Persistent Agent Workspace

AI agents that remember, collaborate, and never start from zero.

---

The Problem

When the task gets complex, you become the coordinator — copying context between tools, dispatching work manually, keeping track of who's doing what. You are the glue holding your AI workflow together.

Multi-agent frameworks tried to fix this. But they isolate every agent in its own sandbox. Separate filesystems. Separate context. One agent can't see what another just built. Nobody picks up where a teammate left off.

That's not a team. That's a room full of people wearing headphones.

What AIPass Does

AIPass is a CLI-native scaffold that adds persistent memory, identity, and coordination to your AI agents. You bring your project — AIPass adds the agent layer on top. No UI, no dashboard. You work in your terminal.

pip install aipass
mkdir my-project && cd my-project
aipass init run

A guided setup creates your project, your first agent, and opens a terminal where that agent is already running. Say "hi" — it knows who it is. Come back tomorrow — it remembers.

This is the base framework. It gives your agents the infrastructure to persist, communicate, and organize — everything else you build on top.

Here's what lands in your project:

my-project/
├── .aipass/                # Project config + prompts
├── .claude/                # Hooks (injected automatically)
├── src/my_project/
│   └── my_agent/
│       ├── .trinity/       # Identity + memory (3 JSON files)
│       ├── .ai_mail.local/ # Local mailbox
│       ├── apps/           # Your agent's code
│       └── README.md       # Domain knowledge
├── CLAUDE.md               # Project instructions
└── MY-PROJECT_REGISTRY.json

Everything is plain files. No daemon, no hidden state. Delete the directory and it's gone.

Start with one agent. Add more when you need them:

aipass init agent my-agent            # Full agent: apps, mail, memory, identity

What makes this different:

• Agents are persistent. They remember across sessions. Expertise develops over time. Nobody starts from zero.
• Bring your own project. AIPass adds agent infrastructure to whatever you're building. It's a scaffold, not a product — you shape it.
• Everything is local. Memory is JSON files. Communication is local mailbox files. No cloud, no external APIs.
• Shared workspace. All agents work on the same filesystem, same project, same time. No sandboxes.
• One command for everything. AIPass ships with drone, a CLI router — drone @agent command reaches any agent. Learn it once, use it everywhere.

Runs on your existing CLI subscription. Claude Pro/Max or Codex — AIPass uses the same CLI binary you already run. No extra API keys, no extra costs for core functionality.

---

Quick Start

Your own project

pip install aipass

mkdir my-project && cd my-project
aipass init run                       # Guided setup — project, first agent, terminal handoff

That's it. Your agent has identity, memory, a mailbox, and access to every AIPass service — planning, quality audits, dispatch, real-time monitoring. All through drone @branch command.

aipass init                           # Just the scaffold (no guided setup)
aipass init agent my_agent            # Add another agent
aipass doctor                         # Check system health

Need help? Ask in Discussions or file feedback — both take 30 seconds.

Explore the full framework

Clone the repo to see all 13 agents working together — the reference implementation:

git clone https://github.com/AIOSAI/AIPass.git
cd AIPass
./setup.sh                            # Creates venv, installs, bootstraps 13 agents

cd src/aipass/devpulse
claude                                # Talk to the orchestrator

# Things you can do:
aipass doctor                            # Check system health
drone @seedgo audit aipass               # Run automated quality checks across all agents
drone @flow create . "Add user auth"     # Create a work plan
drone @ai_mail dispatch @agent "Sub" "Body"  # Send task + wake an agent

---

How It Works

One agent: Run aipass init run and in 5 minutes you have a project with an agent that reads .trinity/ on startup and picks up where it left off. Memory starts as plain JSON files — no setup required. When they fill up, older entries automatically archive into ChromaDB for long-term search. Nothing is lost.

A team: When one agent isn't enough, every agent shares the same structure:

src/my-project/<agent>/
├── .trinity/           # Identity + memory (persists across sessions)
├── .ai_mail.local/     # Mailbox (receives tasks, sends results)
├── apps/               # Entry point → modules → handlers
└── README.md           # Domain knowledge (the agent reads this on startup)

Identical layout everywhere. If you know one agent, you know all of them. drone is the single command that routes to any agent:

drone @branch command [args]    # Every agent, every task. Drone handles routing.

drone @seedgo audit my_project               # Run quality checks on everything
drone @flow create . "Refactor auth module"  # Create a work plan
drone @ai_mail dispatch @agent "Archive old sessions" "Find sessions older than 30 days"

Two ways to use AIPass:

• Your own project: aipass init run sets up a new project with your first agent. Add more agents as you need them. Your first agent is the orchestrator — it coordinates the others.
• The full framework: Clone the repo to work with all 13 core agents. Talk to devpulse (the orchestrator), dispatch work across specialists. Agents work in parallel and report back.

---

The Reference Implementation

AIPass ships with 13 core agents that maintain and develop the framework itself — proving the architecture works at scale. You don't need any of these to use AIPass in your own project. They're here as examples and as services your project can call.

devpulse (orchestrator)
   ├── aipass   — concierge + onboarding (aipass init, doctor, profile)
   ├── drone    — command routing + @agent resolution
   ├── seedgo   — automated quality standards
   ├── prax     — real-time monitoring across all agents
   ├── ai_mail  — agent-to-agent communication + task dispatch
   ├── flow     — plan lifecycle, templates, auto-archival
   ├── spawn    — creates new agents anywhere on your filesystem
   ├── hooks    — hook engine, sound control, per-project config
   ├── memory   — automatic archival, ChromaDB, semantic search
   ├── api      — LLM access layer (OpenRouter, multi-provider)
   ├── trigger  — event-driven automation + self-healing
   └── cli      — terminal formatting and rich output

These agents work on the same filesystem, same project, same time — no sandboxes, no worktrees. This is the pattern your projects inherit.

Agent details

You interact with one: devpulse — the orchestrator. You talk to it, it coordinates everyone else.

Core infrastructure — how agents connect:

Agent	Role
aipass	Concierge — aipass init, doctor, profile, onboarding
drone	Routes drone @branch command to the right agent
ai_mail	Agent-to-agent messaging and task dispatch
memory	Memory lifecycle — automatic archival, ChromaDB vectors, semantic search
api	LLM access layer — multi-provider routing (OpenRouter)
spawn	Creates new agents from templates

Quality and operations — how the system stays healthy:

Agent	Role
seedgo	Automated quality standards, enforced across all agents
prax	Real-time monitoring, logs, dashboards
flow	Plan lifecycle — multiple template types, auto-archival, vector verification
hooks	Hook engine — per-project config, sound control, event dispatch
trigger	Event-driven automation + self-healing
cli	Terminal formatting and rich output

---

CLI Support

AIPass is built and tested with Claude Code on Linux/WSL.

CLI	Autonomous Mode	Status
Claude Code	claude -p "prompt" --permission-mode bypassPermissions	Fully tested
Codex	codex exec "prompt" --dangerously-bypass-approvals-and-sandbox	Experimental

setup.sh auto-detects which CLIs are installed and configures hooks for each.

---

Project Status

Beta. Actively developed by a solo developer working with the AI agents themselves — every PR, every test, every fix is human-AI collaboration.

Metric	Value
Version
Agents	13 core + user-created
Quality	Automated standards enforced across every agent
Coverage	— 75% minimum, CI-gated
Tests	Extensive — every agent ships its own suite

Each agent documents its own operational status in its branch README — what works, what doesn't, and why.

---

Requirements

• Python 3.10+
• Claude Code
• Linux, macOS, or WSL (all CI-tested)
• sudo access optional (for /usr/local/bin symlinks — falls back to ~/.local/bin without sudo)
• API keys optional (OpenRouter/OpenAI — for optional add-on agents)

Roadmap

These items have partial work done and are under ongoing testing:

• macOS support — CI green, full test suite passing (#360)
• Windows native — CI green, full test suite passing
• Codex CLI — hooks and AGENTS.md wired, needs end-to-end testing
• Fork contributor workflow — improved error handling for fork-based PRs (#329)

---

Uninstall

Remove AIPass from a project

AIPass stores everything locally in your project directory. To remove it:

# Remove AIPass files from your project
rm -rf .aipass/ .claude/ .ai_mail.local/ hooks/ src/
rm -f CLAUDE.md AGENTS.md *_REGISTRY.json .gitignore

# If you installed via pip
pip uninstall aipass

No cloud accounts, no external services, no cleanup beyond your local filesystem.

Remove a single agent

Use spawn's delete command to cleanly archive and deregister:

drone @spawn delete @agent_name

This archives the agent's directory and removes it from the registry.

Subscriptions & Compliance

Use your existing subscription

AIPass runs on your existing CLI subscription — Claude Pro/Max or Codex. No API keys required for core functionality. No extra costs beyond your existing subscription.

This works because AIPass runs each CLI as an official subprocess — the same binary you'd run yourself in a terminal. It doesn't extract credentials, proxy API calls, or intercept tokens. Your subscription stays within the provider's infrastructure at all times.

What AIPass does NOT do

• Extract or redirect subscription OAuth tokens
• Intercept CLI-to-provider communication
• Bypass rate limits or prompt caching
• Impersonate official CLI clients

Claude Code is proprietary but officially supports hooks and subprocess usage. Codex CLI is open source (Apache 2.0).

API keys are only needed for optional add-on agents (OpenRouter/OpenAI). For server/automated deployments, API key authentication is recommended per Anthropic's guidance.