Spine

Repository architecture with file-level proof

Spine gives an agent a local map of a repository before it edits:

• What is this repository?

• Where does this behavior live?

• How does a request reach that module?

• What does this change affect?

• Which file and lines support the answer?

The graph stays on the machine. Spine does not call a model, and it does not treat generated prose as proof.

npx -y @sylphx/spine

This starts the MCP server on stdio. For Claude Code:

claude mcp add spine -- npx -y @sylphx/spine

Package @sylphx/spine · bin spine · MCP io.github.SylphxAI/spine · docs https://sylphxai.github.io/spine/

The default is a refresh

Omit mode. Spine reuses the index when nothing changed, updates only the files that changed when the delta is small, and scans the repository when it has to. The response names which one happened: cache_hit, incremental, or full.

{
  "root": "/absolute/path/to/repo"
}

A complete rescan is named explicitly:

{
  "root": "/absolute/path/to/repo",
  "mode": "full"
}

• refresh is the default. auto is the old name for that same behavior.

• status_only checks freshness and does not index.

• Any other mode forces a full scan.

• No remote provider and no model call.

• Unknown languages and incomplete coverage come back as gaps.

Related MCP server: Paparats MCP

Ask one question

{
  "query": "authMiddleware"
}

Search matches a label or a path fragment, such as a symbol name. It returns ranked nodes with path, line range, extraction source, freshness, confidence, and known gaps.

Jobs Spine is built for

Tool surface

architecture_context_pack is an advanced composition tool. Start with the question-oriented tools above.

Why an agent can check the answer

Every result carries:

• repository root and indexed commit;

• current commit and freshness (fresh, stale, dirty, or unknown);

• evidence with file and line locators;

• extracted versus inferred labels;

• explicit gaps instead of a silent guess.

What Spine is not

Locus owns code-chunk retrieval. Spine owns architecture claims. Neither replaces the other.

Companion MCP tools

Each product is independent. Install only the tools the agent needs.

Development

bun install
bun run build
bun test
cargo test

For a release candidate:

bun run benchmark:public-proof
bun run benchmark:release-gate

License

MIT