Getting started

What you get out of the box

Every PR opened or updated on a connected repo gets reviewed by six specialist LLM agents in parallel plus a deterministic CI-security agent. Inline comments post to the PR with tier tags (🎯 Actionable / 💡 Suggestion / 🔍 Nitpick), a verdict lands (approve / request changes / comment), and a full report is reachable from the dashboard.

Four-step setup

The dashboard shows this same checklist as a progress banner when any step is incomplete — auto-hides when everything is done.

1. Install the GitHub App

Open app.looksgoodtomeow.in → Sign in with GitHub, then install the tarin-lgtmGitHub App on the org/user account that owns the repo you want reviewed. Fine-grained access — you pick the repos. Webhooks are configured automatically — nothing to click on GitHub's side.

2. Connect a repository

Repositories → Connect Repo. Pick one or more repos from the authorized set. On connect, LGTM auto-enqueues a background context workflow — tree-sitter parses your codebase, extracts conventions, detects frameworks, and builds a PageRank- ranked repo map. Zero manual click required. A toast tells you the first PR review will start working in ~2-5 minutes.

3. Wait for indexing (auto)

Indexing runs in the background — real-time progress on the Repos page. Once complete, the repo status flips to Ready. Subsequent pushes to the default branch re-index touched files automatically (SHA-diff aware — no wasted work). See Context Indexer for what gets cached.

4. Add an AI provider key

Settings → AI Provider. Bring your own OpenAI key, or an OpenRouter key for unified access to Claude Opus 5, Sonnet 5, Gemini 3.1 Pro, Grok 4.5, DeepSeek V4, Llama 4, and 100+ more models through one gateway. Keys are AES-256-GCM encrypted at rest, decrypted only inside the review worker at review time. Never proxied through our infra, never billed to us.

Open a PR and watch the review

If auto-review is on (default for paid tiers), just open a PR. Within seconds the dashboard shows agents starting one by one; the full review posts to GitHub as soon as the synthesizer completes. Free-tier users click Review manually on the PR card, or use the CLI:

# review the local diff before pushing
lgtm review

# or trigger a review on an already-open PR
lgtm review --pr 42

Key concepts

Four ideas explain most of LGTM's design. Skim these and the rest of the docs will click.

BYOK (Bring Your Own Key)

You provide your own OpenAI key, or an OpenRouter key for unified access to Claude, Gemini, Grok, DeepSeek, Llama, Mistral, Qwen, and 100+ more models via one API. Code goes directly from our worker to the provider — never stored past the review session. LGTM's platform pricing (₹0 / ₹399 / ₹999 for Free / Hobby / Pro) is separate from the model spend that hits your provider bill.

Multi-agent review (v2 merged pipeline)

Six specialist lenses — bugs, security, performance, documentation, readability, best-practices — plus a deterministic ci-security agent when CI/CD files change. Under the hood, the 6 lenses are batched into two structured-output LLM calls (Correctness + Style) followed by a batched verifier — same review depth as running each lens separately, 5-10× cheaper, ~4× lower latency. A synthesizer merges surviving findings into one verdict + a capped set of inline comments.

Context-aware

The context indexer parses your repo with tree-sitter, builds a symbol dependency graph, ranks files with PageRank, and ships a compact repo map + smart file windows into every review prompt. Reviews aren't "diff-only" — they see the neighbourhood.

Real-time updates

Socket.IO events stream progress: agent-by-agent completion, synthesizer running, verdict landed. The dashboard PR-detail page updates live — no polling, no page refresh.

Next steps

• Agent architecture — what each agent does and how findings survive to the verdict
• .lgtm.yml Config — repo-level overrides for teams that need them
• Recipes — copy-paste configs for common problems
• CLI — install @tarin/lgtm-cli to review local changes before you push