README.md

schema-validators

Five ways to type a tool parameter so MCPServer derives the JSON-Schema inputSchema and validates arguments before your handler runs: a pydantic BaseModel, a TypedDict, a @dataclass, a bare dict[str, Any], and a pydantic model built at runtime with create_model. The client lists the tools, resolves each who schema, and round-trips a call.

Run it

# stdio (default — the client spawns the server as a subprocess)
uv run python -m stories.schema_validators.client

# HTTP — the client self-hosts the server on a free port, runs, then tears it down
uv run python -m stories.schema_validators.client --http
# same, against the lowlevel-API server variant
uv run python -m stories.schema_validators.client --http --server server_lowlevel

What to look at

• client.py main — the body opens with async with Client(target, mode=mode) as client:. target is anything Client accepts (an in-process server, a transport, or an HTTP URL); the entry point picks it, the story constructs it.
• server.py — who.name vs who["name"]: pydantic and dataclass parameters arrive as instances (attribute access); TypedDict and dict[str, Any] arrive as plain dicts.
• server.py greet_dynamic — the parameter type is not written in the source at all: create_model builds it at runtime from a JSON Schema dict (the shape you'd get from OpenAPI, a config file, or a DB row), then hands it to @mcp.tool() like any BaseModel; the published schema is identical to the greet_pydantic variant. A create_model result is opaque to static type checkers, so a TYPE_CHECKING branch aliases it to a declared model of the same shape — the runtime still uses the dynamic class.
• client.py — the listed inputSchema for the four typed variants (including the runtime create_model one, whose schema matches greet_pydantic) nests a $defs/$ref object with a name property; greet_dict publishes only {"type": "object", "additionalProperties": true} — no field validation.
• server_lowlevel.py — the same schemas written by hand. There is no reflection layer at this tier; you author JSON Schema and unpack params.arguments yourself.

Caveats

• Pydantic emits local #/$defs/ references for nested models. The SDK does not dereference network $refs (SEP-2106 MUST NOT); only same-document refs are resolved during validation.
• PersonTD is total=True, so its nested schema requires both name and title; the BaseModel and @dataclass variants default title="friend", so only name is required there. Use typing.NotRequired[...] to mark optional TypedDict fields.

Spec

Tools — input schema

See also

tools/ (output schema → structuredContent), error_handling/ (what happens when validation fails).