got rate-limited
- Date-sorted YouTube search returned nothing
Verdict It's how my agents get public video files and subtitles. Titles and view counts now come from the YouTube Data API, where yt-dlp falls short.
“Huh? I've been making videos automatically on the Mac by combining HyperFrames, ffmpeg, yt-dlp and Flow (credits).”
-
Kept for later
Sends mouse and keyboard input on Linux, Wayland desktops included.
I wanted it for Giving agents mouse and keyboard control of my Linux desktop, like computer use on the Mac.
Good
- One small daemon, no agent framework
- Works on my Hyprland (Wayland) desktop
- The click and typing test passed the same day; setup took about 6 minutes
Not good
- Needs root and access to the kernel input device
- Input only: it can't see the screen
- Can't focus a window by itself
Verdict It works and it's still running, but nothing has needed it since the test day. Browser work goes through ego lite instead.
-
Kept for later
An agent skill that researches what people said about a topic in the last 30 days across Reddit, X, YouTube, Hacker News and the web.
I wanted it for Quick scans of recent talk on topics I was curious about.
Good
- Fast, linked summaries of recent discussion
Not good
- X search needs a logged-in session it didn't have
Verdict Still on the Mac, but I used it on only two days.
-
Dropped
An open-source Chromium browser with built-in AI agents and an MCP server that coding agents can drive.
I wanted it for The agent browser on my new Linux PC, once ego lite turned out to be Mac-only.
Good
- The Linux AppImage ran from my home folder without admin rights
- Its MCP server worked with Claude Code on the first try
- It handled real logged-in work, like scheduling Naver blog posts
Not good
- Running several copies meant identical windows and about 5 GB of RAM for three
- Logins copied between the copies kept breaking, so I signed in again and again
- Some of its control ports listened on every network interface, not just my machine
Verdict After a week of signing in over and over, Google Flow refused every job sent from the copied profiles on Oct 9. I retired it that afternoon.
“What's really annoying is logging into a new browser like this all the time (if it's going to be like this, I won't use BrowserOS). Find another way. egolite didn't have this problem.”
Now ego lite
-
Dropped
A community fork that runs the ego lite agent harness on Linux with a stock Chromium.
I wanted it for A Linux stand-in for ego lite after BrowserOS.
Good
- More complete than the other Linux fork I tried: installer, MCP server, agent dashboard
- Fast: about 1.1 s to open and read a page
- Two agents ran side by side once its isolation bugs were patched
Not good
- Stuck on the old v1 API while upstream ego lite moved to v2
- Background tabs froze, and its window mode copied logins into each space; four Google accounts were signed out at once on Oct 12
- Parallel agents moved each other's tabs until I patched it locally
Verdict It ran every agent and timer on my PC from Oct 9 to Oct 14. Logins kept breaking and agents kept clashing, so I had my agents port upstream v2 directly.
“When I run several projects, the agents keep interfering with each other in the browser. Fix it. Sometimes it's fine, sometimes there's a problem.”
Now My own Linux port of ego lite v2
-
Dropped
A small community fork, about 1,200 lines, that lets ego lite's agent harness drive a stock Chromium on Linux.
I wanted it for The same job: ego lite on the Linux PC.
Good
- Uses the current ego lite v2 API
- Small enough to read in one sitting
- Worked after a 14-line fix
Not good
- A one-day project with no updates after Oct 2
- No installer, no MCP server, no way to tell agent windows apart
- Out of the box, picking up work from a new shell failed
Verdict It worked, but agenticbrowser had more of what I needed that same evening, so I went with that one.
Now My own Linux port of ego lite v2
-
Dropped
Open-source computer-use drivers: an agent reads and clicks app windows in the background.
I wanted it for Letting the Hermes agent on my Mac drive my desktop apps (ChatGPT, Muse, T3 Code, Aside) without approving every click.
Good
- Clicks and scrolls in the background without moving my cursor
- The per-app allow-list held: Finder, which wasn't on it, was refused
Not good
- Unattended runs couldn't click anything until each action type was allow-listed
- Turning off full-display observation broke every action
- A web page shown in one app could carry instructions into another app that runs agents
Verdict It was only ever the driver under Hermes. When I deleted the Hermes bots on Oct 14, the plan of one agent driving every app went with them.
Now OpenMausBot for the bots, T3 Code for managing agents
-
Dropped
A social media scheduler you can host yourself, with a paid cloud version.
I wanted it for One free place to post and schedule to X, Threads and Instagram across several accounts.
Good
- Free to host yourself
- Ran on a small 2 GB server
Not good
- Only schedules posts: no commenting, outreach or finding customers
- Still needs your own developer app keys for every platform
- Heavy to run (it needs Temporal next to it), and AGPL-3.0 if you change a hosted copy
Verdict I wanted one tool that posts, promotes, comments and finds customers. Postiz only schedules and I couldn't shape it to my workflow, so it was gone within days and my agents post through my own scripts.
“I want to use it for posting content all the time, promoting, commenting and finding customers. Postiz doesn't satisfy me, and it isn't customized for me.”
-
Dropped
An installer and health checker that sets up command-line readers for X, Reddit, YouTube and other sites.
I wanted it for A personal digest of AI and hardware posts from X and Reddit, minus memes and ads.
Good
- One doctor command showed which sources worked
- My download matched the official release byte for byte
Not good
- Reddit needs a logged-in browser session; there's no zero-setup path
- It reads browser cookies, so a spare account is safer
Verdict The digest it fed stopped after about a week of rate limits and stale results.
Now OpenMausBot research bots
-
Dropped
A CLI plus Chrome extension that reads sites like X and Reddit through your logged-in browser.
I wanted it for Reading Reddit and my X timeline for that digest.
Good
- Worked with my logged-in browser once the extension was connected
Not good
- Kept opening new Chrome windows during scheduled runs
- Subreddit listings returned old posts without extra filtering
- Its dependency audit flagged several high-severity advisories
Verdict Annoying to run unattended, and the digest stopped in late June.
“It keeps opening new Chrome windows, which is annoying. Is there another way?”
-
Dropped
A CLI that reads and searches X with the cookies of a logged-in browser session.
I wanted it for Pulling posts from chosen AI accounts and searches into the digest.
Good
- Simple read and search commands with YAML or JSON output
Not good
- Hourly use got most X routes rate-limited or blocked within a day
- Repeated macOS keychain prompts while it read browser cookies
Verdict X blocked most of it within a day of hourly use. X search goes through Grok Bot now.
05
Guardrails, security and backups
When agents run commands without asking, these are what stop a bad day.
-
Still using
A hook that checks every shell command an AI agent wants to run and blocks destructive ones like rm -rf or git reset --hard.
I wanted it for Making sure no coding agent on my PC can wipe files by accident, even when agents run commands without asking.
Good
- One install guards Claude Code, Codex and the other agents
- Catches deletes hidden inside bash -c or python -c, and lets a grep that only mentions a command pass
- One-time allow codes make an approved exception easy
Not good
- The Codex and Antigravity CLI hooks didn't fire after the official install and needed manual fixes
- False positives: a >= inside a heredoc, deleting a merged branch, some redirects
- It checks command text, so a script file can still get around it
Verdict It does the one job I asked for: no agent can casually rm -rf my machine. About 450 blocks in its first 11 days. The false positives cost my agents some time, and that's fine.
“There's an open-source tool that stops an AI from wiping everything with rm -rf at the root. Find it.”
-
Still using
A scanner that finds API keys, passwords and tokens in files and git history.
I wanted it for Keeping keys out of anything I back up or publish, then checking the whole PC.
Good
- Fast, one binary, scans files and full git history
- Easy to put inside a publish script so nothing goes public unchecked
- A full-PC scan found old keys in backups of servers I had already shut down
Not good
- A whole-PC scan gives a lot of false positives to sort by hand
- Its quiet “no leaks found” output was once misread by an agent as “not installed”
Verdict It runs before every publish and every backup push. The full scan turned up keys I didn't know were still on disk; those old backups (188 files) were deleted the same morning.
“Yeah, I'll need this too. Install it.”
-
Still using
An encrypted, deduplicating command-line backup program.
I wanted it for A daily backup of my projects, agent sessions, settings and keys, so the work survives a dead disk.
Good
- Incremental: daily runs take about 4 minutes
- Encrypted; the backup is unreadable without the password
- The same tool worked with Google Drive and then with SSH to my Mac
Not good
- Lose the password and you lose the backup
- Skips the day when the target computer is away
- Command line only, so I never see it working
Verdict It does what I asked for, quietly. A week in, I had to ask what restic was.
“I'd like things like workflows and sessions backed up. If they're lost, it's a pain to start over from scratch.”
-
Still using
An MCP server and CLI that lets coding agents run reverse-engineering tools on apps, packages and websites without running them.
I wanted it for Seeing how an Android scanner app handles its settings, and checking software I run for backdoors.
Good
- Read-only: nothing from the target is executed
- One npx command to try it
- Says where it's blind and saves evidence files
Not good
- Each CLI call re-read the APK, about 23 seconds, until my agent kept one session open (0.4 s per call)
- Can't see inside native code without Ghidra, IDA or Hopper
- On its own it can't prove there's no backdoor; I also needed provenance checks and a network log
Verdict Three real jobs in ten days, including two backdoor checks that found nothing. A good structured start, not the whole answer. Many apps' terms forbid reverse engineering, so check before you point it at one.
-
Kept for later
Decodes Android APK files back into readable resources and a manifest.
I wanted it for Reading an app's strings and settings, as REA's resource decoder.
Good
- Decoded a 165 MB APK's resources in about 23 seconds
- Installs from a GitHub release with no admin rights
Not good
- Ships as a jar, so it needs Java and a launcher script
Verdict It did its part in one APK job and stays installed next to REA. I haven't had Android work since.
-
Kept for later
A headless MCP server for decompiling Android apps, built on the JADX core.
I wanted it for Decompiling Java and Kotlin methods through REA.
Good
- No JADX window needed
- Handled an app with 49,519 classes
- About 0.4 s per call once loaded
Not good
- Loading the APK takes over 20 seconds, every time unless a session stays open
- Needed a 10 to 12 GB Java heap
Verdict It worked on the one APK job and stays installed for the next one.
06
Voice, speech and music
For two weeks a clone of my voice narrated my videos. Since Oct 16 I film myself. These are the voice, transcription and music models from both phases.
-
Still using
A serving runtime for speech and multimodal models with an OpenAI-style audio API.
I wanted it for Running Higgs as a local voice server on my PC.
Good
- Served Higgs behind a simple speech API for every narration job
- Short lines in 2.6–4.7 seconds once running
Not good
- Hard to install here: several CUDA, compiler and kernel workarounds
- FP8 models like Drama 3 don't run on an RTX 3090 Ti
- Ignored the --port flag when a config file was given
Verdict It's the server behind Higgs. It works on my GPU, but only with workarounds.
-
Still using
Whisper speech-to-text plus forced alignment that gives a start and end time for every word.
I wanted it for Cutting clips exactly on the words of a known script, so lines don't start late or end on an idle face.
Good
- On 120 clips, lines ending with an idle face went from 98 to 1, and cut-off last words from 6 to 0
- Runs on the CPU at about 2 s a clip, no GPU fan
- Its word scores settle which of two readings was really said
Not good
- Can't align digits and symbols, so numbers must be spelled out in the script
- Still misses some long final syllables: a cut in one published video made “25,000” sound like “five thousand”
Verdict It fixed my clipped and late cuts, and now every video job uses it without my asking.
“Can you set it up so I don't have to tell you to use WhisperX?”
-
Still using
Alibaba Qwen's open speech-to-text model and a forced aligner for word timestamps, with Korean support.
I wanted it for Checking that generated voice clips say exactly the script, and timing cuts in Korean Shorts.
Good
- Far tighter Korean timing than whisper.cpp: 2 bad cuts vs 32 out of 154
- About 12 times faster on the same audio (10 s vs 119 s for 11 minutes)
- Caught real Whisper mistakes
Not good
- Needs a newer transformers release, so it lives in its own environment
- The aligner sometimes glues the pause before a line onto its first word
- Aligner input is limited to 5 minutes
Verdict I noticed the Qwen-timed videos looked better, and the numbers agreed. It's still my transcript check for generated voices.
“When we did it with Qwen the video came out better, didn't it? (the timing)”
-
Still using
A fast Whisper implementation on CTranslate2, for GPU or CPU.
I wanted it for Bulk-transcribing a day of trending Shorts on my GPU, and quick checks of finished edits.
Good
- 35 to 38 times real time on a 3090 Ti: 2,622 Shorts in 2,975 s
- Closer than YouTube's auto-captions in 7 of 10 spot checks
- Fine on the CPU for short checks
Not good
- Its voice filter dropped the last words or songs in 3 of 10 clips
- Long GPU runs are loud
- small.en misheard a number under background music
Verdict It's the quick CPU transcript check in my video engine. The overnight GPU job stopped with the rest of that job, because of the fan noise.
-
Still using
Microsoft's WavLM model, used here to turn a voice into a vector so two recordings can be compared.
I wanted it for Measuring whether a generated voice really sounds like me, instead of trusting an AI reviewer.
Good
- Stable numbers where an AI reviewer rated the same voice 8.5/10 once and 2/10 the next time
- Clear separation: real 0.912, avatar 0.881, another voice 0.614
- Small and runs on the CPU
Not good
- A rough guide only; it can't say whether a voice sounds natural
Verdict It gives me a repeatable same-speaker number, which AI listening reviews couldn't.
-
Still using
A free, open-source Mac dictation app that transcribes on the device and cleans up the text with a local model.
I wanted it for Talking to my AI agents instead of typing, without paying for a dictation app.
Good
- Fast: about 0.09 s to transcribe and 0.66 s to clean up per dictation
- Shows words live and fixes them as I talk
- 115 dictations in its first day
Not good
- Product names come out wrong: “Eco light” for ego lite, “Omar key” for Omarchy
- The cleanup model is closed source
- Mac only, and its fast models don't do Korean
Verdict Better than I expected, and good enough that I don't need a paid dictation app for now.
“I'm surprised it's a lot better than what I have expected. … I don't really need the paid ones as of now.”
-
Still using
An open model that turns lyrics and a short style prompt into a full song with vocals, with a no-vocals mode for background music.
I wanted it for Making songs and background music locally instead of paying for Suno.
Good
- Full songs with clear lyrics in 1–2 minutes on a 24 GB card, about 9 GB of VRAM
- Korean lyrics came out clear, though only Chinese and English are officially supported
- Writes a melody and chord score with each song, so it can become sheet music
Not good
- Takes vary a lot, so you make 2–3 and listen to each
- No-vocals mode still put a human voice into 2 of 6 background tracks
- Weights are CC BY-NC 4.0, with a later exception for individual creators
Verdict It's my go-to for local music now. My notes were about the songs, not the model: make it more upbeat, make it easier to sing.
“Make it a more upbeat tone. It seems too hard for ordinary people to sing.”
-
Kept for later
An open-weight text-to-speech model with zero-shot voice cloning and inline emotion tags.
I wanted it for Cloning my own voice to narrate English videos, with some feeling in it.
Good
- The clone sounded like me: speaker similarity 0.895, where two real recordings of me score 0.912
- Emotion tags gave clearly more feeling than Qwen3-TTS
- Close to my natural pace (120–142 words a minute; I talk at 133)
Not good
- Pitch came out about 3 semitones lower than my real voice
- Emotion-tagged lines sound like a different person when you hear them back to back
- Needs most of a 24 GB GPU, and the license is research and non-commercial with a credit line
Verdict It narrated everything from Oct 1 to Oct 15 and beat Qwen3-TTS on emotion. Since Oct 16 I film and voice my own videos, so it only comes out for voice tests.
“h2 definitely has emotion, it's usable. Though sometimes Qwen seems better (for calm parts) and h2 when there's emotion.”
Now My own voice, on camera
-
Kept for later
Fish Audio's open-weight expressive text-to-speech model (Drama 3, 35B mixture-of-experts).
I wanted it for Checking whether the new top-ranked open voice model beats Higgs for my cloned narration.
Good
- The highest clone similarity I measured: 0.955–0.979
- Emotion directions in brackets were audible line by line
Not good
- The official weights need a bigger, newer GPU; only an unofficial Q4 build ran on mine
- Dropped whole lines and left up to 25 seconds of silence on longer chunks
- Uses about 22 GB of VRAM
Verdict The numbers favored it, but by ear it was sometimes better and sometimes worse than Higgs. Not worth the extra checks and GPU.
“It's kind of unclear. Sometimes Drama 3 is better, and some are better with Higgs...”
Now Higgs Audio
-
Kept for later
A C++ (ggml) inference engine for audio models: speech, recognition, voice conversion.
I wanted it for Running Drama 3 on a 24 GB card, which the official runtime couldn't.
Good
- The only way I found to run Drama 3 on an RTX 3090 Ti
- A native binary and HTTP server, no Python process
- Faster than real time with the GPU codec: 11 s of audio in 7.6 s
Not good
- Needed its own compiler toolchain to build
- The documented settings for 24 GB cards ran out of memory; I tuned them by hand
- About 27 s to load the model on every run
Verdict It worked, but the model it was for went on hold after listening, so the build sits unused.
-
Good, not for me
A local voice studio app that bundles about 17 voice and cloning engines behind one UI and API.
I wanted it for Finding out whether it would clone my voice better than Higgs, or was just a UI.
Good
- Extras Higgs lacks: designing a voice from a description, dubbing, transcription
- OpenAI-compatible local API
Not good
- The same short-clip cloning method I already had
- The default model's weights are non-commercial, even with the paid tier
Verdict It wouldn't improve my narration, and the default model's license blocks a monetized channel.
-
Good, not for me
A small (164 MB) English speech-to-text model distilled from NVIDIA Parakeet, with word timestamps.
I wanted it for Checking whether a newer open model beats Whisper for narration timing.
Good
- Much better English timing than Whisper: 50 ms vs 285 ms median start error
- Small and fast: a minute of audio decodes in 0.6 s
Not good
- English only; Korean came out as gibberish
- About 10 s of model loading per command-line call
Verdict It won on English, but the channel that needed alignment then was Korean.
“Something more SOTA than Whisper came out (open source), Phonon or something? Compare it.”
-
Good, not for me
A speech noise-removal model that also ships as a small prebuilt command-line program for the CPU.
I wanted it for Cleaning background noise out of two videos recorded with my iPhone's built-in mic.
Good
- One 35 MB binary on the CPU, no Python environment
- Removed noise without changing my voice (similarity 0.98 to 1.00)
- Both full videos, about 5.7 minutes of audio, in about a minute
Not good
- A smaller quality gain than resemble-enhance's enhance mode
- Only removes noise; it doesn't fix the thin sound of a phone mic
Verdict The best of the four I compared that day, but a real microphone is less hassle than cleaning every recording.
“I'll just use a mic myself. It's a hassle, don't spend compute on this.”
Now A real microphone
-
Good, not for me
A speech denoiser plus a generative enhancer that tries to restore studio-like sound.
I wanted it for The same phone-mic cleanup.
Good
- The highest quality score of the day in enhance mode
- Denoise-only mode kept my voice intact
- 11 seconds for two short clips on the GPU
Not good
- Enhance mode changed my voice (similarity dropped to 0.913)
- Its built-in model download needs git-lfs, so the weights had to be fetched by hand
Verdict The cleanest sound came at the cost of my voice, and I decided on a microphone anyway.
Now A real microphone
-
Good, not for me
A speech toolkit with models for enhancement, speaker separation and super-resolution.
I wanted it for The same phone-mic cleanup.
Good
- Scores close to DeepFilterNet, with slightly better voice similarity on one clip
- Works at 48 kHz
Not good
- About 5.9 GB of environment for one model
- Not clearly better than the 35 MB DeepFilterNet binary
Verdict About as good as DeepFilterNet, much heavier, and I went with a microphone.
Now A real microphone
-
Dropped
Alibaba Qwen's open text-to-speech models with voice cloning and voice design.
I wanted it for My first test of cloning my voice for English narration on my own GPU.
Good
- The timbre was close to my voice
- Very few word errors (0–0.9%)
- Fine for calm lines
Not good
- No emotion in the delivery
- Read 33–40% faster than I talk, with no speed setting
- One take flipped “can't” into “can”, so every line needs a transcript check
Verdict Similar, but flat. Higgs gave more emotion and I wanted one model for everything, so I switched the same evening.
“It's similar. It's not the timbre, it just seems to have no emotion. What should we do?”
Now Higgs Audio
-
Dropped
A Python tool that uses Microsoft Edge's online text-to-speech voices without the browser.
I wanted it for A one-hour, two-host research podcast to listen to on my phone.
Good
- Free, no GPU, no model download
- A full hour of audio in about 15 minutes
Not good
- Sounded robotic
- Two voices trading long paragraphs felt stiff
Verdict Robotic. My podcast project doesn't use it anymore.
“The voice is too robotic. Next time use a more natural TTS, like NotebookLM. Microsoft Edge TTS is really bad.”
-
Dropped
OpenAI's open speech-to-text models and reference Python package.
I wanted it for The reference transcript of my voice for a cloning test.
Good
- Easy to run, and its models live on inside faster-whisper and WhisperX
Not good
- Wrote “moat” as “mode” and dropped a phrase in a 20-second clip
Verdict It isn't the best open model anymore, and Qwen3-ASR caught its mistakes, so Qwen3-ASR does the checks.
“Whisper isn't SOTA (among open-source ones).”
-
Dropped
A C/C++ port of Whisper that runs locally on CPU or Apple silicon.
I wanted it for Finding where each narration line starts, so scene cuts land between sentences.
Good
- Runs locally on Mac and PC with no account
- Fine as a second check of what was said
Not good
- Line starts often 0.3 to 1 s early; 32 of 154 cuts landed inside sentences
- Skipped speech when gaps were pure digital silence
- Misheard Korean names and words
Verdict Its timing caused cuts mid-sentence and clipped endings. The Qwen aligner fixed both.
Now Qwen3 ForcedAligner for Korean, WhisperX for English
-
Dropped
An open text-to-music model (MIT code and weights) that makes songs or instrumentals from a description, BPM and length.
I wanted it for Background music for three contest ads that had to run outside YouTube, where library tracks weren't licensed.
Good
- MIT license, so the music could go into ads anywhere
- About 10 seconds per 30–40 second take on a 3090 Ti
- With 3 takes per cue, every cue had a usable one
Not good
- About 36 GB on disk with models and environment
- About 40 minutes of model downloads before the first sound
Verdict It did the job and I approved the ads. I deleted the 36 GB install after the contest for space, not quality, and I use YuE for new music.
“Yes, go ahead, I think it's fine.”
Now YuE for new music; the tracks I made are still in use
-
Dropped
Zero-shot voice conversion: makes a recording sound like a short reference voice, with no training. The repo is now archived.
I wanted it for Fixing one line in an AI-generated ad where a mother's voice sounded far too young, without breaking lip sync.
Good
- A 3 to 4 second reference was enough
- Keeps the exact timing, so lip sync stays intact
- Under a second per 4-second line on the GPU
Not good
- Changes tone only; it can't fix cut-off or mispronounced words
- Repository archived
Verdict It lowered the voice as intended, but I had the weak shots redone instead of patched, so the conversion wasn't needed in the final cut.
07
Video, editing and effects
I make 1–3 videos a day with agents doing the edit. Most of these were tested against that: does it beat what I already run?
-
Still using
HeyGen's open-source framework that turns HTML and animation into MP4, made to be driven by AI agents.
I wanted it for Captions, titles, motion graphics and the final render for every Short and episode, with agents writing the edit.
Good
- Agents write and fix whole edits as HTML, so I only give notes
- One tool for captions, motion graphics, preview, checks and the render
- Fast enough for 1–3 videos a day
Not good
- Slow next to a compiled renderer: 75 s vs about 14 s for the same 22-second clip in FFrames
hyperframes checkhas a fixed 2-minute limit that heavy episodes hit- Its built-in background remover took 2 min 31 s on CPU for 8 seconds of video and cut off part of a finger
Verdict It edits and renders everything I make: 38 jamin labs project folders by Oct 20. I tested a faster renderer and kept this one, because render time isn't my bottleneck.
“Instead of recording my video, use hyperframes to show how I'm using the actual AI products.”
-
Still using
A video matting model that cuts a person out of the background, frame by frame, on the GPU.
I wanted it for Putting me on a branded backdrop in episodes.
Good
- 4.3 s vs 2 min 31 s for HyperFrames' CPU remover on the same 8-second clip
- Kept my fingers whole
- Clean results outdoors
Not good
- It decided a strap hanging on the wall was part of me
- Transparent video is heavy, and it slowed or broke my checks and renders
Verdict After comparing two cuts of episode 6, I kept the real background in a rounded window. Matting now only runs on shots where the background shows nothing.
“The one that keeps the background is better.”
Now My real background in a rounded window, for most shots
-
Kept for later
Meta's model that masks objects in images and video from a text prompt.
I wanted it for Masking my clothes so the try-on model could pick better keyframes.
Good
- Fast and light: 33 frames in 10 s with about 2.3 GB of VRAM
- The prompt “clothing” found the garment on all 33 frames
Not good
- Gated download that asks for your birth date and affiliation
- Very prompt-sensitive: “shirt” found nothing
Verdict The masks were right, but they didn't make the clothes swap better. The model stays downloaded.
-
Good, not for me
A Rust framework for rendering video from code, built for speed.
I wanted it for Seeing whether my Shorts render faster or better than in HyperFrames.
Good
- About 5 times faster: 5.7 s plus 8.3 s of encoding vs 75 s for the same clip
- Output nearly identical to the HyperFrames render (SSIM 0.924)
Not good
- Every edit means a Rust rebuild of about 14 s
- The build broke over an x264 version mismatch and needed a forced FFmpeg build
- Setup kept asking for sudo
Verdict Much faster, but at 1–3 Shorts a day rendering isn't what slows me down. It would pay off for mass renders, like 100 ad variants.
“But having to do sudo like this every time is annoying. Can't we avoid it?”
Now HyperFrames
-
Good, not for me
ByteDance's open lip-sync model: it redraws the mouth in a video to match new audio.
I wanted it for Fixing AI-generated contest ad shots where the mouth didn't match the line, without paying for a lip-sync service.
Good
- Free and local, and the weights allow commercial use
- Fixed 5 shots across 3 ads with a clean mouth area
- Ready in about 15 minutes
Not good
- A step below the best paid lip-sync services
- Needs 25 fps input
- Not made for shots with several people
Verdict It did the job, but I film myself now, so there's nothing to lip-sync. I removed it in a disk cleanup on Oct 15.
-
Good, not for me
An open-source agent toolkit for making videos end to end with many generation services.
I wanted it for Seeing whether it beat my own Shorts pipeline.
Good
- A real codebase with a real test suite: 1,875 tests passed
- Clean security basics
- Two ideas I borrowed: cheap machine checks before a human looks, and searching public-domain footage
Not good
- The budget cap and approval gates in the README weren't enforced in the code
- Retrying a paid video job could charge twice
- It stops at the MP4: no upload, analytics or feedback loop
Verdict My own pipeline already goes further, from upload to results that feed the next script, and it runs on a subscription.
“Interesting. Then just delete it (if you installed it).”
-
Didn't work for me
Research code for video virtual try-on: it changes the clothes a person wears in a video.
I wanted it for Swapping the pajamas I sometimes wear on camera for a real-looking shirt.
Good
- Recent and built for exactly this, with code and weights public
- Ran on a 24 GB card once the models were quantized
Not good
- My face and hands changed too, which rules it out for a talking-head channel
- Dark blotches in the fabric
- Non-commercial license (CC BY-NC-SA 4.0)
Verdict The swap ran, but at the settings a 24 GB card allows it wasn't real-looking enough for my videos.
08
Running models on my own GPU
An RTX 3090 Ti with 24 GB of memory. It runs most open models, but what I have spare is subscription usage, not GPU time.
-
Kept for later
A node-graph app and API for running image and video models locally.
I wanted it for Running open image and video models on my own card instead of paying per generation.
Good
- Runs current open models on a 24 GB card with official workflows
- Scriptable through its API, not just the web UI
- Nothing leaves the PC
Not good
- Uses nearly all 24 GB of VRAM
- Model files are huge: 17 GB for one image model
- Updates and custom nodes can break workflows, so I keep my install frozen
Verdict I start it for the occasional local model test. It isn't part of my daily pipeline.
-
Kept for later
A fork of the GGUF loader node that adds newer formats such as Qwen-Image 2.1.
I wanted it for Loading Qwen-Image 2.1 after the original node failed.
Good
- Loaded the model the original couldn't, at about 1.7 s per step at 1024x1024
- A small change, checked by diff before I used it
Not good
- A fork, so it can drift from the original
Verdict It works and stays installed, but I don't use the image model I needed it for anymore.
-
Kept for later
An open model that moves the motion in a driving video onto the people in a photo.
I wanted it for Making a two-person photo do a trending dance without a paid tool, after Google Flow kept failing at it.
Good
- Followed the choreography and kept two people distinct in one pass
- Kept the clothes and body proportions from the photo
- Fits a 24 GB card at 512x512
Not good
- Faces drift with expressions; hands and feet blur
- Clothes deform on fast turns
- One 3.4-second clip took 161 s and 22.8 GB of GPU memory
Verdict A promising first test, not a usable clip. I kept the setup but haven't tried longer or 720p runs.
-
Didn't work for me
The original ComfyUI node for loading smaller, quantized GGUF model files.
I wanted it for Loading a GGUF build of Qwen-Image 2.1.
Good
- The widely used original GGUF loader
Not good
- It didn't know the Qwen-Image 2.1 format on Oct 1, so the model wouldn't load
Verdict The model card pointed to a fork that only adds the new format. My agent diffed the two and swapped it in.
-
Didn't work for me
The download and upload client the Hugging Face hub library uses by default.
I wanted it for Downloading large open model weights, 7.3 GB and 24.6 GB files.
Good
- The partial file it left was in order, so I could resume over HTTP and check it by hash
Not good
- Stalled at 0 KB/s on two large downloads, three times in total
- Restarting threw away the partial download
Verdict Both big downloads stalled on my connection, so I turn Xet off for large models.
Now Plain HTTP downloads with Xet turned off, resumed and checked by sha256 (my field note on the fix)
-
Dropped
Alibaba Qwen's open image generation and editing model.
I wanted it for Free, unlimited local images instead of ChatGPT's.
Good
- No per-image cost on a 24 GB card
- Long, detailed prompts improved results a lot
Not good
- Worse than ChatGPT's images, to my eye
- Korean text in images came out as gibberish
- 1–2 minutes per image and 17 GB of files
Verdict I went back to hosted image models.
“The quality is worse than ChatGPT's. Do I need to hook it up to other nodes to get its full performance?”
Now GPT Image, through ChatGPT and Codex
-
Dropped
A fast server for running open-weight LLMs on your own GPU behind an OpenAI-style API.
I wanted it for Sorting about 8,000 collected social posts a day into topics and summarizing short videos, instead of paying for an API.
Good
- Easy OpenAI-compatible server that fit a 24 GB card
- Qwen3.8-27B agreed with Claude's labels on 93% of posts
- A full day of 8,047 posts took 23.5 minutes
Not good
- Long runs kept the GPU at 100% and the fan at full speed for over 90 minutes
- Takes about 21 GiB of GPU memory, so nothing else can use the card
- The money saved was tiny, because I already pay flat-rate AI plans
Verdict It worked, but what I have spare is subscription usage, not GPU time, and the fan noise wasn't worth it.
“The Qwen LLM is useless to me and just makes the GPU fan loud. Don't use it from now on.”
Now The AI plans I already pay for
09
Building my own apps
Parts I tried while building my own apps: CueRoom (my iPhone camera app), a book-layout desktop app, a YouTube summary extension, Torah Spark and a few small games.
-
Still using
Google's on-device machine learning toolkit; its Face Landmarker finds face points in images and video.
I wanted it for Tracking my face on the PC so CueRoom's skin touch-up can run there after the phone sends a raw recording.
Good
- Free and local, no paid SDK or cloud
- Found my face on 26,178 of 26,179 frames of a 14.5-minute take
- Video mode keeps tracking steady across frames
Not good
- On my setup it runs on the CPU, so that 14.5-minute take took 72 minutes to render
- Its face points differ from Apple Vision's, so the masks had to be tuned again
Verdict It's the face tracker for every new CueRoom recording. Slow here, but it works.
-
Still using
Generates an Xcode project from a YAML file and the folder structure.
I wanted it for Building iPhone apps when the code is written on my Linux PC and built on my Mac.
Good
- The project is a short text file, so code edited on Linux syncs to the Mac and rebuilds
- No hand-edited Xcode project files to conflict
Not good
- It compiles every file in the source folder, so parallel agents had to work in separate folders
Verdict It's part of every CueRoom build, 0.3.1 through 0.3.27.
-
Still using
Python bindings for FFmpeg, for decoding and encoding audio and video frame by frame.
I wanted it for Decoding iPhone recordings and re-encoding the touched-up video on the PC with the original timing and audio.
Good
- Frame-level control with exact timestamps
- Works with GPU H.264 encoding
Not good
- Its bundled FFmpeg shifted the iPhone audio timing by 18.7 ms
- Its rotation helper wrote a different matrix than the iPhone file had
Verdict Still decoding and encoding the video. The audio step moved to the system ffmpeg, and both real takes then passed.
-
Still using
A framework and toolchain for building iOS and Android apps from one React Native codebase.
I wanted it for Torah Spark, a Bible-reading app for churches: daily reading, read checks, monthly study.
Good
- One codebase for iOS and Android, and most church members are likely on Android
- I can work on it from Linux and test on a real iPhone
- The cloud build service is optional; local builds cost nothing
Not good
- Breaking changes every SDK release
- iOS builds still need a Mac or the paid cloud service
Verdict I tried a native SwiftUI port, then came back to Expo so there's one app to keep up, built from Linux.
“I'm just curious in long term that if I would just build ios and kotlin, or comeback to expo, because I don't want to pay stupid build fee”
-
Still using
A starter template for desktop apps with Electron Forge, Vite, React, TypeScript, Tailwind and shadcn/ui.
I wanted it for Turning my book-layout web app into a local-first desktop app, mostly so PDF export matches the editor.
Good
- A fast start for an Electron app with shadcn/ui
- Its lint and test checks became my project's gate (1,500+ tests by mid-July)
Not good
- Electron's binary install broke once and needed a manual reinstall
Verdict It's the base of my book-layout desktop app, which is paused since July.
“I git-cloned luanRoger's and changed it myself. I don't think I ever pushed it.”
-
Still using
A crowdsourced database and extension that marks sponsor segments in YouTube videos, with a public API.
I wanted it for Cutting sponsor reads out of transcripts before my YouTube summary extension sends them to an LLM.
Good
- Its time ranges line up directly with timestamped transcripts
- Keeps ad reads out of summaries and saves tokens
Not good
- Only helps on videos where people have submitted segments
Verdict It's part of that extension's summary step, with a 2-second timeout so a slow lookup never blocks a summary.
“I want to add this: when summarizing, leave the sponsor parts out before passing to the LLM. How do I do it?”
-
Still using
Streams SQLite changes to S3-compatible storage for continuous backups and point-in-time restores.
I wanted it for Backups for the SQLite databases behind my Rails apps.
Good
- Ran for months without attention
- A restore test worked
Not good
- The Kamal accessory setup took many tries
- Its container has no shell, which made debugging harder
Verdict It's the backup for my production SQLite apps.
“I need a backup plan for my backend. I have installed gem "litestream", "~> 0.14.0" use it.”
-
Kept for later
The open-source 3D suite; agents drive it with Python scripts and headless renders.
I wanted it for 3D assets for my own iOS games: an AR face mask, a house from a floor plan, a Bible-story scene.
Good
- Free, I own the source assets, and it exports USDZ for ARKit
- Fully scriptable, and headless renders work for agents
Not good
- A version change broke one script setting
- Headless runs crashed inside the Codex sandbox
Verdict It's my 3D tool of choice for game assets; last used Sep 8.
“so it looks like blender is my best option?”
-
Good, not for me
A cross-platform C++ GPU filter library with beauty effects: skin smoothing, lipstick, blush, face reshaping.
I wanted it for The same free skin touch-up for CueRoom.
Good
- Apache-2.0 core with many ready-made beauty filters
Not good
- Ships a closed, prebuilt face-tracking library with no license file I could find
- The makeup effects depend on that library
Verdict I couldn't confirm the terms of the bundled face tracker, so I went with MetalPetal and Apple Vision.
Now MetalPetal, then the PC
-
Didn't work for me
An open-source web builder framework for drag-and-drop HTML and CSS editing.
I wanted it for Visually editing AI-generated pages inside my book-layout app.
Good
- A mature visual editor: canvas, style manager, undo, HTML and CSS export
Not good
- It re-serializes imported HTML and CSS, so exact round trips and stable element IDs aren't guaranteed
- Failed my editor's requirements on all 11 sample pages
Verdict My agents' test showed it couldn't keep the stable element IDs the editor needs, so we wrote our own editor layer.
Now A custom HTML editor layer inside the app
-
Dropped
A GPU image and video processing framework for iOS and macOS, built on Metal.
I wanted it for A free, natural skin touch-up for CueRoom's front camera, after a paid beauty SDK cost too much.
Good
- MIT, no fees or license keys
- Live 1080p front-camera preview at about 30 fps on an iPhone 15
- The same filter code worked for the live preview and the saved video
Not good
- Its built-in skin smoothing was too weak to see, even at full strength; custom shaders were needed
- No face detection of its own, so masks came from Apple Vision
- Processing a 10-minute take on the phone took minutes and held up sending it
Verdict It worked after six builds of custom shaders, but long takes were too slow on the phone. New recordings are now touched up on the PC.
“Natural processing takes too long, recommend the best way. I shot about 10 minutes and the processing seems to take several minutes.”
Now PC-side touch-up with MediaPipe
-
Dropped
Korea Investment & Securities' official Python samples for its Open API: quotes, investor flows, derivatives, orders.
I wanted it for A paper-trading bot for Korean stocks that watches foreign and institutional buying.
Good
- The official samples made the right endpoints quick to find
- Quotes, investor flows and derivatives in one API
Not good
- Request limits too tight for whole-market collection
- Timeouts and dropped connections during market hours when jobs overlapped
Verdict The limits blocked the whole-market data I wanted, and I switched the bot off in July.
“Besides the KIS portal, isn't there somewhere I can pull data with a generous rate limit?”
10
My Mac and Linux desktop
A MacBook Pro I carry and a Linux PC that runs the agents day and night.
-
Still using
An opinionated Arch Linux setup built around the Hyprland tiling window manager.
I wanted it for The operating system for the always-on PC that runs my agents, bots and renders.
Good
- Shortcuts, bar widgets and window rules are plain config files an agent can change
- Plugins exist, like Familiar Desktop below
- Actively developed; it keeps getting better
Not good
- Tiling windows and workspaces felt harder to live with than Debian at first
- After suspend the screen sometimes stayed black on my NVIDIA card, until a fix script
- System packages need the admin password, so agents can't install them alone
Verdict It runs my agent PC. After Mac-style shortcuts and floating windows, I'd recommend it to people who know what they're doing, and Ubuntu to first-timers.
“Omarchy is more annoying to run than Debian (windows and spaces). How do I fix that? Can it be customized?”
-
Still using
An Omarchy plugin that adds a mouse-friendly dock and Mac- or Windows-style window title bars.
I wanted it for Making the new Linux desktop feel more like my Mac.
Good
- A click-to-launch dock that also lists each app's open windows
- Mac-style title bars with close, minimize and maximize
- Checks its downloads against checksums and stops on unsupported versions
Not good
- Tied to one exact Hyprland version, so desktop updates can break it
- Its author retired it on Oct 13, as Omarchy adds similar features to its core
- Omarchy's normal plugin removal leaves its config behind
Verdict Still on my PC, and still the reason it feels like a Mac. Existing installs keep working.
-
Still using
A fast, GPU-accelerated terminal for macOS and Linux.
I wanted it for A terminal for agents and shell work where I type both Korean and English.
Good
- Fast and native on both Mac and Linux
- Plain-text config an agent changes in seconds
- Picked over Warp for Korean and English input and for keeping everything local
Not good
- On the Mac, several lines of Korean were hard to read, even with the D2Coding font
Verdict It's the default terminal on my Linux PC. On the Mac it didn't stick because of the Korean text.
“I still don't like it. One line is fine, but when there are several lines of Korean it's somehow hard to read comfortably.”
-
Still using
A WireGuard-based private network that links your devices as if they were on one LAN. The client is open source.
I wanted it for Connecting my Mac, Linux PC and iPhone, so I can drive the PC's agents from anywhere and send phone videos to it.
Good
- The PC is reachable from anywhere, free for personal use
- Taildrop sends files between all my devices
- Serves a local web app to my own devices over HTTPS
Not good
- iPhone-to-PC Taildrop looked sent, but the file never arrived when iOS restarted the app mid-transfer
- Straight from the phone it ran at about 0.9 MB/s
Verdict It's the link between all three devices, and the reason work keeps running on the PC while my laptop is closed. Phone videos now go to the Mac first.
“Videos don't really go from iOS to the PC over Tailscale (it pretends to send, but nothing actually arrives. Why? iOS to Mac works).”
-
Still using
A command-line tool for Apple Silicon Macs that sets fan speeds and battery charge limits.
I wanted it for Cooling my MacBook, which got hot running several coding agents.
Good
- Open source (MIT) and scriptable, so a small app can sit on top of it
- Fan changes took effect within seconds: 4,100 RPM asked, 4,088 and 4,122 read back
- Also sets a battery charge limit
Not good
- Needs a root helper, installed once with the admin password
- Its default sensor list missed four efficiency-core temperature sensors on my M2 Pro
- No menu bar app of its own
Verdict It's the engine under MacFanSpeedy, the fan app Codex built for me on top of it in about 36 minutes.
Now MacFanSpeedy, my own app on top of it
-
Still using
A command-line cleaner for macOS: caches, apps and their leftovers, big files, old build folders.
I wanted it for Freeing disk space on the MacBook.
Good
- One command for cleaning, uninstalling and finding large files
- Can report first, before anything is deleted; one report found 38.4 GB in Downloads, much of it duplicate archives
- Free and open source
Not good
- No Linux version, and the author doesn't plan one
- Big finds like duplicate archives still need me to decide
Verdict Still on the Mac; I ran it about 45 times between January and October. On Linux I use kondo for the build-folder part.
-
Still using
Finds and deletes build folders and dependencies (node_modules, target, .venv) inside projects.
I wanted it for Clearing old build output on the Linux PC, in place of Mole.
Good
- Lists what it would delete before deleting
- Can stick to projects untouched for a while
- One binary, no admin rights
Not good
- It can't know which folders a scheduled job still uses; I checked those by hand
- On my compressed file system I got back less than it listed: 3.4 GB listed, about 1–2 GB freed
Verdict It's the build-folder cleaner on my PC. One cleanup so far.
-
Didn't work for me
A command-line tool that lets AI agents read and send KakaoTalk messages on a Mac.
I wanted it for Letting my agents read and send my KakaoTalk messages.
Good
- Found the KakaoTalk app, its database and the permissions correctly
- MIT license
Not good
- Auth failed: it couldn't work out my account's user ID with the current KakaoTalk for Mac (a known upstream issue)
- The version the binary printed didn't match the Homebrew package
- Needs Full Disk Access
Verdict It never got past the auth step, and I removed it five weeks later.
-
Dropped
A free Mac app for watching temperatures and setting fan speeds. The GitHub repo holds releases and issues; the source isn't public.
I wanted it for Keeping the MacBook cool while agents run.
Good
- Free, and shows live fan speeds and temperatures
- Can follow a temperature sensor
- Puts the fans back on automatic when it quits
Not good
- No global keyboard shortcut; requests for one have been open since 2020
- Not open source, and its license forbids changing it
- Its root helper kept running after the app was gone
Verdict I couldn't change it and it had no shortcut. I wanted one simple shortcut, not per-fan controls, so I had my own app built.
“Yes, I made it because I wanted something simpler. It's free software, but I can't change the code and there's no shortcut. Instead of controlling both fans, I just wanted to use it simply with a shortcut.”
Now MacFanSpeedy
-
Dropped
A menu bar manager for macOS that hides and arranges menu bar icons.
I wanted it for Tidying a crowded Mac menu bar.
Verdict I'd stopped using it, and it went in a July cleanup of unused apps.
“I don't use Ice either, delete it.”
Not sponsored, and nobody here asked to be reviewed. These are my results on my machines (an M2 Pro MacBook and a Linux PC with an RTX 3090 Ti) on the dates shown; most of these projects move fast, so check the repo before you decide.