Swipr

SwiPR

Swipe-to-review GitHub PRs with AI context. Works as a Claude MCP plugin.

Try the live demo → · Add to Claude Desktop ↓ · Listed on smithery.ai

---

Why this exists

Zeno Rocha (CEO of Resend) on X:

"the cost of opening a PR has dropped to zero. now, we have tons of draft PRs waiting for a finite (and ultra precious) resource: attention. turns out the bottleneck is no longer creation. it's reviewing." — Feb 2026

"before our main repo had an average of ~20–40 open PRs on any given day. now, we average ~130–200 open PRs." — May 2026

PR review is a context problem. Whether a maintainer is triaging or an AI agent is deciding whether to merge, the question is the same: is this safe to ship? Answering it well requires knowing the risk level, contributor history, similar past changes, and which tests cover the diff — not just reading the lines changed.

SwiPR surfaces that context as a swipe UI for humans and an MCP server for agents.

---

How it works

1. Paste any public GitHub repo — SwiPR fetches open PRs and stores them with embeddings
2. Swipe right to approve, left to request changes, down to skip — or use J / F / Space
3. The right panel surfaces: risk score, AI summary, similar past PRs, contributor history
4. Hit "Why is this risky?", "Show me callers", or "What tests cover this?" for deeper context on demand
5. Ask anything in the chat — the AI has access to the full diff and codebase context

AI context is cached per PR — no repeated API calls on every card view.

---

MCP server

SwiPR exposes the same PR context as an MCP server. Add it to Claude Desktop or Cursor and review PRs directly from chat — no browser required.

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "swipr": {
      "command": "npx",
      "args": ["-y", "supergateway", "--streamableHttp", "https://v0-swipr-build.vercel.app/api/mcp"]
    }
  }
}

Then ask Claude:

• "Look up resend/resend-node PR #1247 — what's the risk and are there similar past changes?"
• "Which open PRs in vercel/next.js touch the router?"
• "Find tests that cover the changed files in PR #892."

All 12 MCP tools are pure database reads — zero AI credits consumed when using the hosted server.

SwiPR is listed on smithery.ai and mcp.so.

Available tools

Tool	What it returns
lookup_pr	Resolve owner/repo#number → internal ID
analyze_pr	Full PR data: files, patches, risk score, contributor stats, cached AI summary
risk_score	0–100 heuristic score with reasons
find_similar_changes	Past PRs in the same repo with semantic similarity
get_contributor_history	PR count, merge rate, first contribution date
inspect_file	Raw file content at HEAD
find_callers	Search patches for usages of a function name
find_related_tests	Test files that likely cover the changed code
git_blame_summary	Recent contributors to a file path
compare_with	File content at an arbitrary git ref
record_decision	Capture approve / changes / skip
summarize_session	End-of-session stats

---

Bring your own AI key

The chat feature works with any of these free-tier providers — click the ⌘ button in the header and paste your key. It's stored in your browser's localStorage only, never sent to the server.

Provider	Key format	Free tier
Groq	gsk_...	Yes — Llama 3.3 70B
Google Gemini	AIza...	Yes — Gemini 2.0 Flash
Anthropic	sk-ant-...	Paid — Claude Sonnet 4.6

---

Customizing risk scoring

The risk heuristic lives entirely in lib/scoring.ts — plain TypeScript, no ML, no external calls. Edit it to match your team's standards.

Default rules:

Signal	Score
Large diff (>500 lines)	+20
Medium diff (200–500 lines)	+10
Many files changed (>20)	+15
Touches config/lock/CI files	+20
No test files changed	+10
Empty PR description	+10
First-time contributor	+15
New contributor (<3 prior PRs)	+8
Low historical merge rate (<40%)	+10

To add a rule, add an if block in computeRiskScore:

const touchesAuth = files.some((f) => f.filename.includes("lib/auth"));
if (touchesAuth) {
  score += 25;
  reasons.push("Touches auth module — requires security review");
}

Score is capped at 100. Color thresholds (green/yellow/red) are at 40 and 70 — adjust in app/swipe/_components/view-helpers.ts.

---

Self-hosting

Requires three services — all have free tiers.

1. Clone and install

git clone https://github.com/nochinxx/SwiPR.git
cd SwiPR
pnpm install

2. Create accounts

Service	Purpose	Free tier
Neon	Postgres + pgvector	512 MB
Vercel AI Gateway	Claude + embeddings	$5 credits
GitHub PAT (optional)	Higher API rate limits	Free

3. Configure environment

cp .env.example .env.local

DATABASE_URL=postgresql://...      # Neon pooled connection string
AI_GATEWAY_API_KEY=...             # Vercel AI Gateway key
GITHUB_TOKEN=...                   # Optional — raises GitHub rate limit to 5000/hr

4. Initialize the database

-- Run in Neon SQL editor
CREATE EXTENSION IF NOT EXISTS vector;

pnpm db:push

-- Add HNSW indexes for fast similarity search
CREATE INDEX IF NOT EXISTS prs_embedding_idx ON prs USING hnsw (embedding vector_cosine_ops);
CREATE INDEX IF NOT EXISTS pr_files_embedding_idx ON pr_files USING hnsw (embedding vector_cosine_ops);

5. Run and deploy

pnpm dev          # local dev
pnpm build        # verify before deploying
pnpm precache     # pre-load repos for a demo

Push to GitHub, connect to a Vercel project, set the same env vars in Vercel settings. Point your Claude Desktop config at your own deployment URL.

---

Storage limits

Neon's free tier holds ~512 MB. Each ingested PR uses ~100 KB. That's around 5,000 PRs or 40–50 mid-sized repos before needing an upgrade. Ingest is capped at 100 open PRs per repo.

---

Stack

• Next.js 16 — App Router, React 19
• Tailwind v4 + shadcn/ui — styling
• Framer Motion — card animations
• Neon Postgres + pgvector — PR storage and similarity search
• Drizzle ORM — schema and queries
• Vercel AI Gateway — Claude Sonnet 4.6 (analysis), text-embedding-3-small (vectors)
• @ai-sdk/anthropic · @ai-sdk/google · @ai-sdk/groq — BYOK chat support
• mcp-handler — MCP server at /api/mcp

---

Contributing

Issues and PRs welcome.

• Stack is Next.js App Router + Drizzle + Neon + Vercel AI SDK — no other abstractions.
• Risk scoring in lib/scoring.ts — heuristic improvements especially welcome.
• Don't add GitHub OAuth or actual PR posting. Read-only access is intentional.
• Run pnpm build before opening a PR.