GitHub PR Review Agent with Claude Code
One pass over a pull request that might span sixty files, where GitHub hands back patches rather than whole files, so a good half of the bar has to be settled from fragments and the rest marked cannot tell. Reading comes first: the commits, the changed files, any review already sitting on the thread. Claude Code works through material before it forms a verdict, which is the order this template is built around.
Watch it work before it's live
Run a staged conversation — no account needed. The agent handles it for real while a simulated world answers its tool calls; nothing touches real accounts, and nothing is actually sent.
Stream tile responses instead of buffering the whole body
Openopened by @marijke-devos
Replaces the buffered writer in the tile handler with a streaming encoder, so a 40MB vector tile stops sitting in memory before the first byte goes out. Nine files, 214 lines added, 118 removed, all under internal/tiles. Adds a benchmark showing peak memory falling from 96MB to 11MB on the Jakarta extract, plus table tests for the partial write path. The description links issue 402 and states the wire format is unchanged.
Set up in minutes
Using this template drops you into a guided setup. It asks exactly this, nothing else:
Connect GitHub
One sign-in. The agent acts through your account, scoped to what this template uses.
Connect Slack
One sign-in. The agent acts through your account, scoped to what this template uses.
Review checklist
What a pull request has to satisfy in this repository, put the way you would explain it to somebody reviewing here for the first time: which items are hard blockers, which paths you treat as sensitive, and what you want asked when the diff alone cannot settle one.
Runs on Claude Code
Preselected for this page — connect your Claude Code account during setup, or switch to NoClick's built-in models with one click.
Watch it handle a test run
A staged conversation against a simulated world — then it’s live.
Why Claude Code for this agent
Patches are partial evidence
A hunk shows twenty lines around a change and nothing else. A harness that reads what it was actually given produces honest cannot tell lines rather than confident claims about code it never saw.
One pass, no second chance
It fires when the pull request opens and stays quiet through every push afterwards, so what gets read in that first minute is the entire basis of the comment. Long structured tasks are where this harness is steadiest.
Before you fork
A sixty file pull request is a lot of reading. What does that cost us?
You pay for the tokens the model you connect consumes, and a large diff is a large read, so the heaviest pull requests are also the dearest runs. It fires once per opened pull request rather than once per push, which keeps a busy repository predictable. Where depth matters less than spend, a smaller model still fills the three headings.
Run it with a different agent
Put GitHub PR Review Agent to work on Claude Code
Free to start. Guided setup, a test run against staged conversations, and it's live.