claude-pyls 0.1.0

Lightweight Python LSP MCP server + grep-intercept hook for Claude Code

claude-pyls

A lightweight Python LSP MCP server + grep-intercept hook for Claude Code. Gives Claude token-efficient symbol navigation in any Python project.

Why

When Claude Code searches for a Python symbol using Grep, it reads large swaths of source into context. A targeted LSP lookup costs a fraction of that.

Measured savings (Haiku, claude -p, 400-file project):

Metric	Without (Grep)	With (pyls)	Savings
Cost	$0.0586	$0.0119	80% cheaper
Cache creation tokens	39,124	2,953	92% less
Duration	10.6s	5.7s	46% faster

Modes

Small (single-repo, zero-startup)

Regex scan over class/def lines. Starts in 0.2s (raw JSON-RPC, no MCP SDK). Optionally enhances with jedi for precise goto/references. Ideal for per-project activation.

Big (environment-wide, daemon)

A persistent LSP daemon that starts on boot and serves an entire Python environment. Longer startup, higher accuracy, multi-repo awareness. Good for large monorepos or shared dev environments. (Planned — not yet implemented.)

Install

pip install claude-pyls          # small mode (regex scan)
pip install claude-pyls[lsp]     # + jedi for enhanced accuracy

Quick start

# Scaffold .mcp.json + hook into your project
claude-pyls setup

# Or manually add to .mcp.json:
# {"mcpServers": {"pyls": {"command": "claude-pyls", "args": ["serve"]}}}

Tools exposed

Tool	Description
find_definition	File + line where a symbol is defined
find_references	All sites where a symbol is used
find_symbols	All definitions matching a name pattern

Known limitations

• Textual scan only (small mode) — won't resolve dynamic attributes, decorated names, or re-exports
• jedi (when enabled) is lazy-loaded on first call (~3–4s first call, fast after)
• File cache is not invalidated mid-session — restart the server after large refactors
• Hook catches bare symbol names (PascalCase, camelCase, snake_case 2+ segments) — prefixed patterns pass through

Releasing

Publishing is manual for now (no CI/CD). The PyPI API token lives in a local .env file (gitignored) in the format pypi=pypi-....

# one-time: install hatch into the dev venv
pip install hatch

# build wheel + sdist
hatch build

# publish — reads the token out of .env WITHOUT exporting every line
# of .env into your shell (avoids leaking unrelated vars). Never commit .env.
HATCH_INDEX_USER=__token__ \
  HATCH_INDEX_AUTH="$(grep '^pypi=' .env | cut -d= -f2-)" \
  hatch publish

# tag the release (replace X.Y.Z with the version you just bumped
# in src/claude_pyls/__init__.py and pyproject.toml — see below)
git tag vX.Y.Z && git push origin vX.Y.Z

Version bump procedure

Before each release, bump the version in both locations (they must stay in sync):

1. src/claude_pyls/__init__.py — __version__ = "x.y.z"
2. pyproject.toml — version = "x.y.z"

Then commit, build, publish, and tag as shown above.

License

MIT