mcpatom

A minimal Python MCP Server SDK.

From Greek atomos: indivisible. The whole library is a single file, mcpatom.py, consumable two ways:

• Copy the file into your project. No dependency, no lockfile entry.

• Install the package: pip install mcpatom - the file itself is the installed module.

Scope

• Tools, resources, and prompts, over stdio or streamable HTTP.

• Protocol versions 2026-07-28 (stateless, per-request _meta) and the initialize-handshake versions 2025-11-25 and 2025-06-18, served side by side.

• stdlib-only Python >= 3.10, no dependencies, ever.

Everything else (sampling, elicitation, subscriptions, auth, etc.) is deliberately omitted.

Related MCP server: bare-mcp

Usage

from mcpatom import Server

srv = Server("my-server")


@srv.tool
def greet(name: str, excited: bool = False) -> str:
    """Return a greeting for the given name."""
    return f"Hello, {name}{'!' if excited else '.'}"


@srv.resource("data://motd")
def motd() -> str:
    """Message of the day."""
    return "Be indivisible."


@srv.prompt
def haiku(topic: str) -> str:
    """Ask for a haiku."""
    return f"Write a haiku about {topic}."


srv.serve_stdio()

For streamable HTTP instead of stdio, end with:

srv.serve_http(8388)  # http://127.0.0.1:8388/mcp

serve_http accepts only localhost (and the bind host) as Origin/Host; allowed_origins=["*.example.com"] replaces that list, ["*"] allows any origin, and host="0.0.0.0" requires it.

More runnable servers in examples/.

Schemas and return types

Each function's name, docstring, and annotations become the tool's name, description, and schema; @srv.tool(name=, description=, input_schema=, output_schema=) override generation, and extra= merges raw fields into the listing. Server() also takes version= and instructions=. Parameters may be annotated with any of these:

• str

• int

• float

• bool

• list[X] (or bare list)

• Literal[...]

• TypedDict

• X | None of any of the above

Annotated[X, "text"] adds a description the model sees. A parameter's default is published in the schema when it is JSON-representable; a None default is omitted unless null is a selectable Literal value.

A tool may return:

• str - one text block

• dict - JSON text plus structuredContent; a TypedDict return annotation publishes the matching outputSchema

• Image(data, mime_type) / Audio(data, mime_type) - one binary block (raw bytes in, base64 on the wire)

• a tuple - one block per item: str text, Image/Audio media, a dict whose "type" names a spec block type verbatim (e.g. resource_link)

• None - empty content; any other JSON value - text

Raising is the error API: any exception becomes isError content the model can read and correct.

Resources return str (text) or bytes (base64 blob).

Prompts take only str arguments and return a str user message, or a list of message dicts passed through verbatim.

Wiring it up

claude mcp add my-server -- /abs/path/.venv/bin/python /abs/path/server.py

or in any mcpServers config:

{"mcpServers": {"my-server": {"command": "/abs/path/.venv/bin/python", "args": ["/abs/path/server.py"]}}}

Smoke test without a client:

echo '{"jsonrpc":"2.0","id":1,"method":"server/discover"}' | python server.py
# {"jsonrpc":"2.0","id":1,"result":{"resultType":"complete","_meta":{...},"supportedVersions":[...],"capabilities":{"tools":{}},"ttlMs":0,"cacheScope":"private"}}