Gitoma

Appearance

GitomaAn autonomous agent that improves your GitHub repo.

Analyzes. Plans. Commits. Opens a pull request. Reviews its own work. Local-first by design — your code, diffs, and secrets never leave your machine.

Get started
View on GitHub
Gitoma

Self-correcting guard stack

14 composable guards (G1–G14) catch specific classes of LLM patch slop — broken syntax, dropped functions, hallucinated frameworks/URLs/configs, README destruction. Plus two planner-time post-processors (Layer-A real-bug task synthesis, Layer-B README banishment) and an opt-in scalar Ψ-lite quality gate. Every guard was added in response to a specific live bench failure.

Local-first LLM

Plans and writes patches with a model running on your machine — LM Studio, Ollama, or any OpenAI-compatible endpoint. No SaaS, no egress, no surprises on the bill.

Grounded by Occam Observer

Optional integration with the companion Occam Observer gateway. Per-run fingerprint snapshot (declared deps, frameworks, manifests) is injected into the planner prompt as ground truth and consumed by the worker-side content-grounding guards. Pre-empts hallucinations at plan time.

Full pipeline, resumable

Analyze → Plan → Execute → PR → Self-Review. Every phase persists state; a crash or kill is one `--resume` away from picking up at the last committed subtask.

Live web cockpit

A read-only dashboard streams every phase over WebSocket, with keyboard shortcuts, command palette, and SSE live logs — no build step, no external deps.

REST + MCP

Everything the CLI does is a Bearer-protected REST endpoint. A built-in MCP server also exposes GitHub context + write tools to Claude Desktop and any MCP client.

Hardened by default

Constant-time auth, process-group isolation, credential redaction, WCAG AA contrast, CSP headers, focus trap in dialogs, prefers-reduced-motion. Audited surface end-to-end.

Structured observability

Each run writes a JSONL trace under `~/.gitoma/logs/`. `gitoma logs --follow` tails it live; the cockpit renders phases, heartbeat, and orphan detection without polling the agent.

From zero to a pull request in three commands ​

bash
pipx install gitoma
gitoma config set GITHUB_TOKEN=ghp_your_fine_grained_token
gitoma run https://github.com/<owner>/<repo>

Open the cockpit while it runs:

bash
gitoma serve &
open http://localhost:8000

Who it's for ​

  • Engineers who maintain a portfolio of repositories and want a consistent baseline — tests, CI, docs, dependencies — without writing the same PR fifty times.
  • Teams that need a trusted agent on a VPN to apply low-risk improvements to internal tools, with every write on an ergonomic audit trail.
  • Researchers exploring autonomous coding agents: every prompt, every LLM I/O, every git op is traceable, inspectable, and replayable offline.

What it does not do ​

  • It does not ship your code to a third-party service. The LLM is yours.
  • It does not merge PRs on its own. You keep the decision; the agent does the boring part.
  • It does not touch protected paths. .git/, .github/workflows/, .env* are denylisted at the patcher level — even if the LLM is prompt-injected through the repo content, those are off-limits.

Released under the MIT License. Local-first by design. No telemetry.

© 2026 Fabrizio Salmi