---
title: "Local reviews · Docs · SourceAnt"
description: "Documentation: Review uncommitted changes in your checkout, inspect findings in the local browser, and request a review from your coding assistant through MCP."
url: "https://sourceant.ai/docs/local-reviews/"
---

# Review changes before you push

SourceAnt can review the work in your checkout before you commit it or open a pull request. Read the diff, check it against applicable skills and recorded knowledge, fix the findings, and run another review from the terminal, local browser or your coding assistant.

## Open your checkout

Follow [Local setup](https://sourceant.ai/docs/local-setup) to install SourceAnt. Then start the agent and register your repository:

```
sourceant ui
sourceant repo add /path/to/checkout --name acme/app
sourceant repos
```

You can also add a repository from the browser. Open **Reviews** and select that repository. You do not need a GitHub App installation or a pull request for a local review.

## Review from the terminal

Run a review from your checkout, or pass its path:

```
sourceant review /path/to/checkout
sourceant review /path/to/checkout --against main
sourceant review /path/to/checkout --no-model
```

The command prints a browser link, waits for the review and prints the result. Use `--no-wait` to return the link immediately, or `--json` for structured output. An exit status of `2` means the review found the change was not ready. `--no-model` inspects the change without asking a model to judge it.

Set a review name with `--title` and select a skill with `--skill`. Repeat `--skill` to select several skills. Configure the model in the local browser before requesting model review.

## Read what changed

Click **Read what changed** to inspect the changed files, commits, applicable skills and recorded knowledge. This makes no model call and does not judge the change. It works before you configure a model.

The comparison includes branch changes, staged and unstaged edits, and untracked files that Git does not ignore. SourceAnt chooses the comparison branch from the checkout's remote default branch when available, then finds the common ancestor with your branch. Check the branch and comparison shown in the review before relying on its scope.

## Run a review

1.  Open **Settings → Model**, choose a model and enter its provider key. Save and check the connection result.
2.  Return to **Reviews**, select your repository and check the skills selected for the change. Add or remove skills as needed.
3.  Click **Review it**. The review runs in the background and opens a page you can return to.
4.  Read the **Overview**, inspect the **Files**, and examine the findings and skill results beside the changed code.
5.  Make your changes, run your project's checks, and request another review of the updated checkout.

Local reviews use the same code reviewer as pull requests, with the context available on that local instance. See [Skills](https://sourceant.ai/docs/skills) for reusable instructions and [review controls](https://sourceant.ai/docs/code-review#controls) for exclusions and nitpicks. Completed reviews remain available in the local review history.

## Ask your coding assistant

Connect your MCP client to `http://127.0.0.1:8930/mcp`, or configure `sourceant mcp` as its standard-input/output command. Keep the local agent running. MCP lets the assistant call SourceAnt's tools on your machine.

Ask the assistant:

> Review the working tree for acme/app with SourceAnt and give me the review link.

The `review_working_tree` tool accepts the registered repository name or a checkout path:

```
{
  "repository": "acme/app",
  "title": "Check the checkout before opening a PR"
}
```

The tool returns a review ID, status and browser link while the review continues in the background. Open that link to read the result. A running status is not a completed review. This tool requires an instance that can reach your checkout; the hosted MCP surface cannot read files on your laptop.

## Use reviews alongside your project checks

Run your project's tests, lint, type checks and build as part of local verification. A review finding can identify a likely problem, but an empty findings list does not prove that the tests pass or every changed path was examined. Review coverage reports what was read and what was excluded or unavailable.

**Read what changed** is an inspection, not a test run or an approval. **Review it** adds model review and the analysis tools available to the instance; it does not replace your project's test suite or CI.

## Where the work goes

The local agent reads your checkout and keeps the review for its local browser. Starting a local review does not publish it as a pull-request comment. When you request model review, code and relevant context are sent to the model provider you configured. Local installation does not mean that a remote model runs offline. Provider usage is charged according to your provider account.

Local knowledge belongs to that instance. SourceAnt Cloud provides the managed team workspace and hosted pull-request workflow; a local review does not automatically make its knowledge available to a cloud review.

## When a review cannot run

-   **No repository is listed:** register the checkout and check `sourceant status` and `sourceant repos`.
-   **“Review it” is unavailable:** choose a repository, then configure the model and provider key in Settings.
-   **No applicable skill:** inspect the Skills page and the skills selected for this change. A message that nothing was judged is not a passing assessment.
-   **The comparison branch cannot be found:** check the checkout's Git history and default branch. Use `sourceant review --against main` to choose the comparison branch.
-   **The MCP tool is missing:** connect to the local agent, check that the code reviewer is loaded, and refresh the client's available tools.