cellpy-mcp

An MCP server for cellpy. It lets an agent — or a chat window — load battery cells, collect them into frames, render figures, export data, look up how any cellpy call works, and set up a batch project. Without writing any Python.

pip install cellpy-mcp
cellpy mcp install        # register it with your chat client

Then restart the client and ask it to plot something.

Who it is for

• People building on cellpy — a GUI, a script that automates cell handling. The tools are the useful API, and describe_api gives real signatures from the version you have installed.
• People who just want to ask. Open a chat window, point it at a file, ask for capacity versus cycle. The prompts below are pickable starting points, so you do not have to know what to ask for.
• People who would rather not use a terminal. new_project is cellpy new without the command line.

Nothing is hosted. Under stdio your chat client starts the server itself, on your machine, reading your files — there is no service to sign up for and nothing leaves the machine except what you paste into the chat.

What it can do

Cells and figures:

The cellpy API — "how does this call work, and what are its arguments":

Batch templating:

Prompts: analyse_cell, start_batch_project, explain_call.

Registering with your client

cellpy mcp install writes the block for you. Pass --client for anything other than Claude Desktop:

cellpy mcp install                     # Claude Desktop (default)
cellpy mcp install --client cursor
cellpy mcp install --client vscode
cellpy mcp install --list-clients      # where each one keeps its config

Add --dry-run to see the target first. Restart the client afterwards — none of them re-read the file while running.

VS Code names the key servers, not mcpServers. If you edit that file by hand, this is the mistake to avoid: the wrong key parses, saves, and does nothing at all.

Claude Code

Claude Code is not registered by editing a file — its servers live in ~/.claude.json alongside your sign-in session and per-project trust decisions, or in a project-scoped .mcp.json. It has a command that handles scopes properly, so use that:

claude mcp add cellpy --env CELLPY_MCP_ROOT=/path/to/cells -- python -m cellpy_mcp

cellpy mcp install --list-clients prints that line filled in with the interpreter and the roots for your machine.

By hand

Any client that speaks stdio can run this; it is an ordinary MCP server. The block is the same everywhere except the top-level key:

{
  "mcpServers": {
    "cellpy": {
      "command": "/path/to/python",
      "args": ["-m", "cellpy_mcp"],
      "env": { "CELLPY_MCP_ROOT": "/path/to/cells" }
    }
  }
}

Use the full path to the interpreter that has cellpy-mcp installed, not a bare python: a desktop client activates no virtualenv and inherits no shell PATH, which is the most common reason a server shows up as failed.

cellpy mcp status says which clients it can see cellpy registered with.

Where it may read and write

Everything is confined to a set of roots, and both reads and writes are checked — an unchecked write would make the read check decoration.

By default the roots are the directories cellpy already knows about (rawdatadir, cellpydatadir, outdatadir, notebookdir), because you told cellpy where your data is when you set it up. Override with CELLPY_MCP_ROOT, which takes several directories separated the way PATH separates them:

CELLPY_MCP_ROOT=/data/cells:/data/out cellpy mcp serve

Roots that are not plain local directories are dropped: rawdatadir is often scp://host/…, and containment here is pathlib-based and cannot express "inside that remote share". If cellpy has no configured paths at all, the single root is ~/cellpy_mcp — never your whole filesystem.

Four things it does on purpose

Handles, not data. Only preview_collection returns rows. A tool result goes into a model's context window, and one collected summary is ~29 kB of CSV while a raw figure can be several MB.

Results carry the traps. render returns trace_types alongside points_plotted and rows_collected, so an agent that asked for a density film can see it got histogram2d rather than lines, and one that forgot direction="both" can see it plotted 891 of 2328 rows. describe_api returns undocumented_parameters for the same reason: cellpy documents about half its arguments, and a model should know when the package never said.

It follows the docstring's own cross-references. CellpyCell.get_cap takes 23 arguments, documents none, and points at cellpy.readers.capacity_curves.get_cap — which documents 22 of 24. Following that takes argument coverage across the documented API from 51% to 72%. Only a docs site resolves those markers; nobody reading a docstring does.

One client per process. State is process-wide, deliberately: the MCP SDK does not give a tool a stable session identity, and under stdio each client spawns its own process anyway. Do not put this behind a shared HTTP endpoint as written.

Limits worth knowing

• load_cell blocks, with no progress and no cancellation, so a slow load can look like a hang to a client. This is the gap most worth closing.
• No quota and no eviction: an agent can fill the sandbox with figures, and cells stay in memory until the process exits.
• new_project has to create the project directory itself, because cellpy new prompts even with no_input=True (cellpy#990). It also downloads a cookiecutter from GitHub on first use.

Running it without cellpy's shim

cellpy mcp arrived in cellpy 2.1.3.post3, which this package requires. The same things are reachable without it:

python -m cellpy_mcp serve
python -m cellpy_mcp install --dry-run
python -m cellpy_mcp status

Background

The design, the measurements behind it, and a two-round prototype log are in cellpy#840.

MIT licensed.