Sending a code diff to an existing LLM API and posting its comments back to a pull request is a genuinely achievable integration — the same shape as this register's GitHub Copilot case, just triggered on review instead of while typing.
Borderline
Effort: a weekendRunning cost: $0/mo to start
Why CanICodeThis says Borderline
A webhook that fires on a new pull request, sends the diff to an LLM API, and posts the response as a comment is a bounded, well-documented integration
Basic checks (missing tests, obvious style issues) can be prompted for directly without extra infrastructure
For your own repos, this is genuinely a weekend project
The parts that'll cause problems
Reliably understanding a large diff in the context of the whole codebase (not just the changed lines) needs real retrieval work beyond a raw diff in the prompt
Avoiding noisy, low-value comments that annoy a team takes real prompt-tuning and iteration
Supporting many different languages and frameworks well needs broad testing, not a single prompt
How I'd build this
frontend
Next.js
hosting
Vercel
MVP scope
A webhook on pull-request creation
Send the diff to an LLM API with a review-focused prompt
Post the response as a PR comment
Postpone to v2
Whole-codebase context retrieval, not just the diff
AI agent — with a subscription (Claude Pro/Max, Cursor, etc.)
$0 marginal
AI agent — pay-per-use API, no subscription
$43–$86 one-time
Hosting, once it's running
$0/mo (free tier)
Domain name, if you want your own
~$12/yr
Existing tools solving a similar problem typically run $10–$20/mo — see how they compare below.
The build prompt
Build me a code review bot, genuinely achievable with an existing model:
1. A webhook that fires when a pull request is opened, using your git host's own webhook API.
2. Send the diff to an existing, established coding-capable LLM API with a prompt asking for a focused review.
3. Post the response back as a PR comment via the same git host's API.
That's a real, working review bot based on the diff alone. Out of scope: whole-codebase context retrieval (understanding your entire repo, not just the changed lines), a genuinely deep retrieval problem, and learning from which past comments were accepted or rejected, which needs a real feedback loop most single-repo bots don't have enough data to train on.