GnosisLab

Umbrella project — a composed research-automation distribution. Subsystems live in subfolders:

• ml-scientist/ — the scientific-experiment MCP stack (hypothesis → experiment → evidence → conclusion): five MCP servers, sealed execution, invariant suite, observability GUIs.
• openshell/ — (planned) NVIDIA OpenShell launch-within-shell subsystem.

Host requirements

• Linux — trial sandboxing uses bubblewrap mount namespaces.

• Python ≥ 3.10 and uv (curl -LsSf https://astral.sh/uv/install.sh | sh), or any pip for the PyPI path below.

• bubblewrap + strace — required by the sealed executor; the launchers refuse to start without them:

  sudo apt install bubblewrap strace    # Debian/Ubuntu
  

• Ports 38050–38091 free (five servers × MCP + GUI; see ml-scientist/ports.env to move them).

Install

From source (this repository)

git clone <this-repo> gnosislab && cd gnosislab
uv sync          # one root env for the whole distribution

ml-scientist/ is also independently runnable — cd ml-scientist && uv sync gives it its own env.

From PyPI

pip install gnosislab        # or, better for CLI use:
uv tool install gnosislab    # isolated env, commands on PATH

This installs the five MCP servers as console commands (ml-episteme-mcp, ml-anamnesis-mcp, ml-zetesis-mcp, ml-arete-mcp, ml-agora-mcp). The gnosislab lifecycle script is repo-only for now (bash, not shipped in the wheel) — a gnosislab CLI entry point is planned.

Run

gnosislab is a symlink at the repo root into ml-scientist/:

./gnosislab start all     # upstreams first, agora last
./gnosislab status        # pid/ports/health per server
./gnosislab logs arete
./gnosislab stop all

(./labloop still works — transition symlink, prints a rename notice.)

To launch from anywhere, put it on PATH:

ln -sf "$PWD/gnosislab" ~/.local/bin/gnosislab
gnosislab start all

Per-server observability GUIs: anamnesis :38091, episteme :38081, zetesis :38071, arete :38061, agora :38051 (lab-level status hub). Each server also responds to ml-<name>-mcp run directly in the foreground via ml-scientist/run-ml-<name>.sh for debugging.

Config: ml-scientist/ml-*.toml (per server) + ports.env. Runtime state lives outside the repo in ~/.ml-<name>/ (databases, logs, pidfiles).

Test

uv run pytest -q     # full suite (1289 tests)

Publish

uv build
uv publish --publish-url https://test.pypi.org/legacy/   # TestPyPI
uv publish                                               # PyPI