Amit.dev
Multi-agent · Live tool

Multi-Agent PR Reviewer

Paste any public GitHub PR. Four specialised reviewers — correctness, security, style, tests — read the diff in parallel. A lead reviewer synthesises a verdict with severity-graded top issues. Real-utility multi-agent orchestration, not a demo of one.

Paste a public GitHub PR

Four specialised agents — correctness, security, style, tests — review it in parallel. A lead reviewer synthesises the verdict.

Try:

How this works under the hood

  1. Fetch — server-side, the URL is parsed and the PR + file diffs are pulled from the GitHub public API. Up to 10 largest files survive into the review window with per-file diff truncation, so the context stays cheap.
  2. Fan out — the browser fires four parallel POSTs to the shared streaming LLM route, each with a tightly scoped system prompt for one review dimension. Cards stream in independently as the model produces tokens.
  3. Synthesise — once all four reviewers finish, a lead-reviewer prompt receives the diff plus all four review outputs and emits a structured verdict (LGTM / LGTM_WITH_NITS / NEEDS_CHANGES / BLOCK) plus the top severity-graded issues.
  4. Rate limit — five LLM calls per review hit the shared per-IP counter; abuse caps out the same way Ask Amit does.

Limits: only public repos, ≤10 largest files per PR, diff snippets truncated to 2 KB each, output capped per reviewer.