reviewer

A CLI that reviews a GitHub pull request using a local LLM — no API key, no cloud billing, runs entirely on your machine via Ollama.

reviewer review --pull-request-url https://github.com/owner/repo/pull/123

How it works

• Fetches the PR diff using pure git over SSH — no GitHub API token needed. It uses your existing GitHub SSH key to fetch refs/pull/<N>/head and refs/pull/<N>/merge directly, falling back to diffing against the repo's default branch if the merge ref isn't available.
• Sends the diff to a local model running in Ollama, in two stages: a free-form review pass (so the model can actually reason about the diff) followed by a cheap extraction pass into structured findings.
• Prints the files reviewed, what was checked, the model's reasoning, and any findings — all to your terminal.

Because everything runs locally, the diff and review never leave your machine, and there's no per-review cost.

Requirements

• Node.js >= 18
• git, with SSH access to GitHub already configured (ssh -T git@github.com should work)
• Ollama — no system install required; see setup below

Setup

npm install
npm run build
npm link          # makes the `reviewer` command available globally

Install Ollama (no sudo required)

mkdir -p ~/ollama-local
curl -sL https://github.com/ollama/ollama/releases/latest/download/ollama-linux-amd64.tar.zst -o /tmp/ollama.tar.zst
tar --zstd -C ~/ollama-local -xf /tmp/ollama.tar.zst

Pull a model

~/ollama-local/bin/ollama serve &   # one-time, to pull the model
~/ollama-local/bin/ollama pull qwen2.5-coder:3b

Usage

reviewer ollama start     # start the local model server (background, idempotent)
reviewer ollama stop      # stop it

reviewer review --pull-request-url <url> [--base-ref <branch>]

• --base-ref <branch> — diff against this branch instead of auto-detecting the PR's base. Only needed if a PR targets a non-default branch and the automatic detection picks the wrong one (you'll see a warning when this happens).

Leave the Ollama server running between reviews — the first request after starting it pays a one-time model-load cost (tens of seconds); every request after that is fast.

Configuration

All optional, via environment variables:

Variable	Default	Purpose
OLLAMA_MODEL	qwen2.5-coder:3b	Which pulled model to use for reviews
OLLAMA_HOST	http://127.0.0.1:11434	Where the Ollama server is listening
OLLAMA_BIN	~/ollama-local/bin/ollama	Path to the Ollama binary, used by reviewer ollama start

Example output

Reviewing owner/repo#123 with a local model (this may take a while on CPU)...

Files reviewed (1):
  src/auth.js  +12 -3

Checked for:
  - Correctness bugs (logic errors, off-by-one, null/undefined handling, race conditions)
  - Security issues (injection, unsafe input handling, secrets)
  - Clear maintainability problems introduced by this diff

Review notes:
The diff adds a checkPassword function that compares passwords with ==
instead of a constant-time comparison, which is a timing-attack risk for
auth code.

src/auth.js:14 [WARNING] Password comparison uses == instead of a constant-time check.

1 finding (1 warning)

Known limitations

• GitHub only for now — no Bitbucket/GitLab support.
• Diff-only context — the model only sees the code diff, not the PR title, description, or discussion, since that would require the GitHub API (and a token) rather than plain git.
• Model quality ceiling — qwen2.5-coder:3b is small and fast but noticeably weaker than a hosted model like Claude; it can miss subtler bugs and occasionally contradicts itself between its reasoning and its conclusion. Swap in a larger model (e.g. qwen2.5-coder:7b) via OLLAMA_MODEL for better quality at the cost of speed and memory.
• CPU inference is the bottleneck — review time is dominated by token generation speed, not the CLI itself. GPU offload was tested on integrated graphics and made no measurable difference; the effective levers are model size and response length.