title: API — the MCP server description: pardesline.mcp_server: build_server, main, skills_dir and the impl_* functions behind every MCP tool.
MCP server¶
pardesline mcp: the package as a Model Context Protocol server.
Any MCP client - Claude Desktop and Claude Code, Cursor, Gemini CLI, Codex,
a LangChain or OpenAI Agents tool loop - gets the 43 bundled datasets, any
hf:// / s3:// / tcia:// source, private folders through
pl.scan, the typed object model and the 73 catalog algorithms, on the
user's machine, without a PardesLine account. Nothing leaves the machine
except the downloads the user asks for.
Run it with pardesline mcp (stdio, the default every client speaks) or
pardesline mcp --transport streamable-http --port 8765. Needs
pip install "pardesline[mcp]".
Design rules for tools:
- every result is plain JSON (paths as strings), small enough for a context
window (lists are paged with limit / offset);
- blocking work (downloads, algorithms) runs in a worker thread so the
server keeps answering;
- the package's own rules hold: nothing is guessed, a question is returned
as data for the agent to relay to the human (scan_dataset ->
confirm_dataset), gated sources explain what the human must do.
build_server(name: str = 'pardesline')
The FastMCP server with every tool, resource and prompt registered.
main(argv: list[str] | None = None) -> int
Run the MCP server (pardesline mcp).
stdio by default; --transport streamable-http --port N serves HTTP.
skills_dir() -> Path
Where the expert skills live: inside the installed package, or the repo's skills/ folder.
impl_list_datasets(tag: str | None = None, scheme: str | None = None, medical_only: bool = False) -> dict
The bundled manifests as summaries.
Filtered by tag, source scheme or medical_only.
impl_describe_dataset(ref: str, count_samples: bool = False) -> dict
Roles, frames, units, labels, licence and citation of a dataset reference.
ref is a bundled name, an hf:// / tcia:// / s3:// URI, a manifest
path or a folder. count_samples lists the source to count samples, which may
take a while on large archives.
impl_list_samples(ref: str, limit: int = 50, offset: int = 0, filters: dict[str, str] | None = None) -> dict
Sample ids and fields of a dataset, paged with limit / offset.
filters restricts to samples whose fields match.
impl_load_sample(ref: str, sample: str | int, role: str, standard: bool = True) -> dict
Fetch one role of one sample and describe the resulting object.
Returns the path, kind, frame, units, labels, licence, provenance, the catalog
methods that accept the file and, when standard is true, standard_path:
a PLY / MetaImage rendition for raw arrays and DICOM folders.
impl_scan_dataset(ref: str) -> dict
Draft a manifest for a folder or a Hub repository and return its open questions.
impl_confirm_dataset(ref: str, answers: dict[str, Any] | None = None, save_to: str | None = None) -> dict
Answer a draft's questions (license, units, labels...) and save the manifest.
impl_export_dataset(ref: str, folder: str, take: int | None = None, overwrite: bool = False) -> dict
Write a dataset (or its first take samples) as standard files.
PLY / MetaImage with sidecars and a manifest for the folder.
impl_list_methods(extension: str | None = None, category: str | None = None) -> dict
The catalog methods, optionally those accepting a file extension or in a category.
impl_describe_method(method: str) -> dict
One method with every parameter.
Type, default, range, description, rationale and validation regime.
impl_run_method(method: str, input_path: str, moving_path: str | None = None, params: dict[str, Any] | None = None, output_path: str | None = None, timeout: float = 600) -> dict
Run a catalog method on a file with validated parameters.
Returns the output path and the recorded provenance step.
impl_describe_object(path: str, layout: dict[str, Any] | None = None, kind: str | None = None) -> dict
Kind, frame, units, labels, sidecar and accepted methods of any file.
Raw arrays (.npy, .bin) need a layout.
impl_to_mesh(path: str, label: int | None = None, isovalue: float | None = None, output_path: str | None = None) -> dict
A label map or volume to a surface.
label gives the exact boundary of one label, isovalue a level set.
impl_list_skills() -> dict
The expert skills shipped with the package: name, description and files.
impl_get_skill(name: str, file: str | None = None) -> dict
The text of a skill's SKILL.md or of one of its files.
Paths outside the skill folder are refused.