registree

An anti-hallucination class registry for coding agents, served over MCP.

Coding agents guess constructor signatures from memory — a keyword that doesn't exist, a required argument left out — and you pay for the guess in a TypeError and a debugging round-trip. registree removes the guess: it walks your codebase with ast (never imports), builds a registry of every class definition, and serves it as MCP tools so the agent can verify the signature before writing the call.

Two principles run through every tool:

• Names map to lists. A duplicated class name returns every definition; the server never silently picks the first match.
• Honesty over confidence. An open-ended constructor (**kwargs, Pydantic extra=/alias=) reports its contract as unknowable, never as an empty list pretending to be an answer.

How it works

1. Scan — an AST walk over your source tree extracts every class: constructor parameters (including **kwargs and positional-only), typed fields with defaults, inheritance, docstrings. Classification is transitive: a model routed through your project's own base class is still recognized as a Pydantic model.
2. Serve — the registry is cached as JSON and exposed over MCP stdio. It maintains itself: generated on first use, regenerated whenever a scanned file is newer than the cache.
3. Answer — the agent queries it at the moment of use, instead of guessing.

MCP tools

tool	use it
get_signature	before writing a constructor call — required args, accepted keywords, every definition of a duplicated name
verify_snippet	after drafting code — checks constructor calls against the registry
search_classes	when unsure of the exact class name
list_duplicates	which names need an explicit import to disambiguate
get_usages	before a rename — every usage, including through import aliases (X as XDB)
server_info	server status and registry size

Install

No install needed with uv — MCP clients launch it with uvx. For direct CLI use:

uv tool install registree   # or: pip install registree

Requires Python 3.12+.

Wire it into your agent

Any MCP client, JSON config form:

{
  "mcpServers": {
    "registree": {
      "command": "uvx",
      "args": ["registree", "serve", "--root", "/path/to/your/project"]
    }
  }
}

Omitting --root serves the directory the client launches the server in, which for most MCP clients is the project root.

• Quickstart (any agent): docs/quickstart.md
• Claude Code setup, including the hook integration: docs/claude-code.md

Agent compatibility

The MCP tools work with any MCP client — Claude Code, Claude Desktop, Cursor, Windsurf, Cline, Zed, VS Code Copilot agent mode, Gemini CLI, and anything else that speaks MCP over stdio. Structured tool output degrades gracefully for clients that only read text content.

The hook adapters (hook-check, hook-regen) target Claude Code's hook protocol, which can intercept a pending file edit and hand the model advisory feedback before the write lands — a deterministic checkpoint the MCP layer alone can't provide. See docs/claude-code.md.

That protocol is spreading: VS Code Copilot agent mode (Preview) reads the same format — same events, same stdin JSON, even .claude/settings.json — and the adapters tolerate its camelCase field names. Codex CLI and Gemini CLI use close-enough hook contracts that ports are straightforward. Agents whose hook systems can't intercept file edits pre-application (Cursor, Windsurf without model feedback, Zed with no hooks yet) still get the full MCP toolset — the hooks just add a deterministic layer where the platform supports one.

CLI

The same engine is available directly:

registree gen                 # build the registry
registree conflicts           # duplicate names: accepted layering vs smells
registree usages SomeClass    # every usage, alias-aware — run before renames
registree hook-check          # Claude Code PreToolUse adapter (advisory)
registree hook-regen          # Claude Code PostToolUse adapter (debounced)

registree conflicts exits non-zero only for duplicate names that are genuine smells — the accepted ORM/domain layered pair passes — so it is safe to wire into CI from day one.

The registry cache

Lives at .registree/registry.json by default — add .registree/ to your .gitignore. Every command that touches it accepts --registry-path to put it anywhere else; relative paths are anchored to the project root.

Development

uv sync            # creates .venv, installs deps + dev tools
uv run pytest      # includes a real stdio JSON-RPC handshake test
uv run mypy src tests
uv run ruff check .
uv run black --check .

Run the server directly (speaks MCP over stdio; exits on EOF):

uv run registree

License

MIT