Code review that knows what your agents were asked.
Install the GitHub App and every pull request is reviewed in a minute or two. Grapim checks the change against what was asked, in the pull request, its issue or the agent’s own prompt, and flags what nobody asked for. Few comments, each one checked again before it is posted.
Works withClaude CodeCodexCursorGemini CLIGitHub
| 1 | -- Tidy up the users table. |
| 2 | ALTER TABLE users DROP COLUMN email; |
Blocker · Drops users.email, which the request did not ask for
The run was asked to add pagination. This line deletes every stored email address, and a rollback will not bring them back.
Written by Claude Code run: “Add pagination to the users list endpoint”
| 3 | DELETE FROM users WHERE last_login_at < now() - interval '90 days'; |
- A minute or two. From push to review.
- Knows the agent. And what each run was asked.
- Free each month. Review credit on every plan.
- Or in your own CI. Nothing reaches Grapim.
How it works
One switch on GitHub
Install the App and turn review on. No workflow to write, nothing to host.
Reviews agent code against the request
Asked to add pagination, Claude Code should not drop a column. Grapim flags work beyond the prompt.
Spends effort where the risk is
The deepest review goes where agents touch sensitive code.
What it found in real pull requests.
From our run on 69 public pull requests written with Claude Code. Nothing was posted, and we told maintainers privately, so no project is named.
-
grapimbotwould have commented
A login rate limit anyone can trip for you
The captcha now limits attempts per client IP, and a helper elsewhere in the project trusts
X-Forwarded-Forby default. Together, someone who knows your IP can keep you off the login page.Public repository · Generated with Claude Code
-
grapimbotwould have commented
A restarted process stops listening
The panel tells a background process to start watching for input when it opens. If that process restarts while the panel is open, nothing tells the new one, and the feature quietly stops until the panel is reopened.
Public repository · Generated with Claude Code
-
grapimbotwould have commented
A company parsed as a person
The new name parser would turn
ABC HOME CARE, LLCinto a person whose first name is LLC.Public repository · Generated with Claude Code
Few comments, and they hold up.
On 95 open-source pull requests, people checked Grapim’s findings by hand against the code.
- 83%of inline comments were correctUp from 47% before the second look.
- 35 → 2wrong comments shownOn the same 95 pull requests.
- 20 of 24real issues still posted inlineThe second look drops a few real ones too.
A second look checks each finding against the code around it and the pull request’s discussion before posting. We tuned it on this set, so read 83% as an upper bound.
Or keep your code in your own CI.
The same review on your runner and your model account. Grapim receives nothing.
- Your GitHub
- Your CI runner
- Your model account
- Grapimreceives nothing
… role-to-assume: arn:aws:iam::123456789012:role/grapim-review aws-region: us-east-1- uses: grapim/grapim/action/review@main with: provider: bedrockModel: claude-opus-5-5 on Amazon Bedrock (us-east-1). 2 requests, 18,412 input and 2,906 output tokens.
Your code was sent only to Amazon Bedrock (us-east-1). Grapim received nothing.
How it compares.
From each product’s own public pages, read on 20 November 2026.
| Feature | Grapim | GitHub Copilot code review | CodeRabbit |
|---|---|---|---|
| Open source | Yes, Apache 2.0 | No | No |
| Runs in your own CI | Yes, a GitHub Action on your runner, on every plan | Not stated on its public pages | A self-hosted container on Enterprise, from 500 seats |
| Your own model account | Anthropic API, Amazon Bedrock, Google Vertex AI or Microsoft Foundry | No, GitHub chooses the models | On self-hosted Enterprise |
| Where your code goes | In your CI: only to your model provider. Cloud: a temporary folder on Grapim’s server, deleted after the review | GitHub’s Copilot service | CodeRabbit’s service, or only your model provider when self-hosted |
| Checks the change against what was asked | The linked issue, the pull request and the agent run’s own prompt | Not stated on its public pages | Linked issues |
| Code hosts | GitHub | GitHub and Azure DevOps | GitHub, GitLab, Bitbucket and Azure DevOps |
| Free on public repositories | Yes | Not on Copilot Free | Yes |
| Price | Free plan; Team $16 per pull request author a month, billed yearly | $19 per seat a month (Copilot Business) | $24 per developer who opens pull requests a month, billed yearly (Essentials) |
A blank cell: the vendor’s public pages do not say, so we left it out. Products change; check the sources: GitHub Docs: Copilot code review, GitHub Docs: Copilot plans, CodeRabbit pricing, CodeRabbit self-hosted, CodeRabbit linked issues, CodeRabbit platforms.
Questions
Which model does it use?
Cloud review on the GitHub App uses Claude Opus 5.5 through Anthropic. In your own CI it uses the Claude model in your own account: the Anthropic API with your key, Amazon Bedrock, Google Vertex AI or Microsoft Foundry.
Is our code stored?
Cloud review fetches the pull request into a temporary folder on Grapim’s server, sends the changes to Anthropic, and deletes the folder when the review ends. For each comment Grapim keeps the file, lines, title and reactions, not the code. Run it in your own CI and Grapim receives nothing at all.
How is it different from Copilot code review?
Grapim checks each change against what was asked: the linked issue, the pull request and, when Grapim recorded the run, the agent’s own prompt. It lists what goes beyond that, such as deleted tests, checks turned off or a migration nobody asked for. It is open source and runs in your own CI on your own model account.
Does it work with Claude Code, Codex and Cursor pull requests?
Yes. Every pull request is reviewed, whoever wrote it. With the Grapim CLI recording runs from Claude Code, Codex CLI, Cursor or Gemini CLI, the review also knows which run wrote each line and what that run was asked.
Can we self-host it?
The review, yes: a GitHub Action on your runner with your own model account, and no Grapim account needed. The hub, the web app, is hosted by Grapim; talk to us if you need it on your own servers.
How is pricing counted?
Per pull request author: a person who opens pull requests Grapim reviews that month. Agents and bots are not authors. Free is $0, Team $16 and Governance $32 per author a month billed yearly ($20 and $40 monthly). Nothing is charged during early access.
Does it review private repositories?
Yes. Cloud review covers private repositories for a team with it turned on, about 80 pull requests a month on Free. Grapim Review in your own CI works on any repository, free. Public repositories are reviewed free on the App.
Try it on one repo.
Install the GitHub App and open a pull request, or run the same review in your own CI.