CLI reference

The lcp command-line tool is installed by pip install lcp. All subcommands accept --help for inline documentation.

lcp --version
lcp <command> --help

lcp scan

Introspect an installed Python package and emit an LCP manifest.

lcp scan PACKAGE [OPTIONS]

Flag	Default	Description
PACKAGE	—	Importable package name (e.g. requests).
-o, --output PATH	stdout	Output file path.
--include-private	off	Include private symbols (starting with _).
--no-recursive	off	Don't scan submodules recursively.
--include-tests	off	Include *.tests subpackages (excluded by default).
--validate / --no-validate	on	Validate output against LCP schema.
--indent INTEGER	2	JSON indentation level.
--coverage PATH	—	Also generate a documentation coverage report to this path.

By default, *.tests subpackages (whose leaf name is exactly tests) are skipped: they are not public API and would otherwise pollute the manifest. Public utilities such as numpy.testing are unaffected. Pass --include-tests to scan them anyway.

Example:

lcp scan requests -o requests.lcp.json
lcp scan numpy --include-private

Machine-mode scanning: python -m lcp.scanjson

For programmatic callers, lcp.scanjson is a machine-mode scan entry point with a strict output contract. It writes the LCP JSON document to stdout, reports errors as a single JSON object on stderr, and distinguishes failures by exit code:

python -m lcp.scanjson PACKAGE [--include-private] [--no-recursive] [--include-tests]

Exit code	Meaning	Output
0	Success	LCP JSON document on stdout
3	Import failure — the package could not be imported	{"type": "import_failure", "message": ...} on stderr
4	Scan failure — the package imported but scanning or generation failed (including SystemExit raised by import-time code)	{"type": "scan_failure", "message": ...} on stderr

Unlike lcp scan, this entry point never prints human-readable progress, and it only needs lcp and pydantic importable in the interpreter that runs it (no click) — so it can run inside a project virtualenv that does not have the full CLI installed. Import-time prints from the scanned package are redirected to stderr so they cannot corrupt the stdout document.

lcp.scanjson is a same-venv entry point: it emits the full LCP document and expects lcp installed in the interpreter running it. Cross-venv scans — the MCP server documenting a package in a different virtualenv — instead go through lcp.subprocess_scan, which runs an isolated child that loads only the scanner by file path (never importing the host's lcp/pydantic) and hands the raw scan back to the host for generation, so the host environment can never contaminate the target scan.

python -m lcp.scanjson requests > requests.lcp.json

lcp validate

Validate an LCP JSON file against the LCP schema.

lcp validate FILE

Flag	Default	Description
FILE	—	Path to an LCP JSON file to validate.

Example:

lcp validate requests.lcp.json

lcp serve (deprecated)

Deprecated

lcp serve is deprecated — use lcp serve-all with --expose <package> instead. The command still works: it starts the same universal server, pre-loaded with the manifest and restricted to that library, and prints a deprecation warning on stderr.

lcp serve MANIFEST [OPTIONS]

Flag	Default	Description
MANIFEST	—	Path to an LCP JSON file to serve.
--name TEXT	lcp-{library-name}	Server name for MCP identification.

lcp serve-all

Start a universal MCP server that resolves any installed Python library on the fly. No pre-built manifest is required — AI agents call the resolve_library tool to load any pip-installed package, then search and get_symbol to explore it. Manifests are cached locally and, when local scanning fails, can be fetched from a remote LCP registry.

Resolution order: local cache → live scan → registry fetch.

lcp serve-all [OPTIONS]

Flag	Default	Description
--cache-dir PATH	~/.lcp/cache/	Cache directory for LCP manifests.
--name TEXT	lcp-universal	Server name for MCP identification.
--no-cache	off	Disable reading from and writing to the local cache.
--registry TEXT	—	Base URL of an LCP registry used as fallback when local scanning fails. Manifests are fetched from {registry}/manifests/{language}/{first_letter}/{slug}/{version}.lcp.json.gz, where {slug} is the hyphenated package name (google.adk → google-adk).
--expose TEXT	all packages	Restrict resolve_library to these package names (repeatable).
--preload TEXT	—	Resolve these packages at startup (repeatable).
--max-response-bytes INT	25000	Byte budget for list-returning tool responses (context blowout guard).
--scan-mode [subprocess\|inprocess]	subprocess	How resolve_library scans installed packages. subprocess isolates package imports in a disposable child process (crash isolation, cross-venv scanning, no import-lock stalls); inprocess imports into the server process, for environments where spawning is restricted.
--scan-python TEXT	server's interpreter	Python interpreter whose environment resolve_library scans. Lets the server document packages installed in a different venv.
--scan-timeout FLOAT	60.0	Seconds before a subprocess scan is killed.

Example:

lcp serve-all
lcp serve-all --registry https://raw.githubusercontent.com/zazza123/lcp-registry/refs/heads/main
lcp serve-all --expose requests --preload requests
lcp serve-all --scan-python /path/to/project/.venv/bin/python

lcp coverage

Generate a documentation coverage report for a Python package, showing which symbols are missing docstrings.

lcp coverage PACKAGE [OPTIONS]

Flag	Default	Description
PACKAGE	—	Importable package name to analyze.
-o, --output PATH	stdout	Output file path.
--format [json\|markdown]	json	Output format.
--include-private	off	Include private symbols (starting with _).
--no-recursive	off	Don't scan submodules recursively.
--include-tests	off	Include *.tests subpackages (excluded by default).

Example:

lcp coverage requests -o coverage.json
lcp coverage numpy -o coverage.md --format markdown

lcp docgen

Generate missing docstrings for the symbols listed in a coverage report, using an LLM provider. Requires the optional AI extra (pip install "lcp[ai]"). Processing is hierarchical bottom-up (methods → classes → modules) with parallel LLM calls; generated docstrings are injected into the package's source files in place.

lcp docgen COVERAGE_JSON [OPTIONS]

Flag	Default	Description
COVERAGE_JSON	—	Path to a coverage report produced by lcp coverage.
--provider [openai\|anthropic]	openai	LLM provider.
--model TEXT	provider default	Model name.
--api-key TEXT	env var	API key; defaults to OPENAI_API_KEY / ANTHROPIC_API_KEY.
--kinds TEXT	all	Filter by symbol kind, comma-separated (e.g. class,function,method).
--description TEXT	—	Package description used to guide the model.
--reasoning	off	Enable reasoning mode for OpenAI models (o1, o3, ...).
--dry-run	off	Show what would be modified without writing files.
--workers INTEGER	4	Max concurrent LLM calls.
--failure-threshold FLOAT	0.5	Ratio (0.0–1.0) of failed children that skips the parent symbol.

Example:

lcp coverage mypackage -o coverage.json
lcp docgen coverage.json --provider openai --dry-run
lcp docgen coverage.json --provider anthropic --workers 8

See the AI DocGen guide for how hierarchical processing works and the Python API.

lcp publish

Publish an LCP manifest to the registry by opening a GitHub Pull Request. The command scans the package (or uses an existing manifest via --file), validates it, then submits the PR to the registry repository.

lcp publish PACKAGE [OPTIONS]

Flag	Default	Description
PACKAGE	—	Importable package name to publish.
--token TEXT	—	GitHub personal access token with repo or public_repo scope. Can also be set via LCP_GITHUB_TOKEN or GITHUB_TOKEN env var.
--registry-repo TEXT	zazza123/lcp-registry	Target registry repository in owner/name format.
--file PATH	—	Use an existing LCP JSON file instead of scanning the package.
--include-private	off	Include private symbols when scanning (starting with _).
--no-recursive	off	Don't scan submodules recursively.
--dry-run	off	Generate the manifest and show what would be submitted without creating a PR.

Example:

lcp publish requests --token ghp_xxxx
lcp publish numpy --dry-run
lcp publish mylib --file mylib.lcp.json --token ghp_xxxx

lcp diff

Compare two LCP files and detect deprecated symbols. Symbols present in the older file but missing in the newer file are reported as removed. The output includes generated deprecation entries that can be merged into the new manifest.

lcp diff OLD NEW [OPTIONS]

Flag	Default	Description
OLD	—	Path to the earlier LCP JSON file.
NEW	—	Path to the later LCP JSON file.
-o, --output PATH	stdout	Output file path.
--indent INTEGER	2	JSON indentation level.
--update	off	Write detected deprecations back into the NEW LCP file.

Example:

lcp diff v1.lcp.json v2.lcp.json
lcp diff v1.lcp.json v2.lcp.json -o diff.json --update

See also

• Quickstart — first-time usage.
• MCP Server guide — using lcp serve and lcp serve-all in depth.
• AI DocGen guide — generating missing docstrings with lcp docgen.
• Publishing guide — using lcp publish to submit to the registry.