title: "The MCP server" description: "Run and use pardesline as a local MCP server (pardesline mcp, pip install "pardesline[mcp]"): the stdio / streamable-http transports, the configuration"

The MCP server¶

The pardesline MCP server¶

pip install "pardesline[mcp]"        # or [all]
pardesline mcp                        # stdio: what every MCP client spawns
pardesline mcp --transport streamable-http --port 8765   # http://127.0.0.1:8765/mcp

Everything runs on the user's machine, without a PardesLine account. The only network traffic is the downloads the user asks for (one sample at a time). Logs go to stderr (PARDESLINE_LOG=DEBUG); stdout is the protocol channel.

Connecting a client¶

Client	Configuration
Claude Code	`claude mcp add pardesline -- pardesline mcp` (or the repo's `.mcp.json`; the plugin declares the server too)
Claude Desktop	`claude_desktop_config.json`: `"mcpServers": {"pardesline": {"command": "pardesline", "args": ["mcp"]}}`
Cursor	`.cursor/mcp.json`: same `mcpServers` block
Gemini CLI	`~/.gemini/settings.json`: `"mcpServers": {"pardesline": {"command": "pardesline", "args": ["mcp"]}}`
Codex CLI	`~/.codex/config.toml`: `[mcp_servers.pardesline] command = "pardesline" args = ["mcp"]`
Any HTTP client	`pardesline mcp --transport streamable-http --port 8765`, endpoint `http://127.0.0.1:8765/mcp`

When pardesline is not on PATH, use python -m pardesline.mcp_server as the command (same arguments).

Tools (all return JSON; errors are `{"error", "message", "code"}`)¶

Tool	What it does
`list_datasets(tag?, scheme?, medical_only?)`	the 43 bundled manifests: name, source, licence, roles
`describe_dataset(ref, count_samples?)`	roles, licence, citation, layout of a bundled name or any `hf://`, `s3://`, `tcia://`, folder
`list_samples(ref, limit?, offset?, filters?)`	paged sample ids, roles, fields; nothing downloaded
`load_sample(ref, sample, role, standard?)`	downloads ONE file or DICOM series; returns `path`, `kind`, `frame`, `units`, `labels`, `license`, `provenance`, accepted `methods`, and `standard_path` (PLY / MetaImage) for raw arrays and DICOM folders
`scan_dataset(ref)`	draft manifest + `questions` for a private folder or unknown source
`confirm_dataset(ref, answers, save_to?)`	applies the human's answers, loads, saves the manifest
`export_dataset(ref, folder, take?, overwrite?)`	standard copy with sidecars and a manifest
`list_methods(extension?, category?)`, `describe_method(method)`	the catalog
`run_method(method, input_path, moving_path?, params?, output_path?, timeout?)`	one algorithm on a local file, parameters validated first
`describe_object(path, layout?, kind?)`	what a local 3D file knows about itself
`to_mesh(path, label?, isovalue?, output_path?)`	exact surface of a label map, or an isosurface

Resources: pardesline://datasets (JSON list), pardesline://manifest/{name} (one manifest), pardesline://catalog. Prompt: start_here.

The question contract (for the agent)¶

scan_dataset and a premature confirm_dataset return {"error": "NeedsConfirmation", "questions": [{"field": "roles.sweep.layout", "message": ...}], "how_to_answer": ...}. Relay the questions to the human, collect the answers, call confirm_dataset(ref, answers) once. Answer keys: license (SPDX id or free text such as "internal, no redistribution"), pairing, units and frame (one value or {role: value}), labels ({role: {"1": "liver"}}), types ({role: "points" | "mesh" | "volume" | "labels" | "image"}), layout ({role: {"dtype", "shape", "columns", "key"}}). Never invent a licence or a layout; say "unknown" and let the human decide. Gated sources return a FetchError naming the step the human must take (accept terms, set HF_TOKEN).

Typical flow¶

list_datasets(medical_only=true) or describe_dataset("tcia://Pancreas-CT").
list_samples(ref, filters={"split": "Tr"}).
load_sample(ref, "la_007", "mask") -> path, labels={"1": "left_atrium"}.
to_mesh(path, label=1) -> surface in LPS mm, or run_method("pyvista_mesh_processing.pv_decimate", path, params={"target_reduction": 0.5}).
Tell the user the licence (license_text) before they redistribute anything.

Implementation facts¶

src/pardesline/mcp_server.py, FastMCP from the official mcp SDK (>= 1.2). Tool bodies are synchronous (impl_* functions, testable without a server) and run in a worker thread; build_server() registers them. Tests: tests/test_mcp_server.py (in-process calls and a stdio round trip with the real client).

_{Source: skills/mcp-server/SKILL.md, one of the twenty skills served by pardesline mcp; published on 2026-10-14.}