I've been building this repo public since day one, roughly 7 weeks now with Claude Code. Here's where it's at. Feels good to be so close.
The short version: AIPass is a local CLI framework where AI agents have persistent identity, memory, and communication. They share the same filesystem, same project, same files - no sandboxes, no isolation. pip install aipass, run two commands, and your agent picks up where it left off tomorrow.
You don't need 11 agents to get value. One agent on one project with persistent memory is already a different experience. Come back the next day, say hi, and it knows what you were working on, what broke, what the plan was. No re-explaining. That alone is worth the install.
What I was actually trying to solve: AI already remembers things now - some setups are good, some are trash. That part's handled. What wasn't handled was me being the coordinator between multiple agents - copying context between tools, keeping track of who's doing what, manually dispatching work. I was the glue holding the workflow together. Most multi-agent frameworks run agents in parallel, but they isolate every agent in its own sandbox. One agent can't see what another just built. That's not a team.
That's a room full of people wearing headphones.
So the core idea: agents get identity files, session history, and collaboration patterns - three JSON files in a .trinity/ directory. Plain text, git diff-able, no database. But the real thing is they share the workspace. One agent sees what another just committed. They message each other through local mailboxes. Work as a team, or alone. Have just one agent helping you on a project, party plan, journal, hobby, school work, dev work - literally anything you can think of. Or go big, 50 agents building a rocketship to Mars lol. Sup Elon.
There's a command router (drone) so one command reaches any agent.
pip install aipass
aipass init
aipass init agent my-agent
cd my-agent
claude codex or gemini too, mostly claude code tested rn
Where it's at now: 11 agents, 4,000+ tests, 400+ PRs (I know), automated quality checks across every branch. Works with Claude Code, Codex, and Gemini CLI. It's on PyPI. Tonight I created a fresh test project, spun up 3 agents, and had them test every service from a real user's perspective - email between agents, plan creation, memory writes, vector search, git commits. Most things just worked. The bugs I found were about the framework not monitoring external projects the same way it monitors itself. Exactly the kind of stuff you only catch by eating your own dogfood.
Recent addition I'm pretty happy with: watchdog. When you dispatch work to an agent, you used to just... hope it finished. Now watchdog monitors the agent's process and wakes you when it's done - whether it succeeded, crashed, or silently exited without finishing. It's the difference between babysitting your agents and actually trusting them to work while you do something else. 5 handlers, 130 tests, replaced a hacky bash one-liner.
Coming soon: an onboarding agent that walks new users through setup interactively - system checks, first agent creation, guided tour. It's feature-complete, just in final testing. Also working on automated README updates so agents keep their own docs current without being told.
I'm a solo dev but every PR is human-AI collaboration - the agents help build and maintain themselves. 105 sessions in and the framework is basically its own best test case.
[link] [comments]