Handles jobs no tool covers
Ask it to check yesterday's logs, fix a failing script or pull this week's numbers. It reads the repo, runs commands and works out the steps itself, the way it would in a terminal.
Piscord
A real coding agent in your team's Discord. 4,042 installs in the past year.
01 / 03npm i -g piscord
↗
0+GitHub stars
0+npm installs, past year
0Small businesses running my team AI
01 — Piscord · Discord gateway · TypeScript · npm
Not a chatbot. A coding agent the whole team can talk to.
Stylised demo · not a live botnpm i -g piscord
# agentsColy · APP · online
C
Crokily
Piscord, introduce yourself.
π
Coly APP
I'm pi, running in this channel. Ask me to read the repo, run a command or write a script, and I'll answer here. If I'm busy, your message waits its turn in the queue.
C
Crokily
/pi statusπ
Coly APP
modelfrom pi's catalogthinkingmediumcwd~/repos/appqueue0 waiting · 0 stuck
Message #agents
RoleAuthor and maintainer, since January 2026
In useTeam AI inside 6 small businesses · 4,042 npm installs in the past year
Community48 stars · issues and pull requests from 10 outside developers
Works when nobody asks
Tell it in chat to send a summary every morning or remind the team at 2pm. It schedules the job itself and posts the result to the channel, files included.
Runs on your machine, and keeps running
One command sets it up as a background service, using your own pi login and models. Every message waits in a queue, so a busy agent or a restart loses nothing.
02 — herdr-lazygit · herdr plugin · Shell · MIT
Git, without leaving your agents.
herdr runs coding agents side by side in the terminal. This plugin puts lazygit in a sidebar next to them, and one key writes the commit message with whichever AI CLI you are already logged into.
Stylised demo · the real thing is a TUIprefix+g
herdr · ~/repos/app · lazygitsidebar 42 cols
M notify.tsM store.tsA tasks.ts
claude · haiku · reading the staged diff
3 candidates · Enter=commit · type to edit · Esc=cancel
› feat: add due dates and overdue task detection
feat: support task deadlines and overdue filtering
feat: track optional task due dates in store
✓ Committed 414bf66 · feat: add due dates and overdue task detection
C commit with AIU expand; settingsp pullP push
RoleAuthor and maintainer · 37 stars
InstallOne herdr command, or hand the README prompt to an agent
Tested onherdr 0.7.0 to 0.9.0, macOS and Ubuntu, every release
Right beside your agents
A narrow sidebar sits next to the agents herdr is running. Press U for the full lazygit layout when you need it, and U again to get your space back.
The AI you already pay for
It calls claude, codex, opencode or gemini through each CLI's own login. No API keys, and your diff only goes to the AI when you press C.
Three messages to choose from
Press C. The AI reads your staged diff and writes three commit messages. Pick one, edit it or type your own, and Enter commits.
03 — Involution Hell · Community · Next.js · Fumadocs
Study notes that students keep current.
A free, open-source place where students share course notes, study guides and project experience, in English and Chinese. I co-founded it in 2025 and still help maintain it.
Community paper · ZH / ENCo-founded 2025 · Sydney
Involution
Hell
Hell
A non-profit, open-source platform where students share study resources, course notes and project experience. Built by current students and recent grads.
165 ★46 forks32 contributors156 commits by me
RoleOne of the founders · core maintainer · second-most commits
Community165 stars · 46 forks · 32 contributors, since September 2025
StackNext.js · Fumadocs · Vercel
Notes that read like a book
Every folder is a chapter. Course notes, study guides and project write-ups, in English and Chinese, free to read.
Anyone can add a page
Contribution guides, pull request checks and content checks let 32 people ship notes, translations and fixes without stepping on each other.
Reachable from chat and from AI
A Discord bot carries links and messages from the community chat to the site. Agent tools and an MCP server let an AI search and publish articles.
More work · 08
Smaller experiments, all open source.
Most start as a question. What happens when an agent maintains my wiki, runs a team, or lives in a mail folder?
01pi-tagPiscord's sibling for Slack. Self-hosted, Socket Mode, 321 installs in the past year.TypeScript · npm↗
02rlm-doc-explorerWatch a Recursive Language Model work through documents step by step. Bring your own key.TypeScript↗
03MyWikiA wiki the LLM writes and I read, on the llm-wiki.md pattern.TypeScript↗
04pi-agent-teamA team of autonomous agents, run with nothing but directories and a few conventions.TypeScript↗
05flue-mailA mail app that lives in a folder you own, with a Flue agent running in the browser. No install.TypeScript↗
06colys-agent-labThe agent skills I use every day, including the design method behind this site.Python↗
07pideployOne-click deploys of a personal OpenClaw assistant, in under 60 seconds.TypeScript↗
08Personal AI Audio HubRealtime transcription and translation, with a native Rust and GPUI captions app.Rust · TypeScript↗
Notes · from my own runs
What I find by running agents hard.
I write up what I test on LinkedIn. Long reads, with the numbers from my own runs.
01Why customers think your AI agent is dumbA friend's agent had more than 70 tools and still felt dumb. Workflows, tool agents and code-mode agents, and when each one is right.Aug 2026↗
02Dynamic Workflows, not Opus 4.8, is the release that matters16 runs, about 194 sub-agents, 2.72 million tokens. Claude writing its own orchestration, read as a recursive language model with depth one.May 2026↗
03The future of app interaction, with CodeAct and MCPApps move data today. With CodeAct and MCP, they could move functionality.Apr 2025↗
About
I build agents that find their own way.
I'm Colin, an AI agent engineer in Sydney. Half my time goes into finding out what the newest agents can really do, usually by running them hard and counting what happens. The other half goes into making that useful for teams who will never open an agent tool.
I have built all three kinds of agent: workflows where the program decides the path, tool agents that pick from tools I wrote, and code-mode agents that write their own way through a problem. Each has its place. The last kind is where the big jumps come from, so most of what I build puts one where a team already works.
- 01Agent runtimesgateways, queues, sessions, sandboxes
- 02Agents at workteam AI in the chat tools people already use
- 03Developer toolsterminal plugins, agent skills
Sydney · UTC+10UNSW · Master of ITOpen to AI engineering roles