Overview

Three days after Manus launched behind invite codes in March 2025, a group from the MetaGPT team put a working open implementation on GitHub. Xinbing Liang and Jinyu Xiang led it, with Zhaoyang Yu, Jiayi Zhang and Sirui Hong, and the prototype took about three hours. The repository now lives under the FoundationAgents organisation, carries 58,100 stars and 10,100 forks, and was last pushed on 22 August 2026. It is not archived and the issue tracker is live. The licence is MIT, full stop, with none of the user-count thresholds or acceptable-use appendices that show up in most Chinese open releases.

Architecturally it is deliberately small. Python 3.12, install with uv or conda, configure models in a config.toml that takes an OpenAI-compatible base_url and key. That single design decision is what makes it useful in China: point it at DeepSeek, at Qwen through DashScope, or at a vLLM server you run yourself, and the agent never touches a foreign endpoint. Browser Use CLI 3.0 starts as the default MCP server, Playwright covers browser automation and BrowserGym is there if you want to run web benchmarks. There is a planning agent, a tool-calling agent and a data analysis path with chart generation. A sibling project, OpenManus-RL, built with researchers at UIUC, works on reinforcement learning based tuning of agent policies.

Two honest caveats. The version tags are misleading: v0.1.0 through v0.3.0 were all cut on 10 April 2025 and nothing has been tagged since, so production users pin a commit hash and own their own testing. And this is a framework, so you supply the sandboxing. An agent with shell access on an unisolated host is a security incident waiting for a prompt injection, and the repo does not solve that for you. Neither point undermines the recommendation. If you want an agent loop you can read end to end, modify and run entirely on domestic infrastructure, this is the reference codebase, and the licence means nobody can change the terms on you later.

Key Features

✓ MIT licence with no user cap and no field-of-use restriction, which is a materially different legal position from the community licences most Chinese AI releases ship under

✓ Model agnostic via a config.toml that accepts any OpenAI-compatible base_url, so DeepSeek, Qwen through DashScope or a local vLLM endpoint all drop in without code changes

✓ Starts Browser Use CLI 3.0 as a default MCP server, with Playwright and BrowserGym available for web-native and benchmark tasks

✓ Tool-calling agent, planning agent and a data analysis path with chart generation, each readable in a few hundred lines rather than buried under abstraction

✓ OpenManus-RL, a sibling repository built with UIUC researchers, for reinforcement learning based tuning of agent policies

✓ Python 3.12, installable with uv or conda, runs on a laptop with no hosted service in the loop

My Take

MIT, no user cap, no field-of-use clause. That one fact matters more than the three-hour build story this repo is famous for, because a great deal of what China ships as "open" arrives wrapped in a community licence you need to read twice. The MetaGPT team kept it alive as well, with a push on 22 August 2026 against 58,100 stars and 10,100 forks. Point config.toml at a DeepSeek or local Qwen endpoint and the whole agent loop runs without a packet leaving your network, which is the real reason to choose it over anything hosted.