metagpt

MetaGPT is a Python multi-agent "software company" framework with the metagpt CLI. It came into the catalog from the agent framework sweep for issue #64; see research/frameworks.md for the frameworks that got no catalog line. It is collected only, with no parser: MetaGPT keeps no transcripts under the home.

Source and evidence

FoundationAgents/MetaGPT (geekan/MetaGPT redirects there), MIT (setup.py:107), open source, commit 11cdf466d042aece04fc6cfd13b28e1a70341b1f (2026-01-21). All claims are from source.

MetaGPT runs shell commands on the host: the Terminal tool spawns a persistent shell (terminal.py:47-52).

Observed

• No real install: nothing was installed or run.

Source reviews

• 2026-10-19: HEAD of FoundationAgents/MetaGPT is still 11cdf46, re-read from source for permission, model, past-use and sync records. It added the run log and the model key (Per-user storage); there is still nothing to parse.

Collection

Per-user storage

Path (home-relative, all OSes)	Holds	Default or opt-in
.metagpt/config2.yaml (Path.home())	LLM configuration including api_key (const.py:39, config2.py:115-118, software_company.py:130-148, README.md:70-73)	read by every run; written by metagpt --init-config
.metagpt/config2.bak	the previous config, renamed on re-init (software_company.py:146-148)	on a second --init-config

CONFIG_ROOT / "config2.yaml" is the only home path in the package (grep for CONFIG_ROOT, Path.home, expanduser, platformdirs). The model is llm.model in that file (llm_config.py:64).

Nothing records a permission decision: the Terminal tool asks nothing before a command (no approve or confirm in terminal.py), and the one question RoleZero asks, whether to continue past the action limit, goes to the human through the environment and is not stored (role_zero.py:328-335). Nothing syncs or relays a session (not found in metagpt/ at 11cdf46, grepped for share_url, upload_session and sync_session), and there is no session store to delete from.

Scratch directories

<METAGPT_ROOT>/workspace (Workspace) is the agent's own output directory for generated projects; nothing under the home records it, so no catalog line or discovery source can find it.

Workspace

Generated projects and serialized team state go to <METAGPT_ROOT>/workspace and workspace/storage/team/team.json, where METAGPT_ROOT is $METAGPT_PROJECT_ROOT, else the package parent if it holds .git (an editable install from a clone), else the working directory (const.py:19-41, const.py:58, workspace_config.py:18-37, team.py:59-65). team.json is written when Team.run ends in an exception or Ctrl-C (common.py:675-684). Every run also writes a DEBUG-level log, logs/<YYYYMMDD>.txt (or <name>_<YYYYMMDD>.txt), under the same METAGPT_ROOT (logs.py:40-51). These are relative to the working directory or the clone, not the home, so they have no catalog line; workspace is also too generic a name to anchor a glob on. The git tool reads, but does not write, ~/.git-credentials (git.py:69-71).

Docker

No named volume.

Credentials

• .metagpt/config2.yaml and .metagpt/config2.bak: flagged.

Exclusions

None.

Project-local files

The workspace/ tree in Workspace, under no MetaGPT-specific name.

Where the project path is recorded

Nowhere under the home.

Parsing

MetaGPT is collected only: the collector gathers its configuration, and there is nothing to parse under the home. The team state and the run logs are written under METAGPT_ROOT, under no MetaGPT-specific name (Workspace), so no catalog line reaches them and a parser would have no collected file to read.

Confidence

• High for every path (source).
• Not determined: the workspace location when the working directory is the home (then ~/workspace, too generic for a glob).