AI PR reviewer with your own API key

What the PR reviewer reads

An AI PR reviewer is useful when it explains a specific failure a change could introduce. InfraStudio Agents reads the pull request against the surrounding repository and its contributor guidance, then posts a summary, a verdict and comments attached to relevant changed lines on GitHub. It reviews your team's pull requests as well as agent-authored work. The purpose is to give the author evidence they can check, and to keep the review alongside the code that prompted it.

The review looks for correctness problems, security concerns, regressions, missing checks and deviations from the repository's own expectations. A useful comment connects an input or situation to the resulting behavior. A vague request to improve quality is less actionable than a reproduction or a proposed assertion. The reviewer can still misunderstand the code, miss a dependency or suggest a test that is unnecessary. Treat its conclusions as review feedback requiring judgment.

Event review can run when a pull request opens and again after new commits. Reviewing drafts is a separate option and is off by default. The audience setting offers members or everyone. Review eligibility against your GitHub repository and organization membership before enabling the agent; selecting members is not a separate verification of write permission. Expanding the audience to everyone can let people outside your team consume your AI tokens by opening pull requests, so make that choice deliberately on public repositories.

A real inline comment from our repository

On 5 November 2026, the InfraStudio Agents bot reviewed PR #152 and attached a comment to the command parser in infrastudio/runner/common.py. The proposed pattern accepted the command docs by itself, but commands with a focus argument could become ordinary questions. That mattered because another part of the product told the owner to send docs anyway.

“This branch only matches docs with nothing after it, so docs anyway … and docs <focus> become questions.”

The reviewer proposed checking the result of parsing docs anyway against the expected command and argument. That is a concrete connection between what the interface tells a person to type and what the parser accepts. This is one observed review, not an accuracy benchmark or evidence that all comments find bugs.

Bring your own key: what a review costs

You choose the AI provider and model and supply your own API key. Supported provider choices include Anthropic, OpenAI, AWS Bedrock and Azure OpenAI. Your provider charges for its tokens; InfraStudio Agents charges for the platform according to the selected plan. The platform adds the key to provider calls through a gateway, rather than exposing the credential to the reviewing agent. See security for the boundaries around runs, keys and repository access.

There is no fixed, universal cost per review. A larger diff, more surrounding code, a longer discussion or a more expensive model changes the bill. Input tokens, output tokens, caching behavior and your provider's prices also matter. Inspect the run's recorded token usage. The platform may show usage without a cost estimate when model rates are unavailable. Any platform estimate uses operator-configured rates, not a customer price setting. For your actual cost, use your provider's applicable rates and bill. We do not have a verified token count or cost for the PR #152 example, so we do not publish an estimate for it.

For an evaluation, begin with a few representative pull requests rather than extrapolating from the smallest one. Compare useful findings, false alarms and reviewer effort alongside token spend. Repeated reviews after new commits consume additional runs and tokens. The pricing page explains platform plans and run allowances; it does not promise a particular AI bill.

Set up a controlled first review

1. Sign in as a workspace admin or owner and choose the repository to connect through the GitHub App. New GitHub projects have their runner enabled and five event agents switched on by default.
2. If the workspace already has an AI key, review the intended trigger behavior before connecting the repository. Otherwise, connect it, open Agents, choose Settings on the project card, and review or disable event agents before adding a key.
3. Choose the PR reviewer audience deliberately and check which GitHub authors qualify for your repository. Choose whether to review drafts and whether new commits should trigger another review. Pick the model that you want to evaluate.
4. Add the AI key under Integrations with the intended project or workspace scope. As an admin or owner, use Run now for the PR reviewer and enter a real pull request number.
5. Read the inline comments and summary on GitHub, inspect the corresponding run, and decide which suggestions deserve changes. Leave the final merge decision with a person.

Manually added GitLab projects need their runner switched on separately. Do not assume every GitHub-specific review presentation or event path is identical on GitLab. The platform's optional Auto-merge agent is GitHub-only and opt-in; enabling the PR reviewer does not enable automatic merging.

How this fits beside CodeRabbit and Copilot

CodeRabbit's own review page describes contextual pull request review for GitHub and GitLab. It is a sensible alternative to evaluate if your main requirement is a dedicated review workflow across those hosts. GitHub's Copilot code review guide describes review comments and suggested changes inside GitHub. It may fit a team that already works in GitHub and wants that native workflow. These are examples of each product's capabilities, not complete inventories.

The fit for InfraStudio Agents is the connection to findings, plans, implementation and CI work, together with your own provider key and per-agent model choices. Choose on your repository's results and operating preferences; no exclusive review capability or universal quality advantage is claimed here. Continue with analytics to pull request, the error triage agent, the CI fixer, or our measured dogfooding case.