setsu

Documentation · Quickstart · Commands · Privacy · Roadmap

npx @setsu-ai/cli init                 # detect agents, build the graph, wire up MCP
setsu context "How does auth work?"    # token-budgeted evidence, not the whole repo

Available on npm as @setsu-ai/cli · latest version 0.1.0-beta.1 (2026-09-08) · for Claude Code, Codex, Cursor, Copilot and Kiro

🧠 What is SETSU?

AI coding agents rediscover your repository from raw files on every task: grep, read, re-read, repeat. That loop burns your context window and your money.

SETSU builds a persistent code graph of your repository once (tree-sitter parse → symbols and typed edges → PageRank and communities → SQLite), keeps it fresh incrementally, and answers questions with token-budgeted evidence packs — the smallest set of signatures, snippets, and graph paths that answers the question, with provenance on every edge.

• 🗺️ Persistent code graph — symbols, calls, imports, implements, tests; typed and provenance-labeled

• 🎯 Token-budgeted retrieval — every pack has a hard budget; a greedy value/cost packer fills it

• 🔀 Adaptive routing — graph traversal, lexical FTS, repo map, or hybrid, chosen per task

• ⚡ Incremental — content-addressed cache; unchanged files are never reparsed

• 🔌 MCP-native — five tools, stdio only, no network listener

• 📉 Instruction doctor — measures what your CLAUDE.md/rules actually cost per session

• 🧪 Learns locally — which strategy works for which task class, on your machine only

• 🔒 No API key. No account. Source never leaves your machine.

Related MCP server: engram-serve

⚙️ How it works

            AI CODING AGENT (Claude / Codex / Cursor / ...)
                              │  setsu_context("How does checkout reach Stripe?")
                              ▼
                    ┌──────────────────┐
                    │   SETSU ROUTER   │  task classifier
                    └────────┬─────────┘
          ┌─────────────────┼──────────────────┐
          ▼                 ▼                  ▼
   Graph traversal    Lexical / FTS5      Repo map (PageRank)
          └─────────────────┼──────────────────┘
                            ▼
                  TOKEN BUDGET PACKER
                            ▼
              ContextPack (~1,300 est. tokens)
     path: CheckoutController → CheckoutService → PaymentGateway → StripeAdapter

Below the retrieval layer: files → content hash → tree-sitter parse → symbol + edge extraction → import resolution → PageRank + Louvain communities → SQLite (FTS5 included) → incremental watcher.

⚡ Quickstart

npm install -g @setsu-ai/cli    # or npx @setsu-ai/cli init

setsu init                      # detect agents, build the graph, offer MCP install
setsu context "How does checkout reach the payment gateway?" --explain
setsu doctor                    # what your agent setup costs per session
setsu optimize --dry-run        # what SETSU would improve, as diffs

🧰 Commands

🔒 Privacy

Everything is local: the graph lives in ~/.setsu/repos/<id>/graph.db, usage learning in ~/.setsu/setsu.db. No network calls, no telemetry, no accounts. Task terms are hashed with a per-device salt before storage; raw prompts are never persisted. setsu privacy inspect lists every stored category and location; setsu privacy purge --yes deletes all of it.

🗺️ Roadmap

🤝 Contributing

Contribution ladder — pick your entry point:

1. Benchmark task — add a task to benchmarks/; retrieval quality lives and dies by real questions

2. Fixture + golden — a code pattern we extract badly makes a great failing test

3. Language extractor — packages/parsers/src/extractors/ has the pattern to follow

4. Adapter depth — deepen Copilot/Kiro config generation in packages/agents

5. Core — router, packer, graph algorithms

npm install && npm test is all it takes to start. Every PR needs tests.

📄 License

Apache-2.0 © 2026 Aritra Ghosh · NOTICE · Third-party notices