wanman  by chekusu

Agent matrix runtime for autonomous multi-agent workflows

Created 5 days ago

New!

388 stars

Top 73.6% on SourcePulse

GitHubView on GitHub
Project Summary

Summary

Wanman is an open-source agent matrix runtime designed for orchestrating autonomous multi-agent workflows, task execution, and artifact management. It targets engineers and power users who need to coordinate AI agents locally, allowing human operators to transition into an observer role while complex agent networks collaborate on tasks. The primary benefit is enabling sophisticated, automated AI workflows on a user's machine.

How It Works

The framework employs an agent matrix architecture, inspired by "one-man trains," where a human user oversees a network of autonomous agents. A JSON-RPC supervisor coordinates multiple agents (e.g., CEO, dev, devops) that run as isolated local CLI subprocesses, utilizing either Claude Code or Codex. Agents communicate via an asynchronous message bus with steer/follow-up priorities. Each agent operates within its own dedicated worktree and $HOME directory, ensuring isolation and preventing interference with the host environment. This CLI-first design emphasizes scriptability, observability, and reproducibility.

Quick Start & Requirements

  • Prerequisites: Node.js 20+, pnpm 9+, Git, and a logged-in Claude Code or Codex CLI.
  • Installation: Clone the repository (git clone git@github.com:chekusu/wanman.git wanman.dev), navigate into the directory, and run pnpm install followed by pnpm build.
  • Execution: The primary command to start an agent matrix on a Git repository is pnpm --filter @wanman/cli exec wanman takeover /path/to/any/git/repo. A single-file CLI bundle is also available via pnpm --filter @wanman/cli standalone. Refer to docs/quickstart.md for a detailed walkthrough.

Highlighted Details

  • Coordinates diverse agent roles (CEO, dev, devops, marketing, feedback) through a prioritized asynchronous message bus.
  • Manages agent lifecycles, spawning them as isolated Claude Code or Codex CLI subprocesses.
  • Ensures agent isolation via per-agent worktrees and per-agent $HOME directories, preventing host environment mutation.
  • Features a CLI-first approach for scriptable, observable, and reproducible agent operations.

Maintenance & Community

No specific details regarding notable contributors, sponsorships, or community channels (like Discord/Slack) were found in the provided README.

Licensing & Compatibility

  • License: Apache-2.0.
  • Compatibility: The Apache-2.0 license is generally permissive for commercial use and integration with closed-source projects. Users must provide their own authentication for the required LLM CLIs.

Limitations & Caveats

The runtime strictly depends on having the Claude Code or Codex CLI installed and authenticated. The idle_cached agent lifecycle is exclusive to Claude. Advanced features like large-scale concurrency and dynamic role extraction are noted as capabilities of the hosted wanman.ai service rather than the core open-source offering.

Health Check
Last Commit

3 days ago

Responsiveness

Inactive

Pull Requests (30d)
2
Issues (30d)
4
Star History
389 stars in the last 5 days

Explore Similar Projects

Starred by Andrew Ng Andrew Ng(Founder of DeepLearning.AI; Cofounder of Coursera; Professor at Stanford), Jack Lukic Jack Lukic(Author of Semantic UI), and
5 more.

ag2 by ag2ai

0.4%
4k
AgentOS for building AI agents and facilitating multi-agent cooperation
Created 1 year ago
Updated 4 hours ago
Starred by Gagan Bansal Gagan Bansal(Coauthor of AutoGen; Research Scientist at Microsoft Research), Elvis Saravia Elvis Saravia(Founder of DAIR.AI), and
1 more.

agent-framework by microsoft

2.2%
10k
AI agent and multi-agent workflow framework
Created 1 year ago
Updated 4 hours ago
Feedback? Help us improve.