project-lore-mcp

A local-first MCP server that routes coding agents to the smallest useful set of project documentation — decisions, intent, and constraints — with provenance and freshness.

Code-intelligence tools (such as codebase-memory-mcp) answer where is this implemented and what depends on it. This project answers the complementary questions: why was it designed this way, which document governs this task, and is that knowledge still current?

What it does

Point it at a repository (plus optional external documentation roots). It:

1. discovers Markdown, text, HTML, YAML, JSON, Mermaid, and Claude instruction files;
2. classifies each document by kind and authority (instructions → canonical → design → planning → historical → generated → proposed);
3. extracts headings, sections, links, frontmatter, ADR status, and explicit "read first" / "supersedes" / task-routing relationships — deterministically, with recorded provenance (structural extraction applies to Markdown; other formats receive plain-text indexing);
4. stores a rebuildable local index in SQLite (FTS5), updated incrementally by content hash;
5. answers task and knowledge queries with bounded excerpts and exact source locations — never whole files, flagging stale or superseded sources instead of hiding them.

MCP tools

| Tool | Purpose | | --- | --- | | route_task | Given a development task, return the ordered documents/sections to read first. Explicit manifest routes win over ranked search. | | search_project_knowledge | Full-text search over documentation, re-ranked by authority and status; freshness is shown per result. | | get_document_context | Sections + typed relationships of one document, without returning the whole file. |

Install and run

Requires Node.js ≥ 22.

npm install -g project-lore-mcp   # or npx -y project-lore-mcp (no install needed)
project-lore index --root /path/to/your/repo   # one-shot index + stats
project-lore serve --root /path/to/your/repo   # stdio MCP server

Claude Code

Run this once from inside the repository you want to index:

# Personal setup — stored in your local Claude Code config, not committed
claude mcp add project-lore --scope local -- \
  npx -y project-lore-mcp serve --root "$PWD"

For teams, commit a shared configuration instead:

# Shared setup — writes .mcp.json into the repository
claude mcp add project-lore --scope project -- \
  npx -y project-lore-mcp serve --root .

The project scope uses a relative root (.) so it works on any checkout. Claude Code will ask each user to approve project-scoped servers before running them.

One repository, one configuration. Project Lore's command contains a fixed root path, so it always indexes one specific repository. Register it separately per repository rather than globally.

It coexists with codebase-memory-mcp — configure both; they own different questions. If your configuration references documentation outside the repository, Claude Code must be granted filesystem access to those paths (add them as additional working directories or approve the permission prompts).

Try the example

npm run index:example        # index examples/documented-project
npm run serve:example        # serve it over stdio

Configuration

Optional project-lore.config.yaml at the repository root:

sources:
  - id: repository
    type: directory
    path: .
    authority: canonical
  - id: design-handoff
    type: directory
    path: ../product-docs      # explicit opt-in external root
    authority: design
classification:
  rules:
    - pattern: "docs/legacy/**"
      authority: historical

See docs/configuration.md for the full reference and docs/concepts.md for the authority model.

Design principles

Store and retrieve documented intent conservatively. Derive structure automatically. Preserve provenance always.

• Local-only by default. No network access, no telemetry, no uploads. The index is regenerable and safe to delete (.project-lore/, gitignored).
• Deterministic before AI. v1 uses no LLM and no embeddings. Every stored relationship records how it was extracted; inferred data can never masquerade as documented fact.
• Evidence, not instructions. Retrieved excerpts are quoted, bounded, and attributed. Document content is treated as untrusted data (see docs/provenance-and-trust.md).

Project status

v0.1.0 — first vertical slice. See docs/architecture.md, the ADRs in docs/adr/, and CONTRIBUTING.md.

License: MIT.