OpenManus
An open, general-purpose agent framework from FoundationAgents — MIT, 57.6k★ and 10k forks off 526 commits, which is the wiki’s steepest star-to-commit ratio and the first thing to explain about it. Its README opens: “No fortress, purely open ground. OpenManus is Coming.”
Built as a jailbreak of a closed product
Manus shipped as a proprietary general agent gated behind an invite code. Core authors Xinbin Liang and Jinyu Xiang, with contributors from the MetaGPT community, had a working prototype in three hours and put the pitch in one line: “OpenManus can achieve any idea without an Invite Code.”
The 57.6k stars are for that, not for 526 commits of engineering. What the repo demonstrates is how thin the moat around a general-purpose agent product turned out to be — the scarce thing was access, not architecture. Compare the corpus’s other open-vs-closed instances (openclaw, hermes-agent): those replaced closed harnesses; this one replaced a closed waitlist.
What it actually is
A Python 3.12 framework (conda or uv install, config.toml for keys) that takes an idea from the
terminal and executes it. Three modes ship: the general OpenManus Agent, a DataAnalysis agent
for visualization work, and an MCP tool version. Capabilities named:
browser automation via Playwright, data analysis, computer vision, and further tools through MCP.
Default model is GPT-4o with base_url/key overrides for any other provider, so it sits on the
provider-agnostic side of the agentic-coding-harness pattern.
The README is thin exactly where this spoke cares most: the agent loop is not documented. Multi-agent support and tool integration are asserted; how the agent decides, retries, or verifies isn’t described. 301 open issues and 192 open PRs against 526 commits is its own signal about maintenance load.
The sibling that argues with this wiki
OpenManus-RL, a collaboration with UIUC researchers, tunes LLM agents with reinforcement learning (GRPO). That is a different theory of where agent competence comes from than the one this spoke has been accumulating evidence for: don’t scaffold the model, train it. The corpus has no other source on RL-tuned agents, and this one is a pointer rather than a study — see synthesis, where it’s filed as a standing tension rather than a settled contradiction.
Related
foundation-agents · agentic-tooling · agentic-coding-harness · model-context-protocol · openclaw · hermes-agent