Wiki article · Published October 2, 2026 · Living document

Star Lab

A free local agent research lab — open source, MIT

Serious agent research does not require serious infrastructure. Star Lab runs agent experiments on ordinary hardware — no cloud bill, no data leaving the bench. The workshop where doctrines get tested before they get published.

Inspect the repository Read the references
Golden Gate Book
Wiki · MMXXVI

Star Lab is a free, local-first laboratory for running agent experiments, built by Marcus Richards with Grok and published as open source under the MIT license. It provides a unified command-line interface, a health-scoring doctor, a mission-control dashboard, and a set of offline-first research modules on ordinary hardware.

It began as the control plane for a fair head-to-head comparison between coding agents: durable memory, safety under always-approve operation, project roots, CLI tooling, and a doctor that scores system health rather than merely reporting it. Grok Star Lab extends that foundation with experiment tracking, a model gym, and a showroom, all offline-first.

Background

Agent research in 2026 typically assumes cloud infrastructure: hosted models, remote experiment tracking, and usage bills that scale with curiosity. Star Lab takes the opposite position, that a capable laboratory can run on the machine already on the desk. The project was commissioned and directed by Richards and built with Grok, and it doubles as the workshop where his other doctrines, leased authority, claim provenance, token governance, are tested before publication.

Design

The entry point is a unified lab command-line interface. lab status gives one-screen vitals, lab doctor runs a full health check that scores the system and exits zero only when operational, and lab dash opens the mission-control dashboard. Installation is idempotent and never touches the machine's safety hooks; mutable lab state lives under the user's home directory.

Research modules are all offline-first:

  • Mission Control and Observatory, dashboards and status embeds from doctor output.
  • Experiment Forge, experiment tracking on a SQLite and Markdown ledger.
  • Model Gym, local model smoke tests and evaluations.
  • Agent Arena, scripted agent pipelines.
  • Knowledge Crucible, full-text indexed local knowledge.
  • Sandbox Range, seatbelt-profiled execution sandboxes.
  • Ship Bay, local ship checks with a showroom inbox.

Limitations

What is not claimed. Star Lab is a laboratory instrument, not a product. Its documentation carries machine-specific paths from the bench it was built on, and its experiment results describe local runs, not benchmarks. It assumes a cooperative local operator; it is not a sandbox against a hostile one.

Development history

Star Lab was created in August 2026 and published under the MIT license. It is under active development as of October 2026, with continuous integration green on the main branch.

See also

  • Interlock, the agent safety substrate whose doctrines are tested in the lab
  • AEGIS, token governance for AI coding work

References

  1. Star Lab source repository. github.com/marsojuji-cmyk/star-lab (public, MIT).
Inspect the repository Back to the workshop