msgspec-fastapi

An experimental FastAPI router that uses msgspec for JSON request bodies and responses. In one benchmark service, moving from Pydantic models to msgspec Structs with this router increased throughput by 68% on a route with a large JSON body and by 11% to 13% on database-backed routes; see the Benchmarks section below.

This is a prototype. Its API may change, and it relies on private FastAPI internals, so it is pinned to FastAPI 0.142.2 to 0.143.x. It is a personal project, not part of msgspec; feedback is welcome in the issues.

Installation

pip install msgspec-fastapi

Usage

import msgspec
from fastapi import FastAPI
from msgspec_fastapi import MsgspecRouter, install_openapi


class City(msgspec.Struct):
    name: str
    population: int


router = MsgspecRouter()


@router.post('/cities')
async def create_city(city: City) -> City:
    return city


app = FastAPI()
app.include_router(router)
install_openapi(app, router)

Benchmarks

The router was measured in a small reference-data service on a cloud VM with dedicated vCPUs (October 2026). Each row compares the same service on FastAPI with Pydantic models against the same service with msgspec Structs and MsgspecRouter. The database queries (SQLAlchemy ORM), event loop and server settings are the same on both sides, and there is no cache.

The first route returns a constant 16-byte response ({"accepted":100}), so the difference on this route comes mostly from request decoding and validation, and the gain is largest there. On routes that read from a database, API CPU time per response falls by about 190 to 260 µs out of roughly 2 ms, so the relative gain is smaller.

In the same service, setting response_model=None on the GET route made no detectable difference in throughput with MsgspecRouter (0.0% [-1.0%, +1.0%]). This is consistent with an instance of exactly the declared Struct class being encoded without revalidation.

The benchmark sources, the data preparation scripts and the results of all 14 comparisons, including direct asyncpg queries without the ORM, uvloop, a Redis cache, Litestar and Go, are in world-directory-bench.

Things to keep in mind:

• On the control side, FastAPI parses the body with json.loads and then validates it with Pydantic, which is FastAPI's standard path.
• The GET and search rows also include changes outside the router: database rows are converted to msgspec Structs instead of Pydantic models and, for GET, the query model is msgspec instead of Pydantic, so they measure a migration, not the router alone.
• Both sides check the same constraints, including custom per-record checks written in Python. Results for other schemas and other routes will differ.
• This is one series on one machine. The host slowed down near the end of the series: a static control was about 10% slower after the series than before it. The blocks of all comparisons ran in 8 rounds, and the lasting slowdown began during the last round. Leaving out any single round, the last one included, changes each estimate by less than 0.32 percentage points. Compare the relative effects rather than the absolute numbers.

What it does

• A msgspec body parameter (a Struct or another type msgspec can decode) is decoded from the raw request bytes by a msgspec.json.Decoder built once at route registration. A body with a non-JSON content type returns 415, malformed JSON or JSON nested too deeply returns 400, and a missing or invalid body returns 422. The 415 and 400 responses are sent before dependencies run; a 422 is sent after them, as in FastAPI, so a failing dependency such as authentication takes precedence. Bodies are decoded in msgspec strict mode: for example, the JSON string "1" is not accepted for an int field, unlike FastAPI with Pydantic.
• A body without a content type follows FastAPI's strict_content_type setting: it is rejected with 415 by default and parsed as JSON when the setting is False. FastAPI returns 422 in the first case.
• A msgspec response type, from the return annotation or response_model, is encoded with msgspec. An instance of exactly the declared Struct class is encoded without revalidation, as FastAPI does for a Pydantic response_model; other values are validated against the declared type. Nested values and type parameters of a trusted instance are not checked. A NaN or infinite value in a float field is encoded as null unless a constraint rejects it. In a Decimal field it is encoded as a string, while FastAPI with Pydantic rejects non-finite Decimal values.
• Responses of routes without a return annotation or response_model, or with -> None, are also encoded with msgspec, so the output can differ from FastAPI's jsonable_encoder (for example, bytes are base64-encoded and Decimal is encoded as a string).
• Path, query, header and cookie parameters keep FastAPI extraction, aliases and defaults, but are converted with msgspec.convert(strict=False) instead of Pydantic, so coercion differs for some inputs (for example, " 1.5" is rejected for float and "nan" is accepted for Decimal). Parameter types must be supported by msgspec; Pydantic-specific types such as AnyUrl raise TypeError at route registration.
• install_openapi adds the msgspec JSON schemas to the OpenAPI document. It is required for a valid document: call it once after including the routers, passing every MsgspecRouter that is not included directly in another MsgspecRouter.
• A Pydantic body or response model is left to FastAPI; the route parameters are still validated with msgspec.

Limitations

• Only one msgspec body parameter per route, and it cannot be combined with other FastAPI body fields. Only JSON bodies are supported: application/json, application/*+json, or a request without a content type when strict_content_type is False. Form() and File() parameters are left to FastAPI.
• When both the parameters and the body are invalid, only the parameter errors are returned; FastAPI returns both.
• Structs, alone or inside containers, are not supported as path, query, header or cookie parameters, or as Form() and File() parameters.
• response_model_include, response_model_exclude, response_model_exclude_* and response_model_by_alias=False are rejected for typed msgspec responses.

Development

uv run --group test pytest