How to run AI code review locally with Ollama.

Local AI code review is the practice of running a language model on your own machine to review a git diff, so the change never goes to a hosted API. With CommitBrief and Ollama it takes four commands, needs no API key, and costs nothing per review beyond the electricity.

In short: install Ollama, pull a coder model, run commitbrief setup --local, then commitbrief --staged. The local model is weaker than a frontier cloud model, so treat the output as a first pass, not a verdict.

This is the how-to. For the argument about why a diff should stay on the machine, see the air-gapped Ollama post. For how this setup compares with hosted reviewers, see AI code review tools: CLI vs GitHub App vs IDE.

What you need

Pick the model size your hardware can serve. This post does not give timings: they depend entirely on your CPU, GPU, and model, and a number from someone else’s laptop would mislead you.

Step 1: pull a model

ollama pull qwen2.5-coder:14b

Step 2: point CommitBrief at it

commitbrief setup --local

In the wizard, choose ollama. It asks for a base URL (default http://localhost:11434), then queries <base_url>/api/tags and lists the models you have pulled. Pick qwen2.5-coder:14b. The wizard finishes with a connection test, and a failure aborts it, so a stopped Ollama server shows up here and not on your first review.

--local writes the config to <repo>/.commitbrief/config.yml and adds the .commitbrief/ directory to .gitignore. Without it the config goes to ~/.commitbrief/config.yml and applies to every repo. The result is this:

provider: ollama
providers:
  ollama:
    base_url: http://localhost:11434
    model: qwen2.5-coder:14b

You can also skip the wizard for a single run:

commitbrief --provider ollama --model qwen2.5-coder:14b --staged

To check connectivity without changing any state:

commitbrief providers test ollama

If Ollama runs on another machine, set providers.ollama.base_url. If OLLAMA_HOST is set in your environment, it always overrides base_url, so check your shell profile when the review goes to an unexpected host. That machine then receives your diff, so the privacy argument now depends on who controls it.

Step 3: review a change

git add .
commitbrief --staged

--staged is the default scope, so a bare commitbrief does the same. The review prints as cards. Add --json for the structured findings, or --fail-on=critical to get a non-zero exit when a critical finding survives. Review scopes covers the other ranges, such as commitbrief diff main...HEAD.

Two properties of the Ollama provider are worth knowing:

• No cost preflight. Ollama reports zero pricing, so the cost check never prompts and the reported cost is 0. Token counts are still shown. See Safety, cost, and cache.
• Cached re-runs. Reviewing the same diff again with the same settings is served from the local response cache and does not call the model.

Step 4: tune it when it is slow or noisy

Timeouts. Ollama’s HTTP client gives up after 5 minutes. A large diff on a mid-size model can exceed that. --timeout raises the cap, and the value also replaces the provider’s own limit:

commitbrief --staged --timeout 15m
commitbrief config set review.timeout 15m   # make it the default

See Timeouts for the resolution order.

Diff size. commitbrief dry-run --staged --verbose runs the same pipeline without a provider call and reports estimated tokens. Use it to see whether a diff is near your model’s context window before you wait on it. If it is, review a smaller range.

Rules. commitbrief init writes a COMMITBRIEF.md that is sent as the system prompt. The rules apply to local models like any other provider. See Review rules.

What this setup does not give you

A local 14B-class model is not a frontier cloud model, and a review from one is not equivalent. Smaller models can fail to return valid JSON. CommitBrief retries once and then degrades to plain markdown. When that happens --fail-on is skipped by design, because failing CI on a flaky model invocation is worse than showing you the text. If you gate a pipeline on a local model, watch for the --fail-on skipped line on stderr. The troubleshooting page lists the probable causes and fixes.

Intent-level problems, such as a function that answers the wrong question, remain a human reviewer’s job with any model.

One action

Run commitbrief providers test ollama now. If it passes, stage your next change and run commitbrief --staged before you commit.

← all posts