A solo developer has shipped AIPass v2.4.0, a CLI-native framework that adds persistent memory, identity, and inter-agent communication to AI workflows. The tool runs entirely local—no cloud dependencies, no extra API keys—and plugs into your existing Claude Code or Codex subscription. Installation takes one pip command and five minutes of guided setup.

Why Current Multi-Agent Setups Fail

The project's documentation cuts straight to the problem: most multi-agent frameworks isolate every agent in its own sandbox with separate filesystems and context windows. When Agent A finishes a task, Agent B has no visibility into what was built or where work left off. Developers become the coordinators—manually copying context between tools and tracking who's doing what. AIPass's creator calls this 'a room full of people wearing headphones.' AIPass attacks this by giving every agent three persistent components stored as plain JSON files: a .trinity/ directory holding identity and memory, a .ai_mail.local/ mailbox for receiving tasks and dispatching results, and apps/ containing the agent's code. Everything lives in the project directory under src/my_project//. No daemon runs in the background—delete the folder and it's gone.

The Core Architecture

The framework ships with 13 pre-built agents that maintain and develop itself, serving as both examples and services you can call from your own projects. devpulse acts as the orchestrator—you talk to it, it coordinates specialists. drone handles routing commands via @agent syntax across any agent in the system. seedgo runs 36 automated quality checks against all agents. ai_mail manages inter-agent messaging while memory automatically archives filled JSON logs into ChromaDB for semantic search. The CLI router is central to the experience: drone @seedgo audit my-project triggers quality checks, drone @flow create . "Refactor auth" generates a work plan, and drone @ai_mail dispatch @agent "Task" "Body" wakes a sleeping agent with new work. Every command follows the same pattern regardless of which agent handles it.

Local-First by Design

AIPass explicitly avoids cloud infrastructure for core functionality. Memory lives in JSON files that you can inspect, edit, or delete directly. Communication uses local mailbox files rather than network calls. The only optional external dependencies are OpenRouter or OpenAI API keys—needed only if you're building add-on agents that require them. The framework runs on Claude Code (fully tested on Linux/WSL) and Codex CLI (experimental status). setup.sh auto-detects installed CLIs and wires hooks for each. With 8,400+ tests passing across all agents and 600+ PRs merged through human-AI collaboration, the architecture has real production miles behind it.

Key Takeaways

  • AIPass v2.4.0 is in beta but actively developed with extensive test coverage
  • All agent state persists as local files—identity, memory, mailbox—no cloud required for core features
  • Shared workspace model lets agents see same filesystem and pick up where teammates left off
  • Drone CLI router provides consistent command syntax across all 13 built-in agents
  • Runs on existing Claude Code/Codex subscriptions without extracting OAuth tokens or proxying API calls

The Bottom Line

This is the infrastructure layer multi-agent frameworks have been missing. No more starting from zero every session, no more manual context handoffs between isolated sandboxes. If you're running AI agents in your workflow and feeling the coordination pain, AIPass is worth a look—just remember it's beta software from a solo dev shipping via human-AI collaboration.