A successful build leaves the MCP connection unverified
FastMCP is worth evaluating if you want to expose existing Python functions through the Model Context Protocol, or MCP. Its README describes servers, clients and interactive apps in one framework. The adoption question is whether those APIs work with your intended client and deployment. Our recorded run does not answer that question.
The README starts with an addition function decorated as a tool and says schemas, validation and documentation are generated automatically. That provides a documented entry point. It does not establish the permissions, authentication or failure behavior of a tool connected to your systems.
What happened when we ran it
Our sandbox installed FastMCP in 63 seconds and built it in 33 seconds, then the test command exited with code 4 after 13 seconds. The run took place on September 29, 2026, at commit 89b00b7. It recorded 95 dependencies and 90 MB on disk. These are source-checkout measurements, not package-install timing promises.
The command stopped while loading tests/conftest.py. The retained output shows an import of TracerProvider from opentelemetry.sdk.trace and ends with ModuleNotFoundError: No module named 'opentelemetry.sdk'. No test total was recorded. This is a collection blocker in that environment, rather than evidence of failing test assertions. The log does not establish the cause or support a specific dependency fix.
The configured sandbox had 3 CPUs, 8 GB RAM and image ghcr.io/astral-sh/uv:python3.12-bookworm. That label does not establish the actual Python or operating-system version.
The audit record reports zero advisories and zero packages; dependency coverage is not established, so this supplies no clean dependency-security finding.
The record does not establish an installed CLI check, a successful tool call, an authenticated connection or interactive app rendering. No new execution was performed for this correction.
Separate the framework from the gateway
The README places servers, clients and interactive apps in FastMCP's application layer. It describes Prefect Horizon separately as an enterprise MCP gateway, with SSO, tool-level RBAC and centralized governance in that discussion. If those controls are purchasing requirements, evaluate the relevant Horizon deployment separately. The README does not establish them as verified features of the Python package in our run.
For a framework trial, choose one existing function whose output you can check independently. Call it through the client you intend to use, then verify denied access and failure behavior. These are recommended acceptance checks, not completed tests.
Pin versions and distinguish the import failures
The installation guide warns that breaking changes may occur in minor versions and recommends exact version pins for production. Check upgrades against the clients and tools you actually use.
The guide also describes a specific pip-upgrade problem when moving from FastMCP 3.2 or earlier to 3.3 or later. It says fresh installations and uv upgrades are unaffected. That warning concerns missing FastMCP imports after an upgrade. Our log concerns a missing OpenTelemetry module while loading test configuration. The supplied evidence does not connect those failures, so the guide's reinstall instructions are not a demonstrated fix for our run.
The current README and installation extract are documentation evidence, not proof that every described API was present at commit 89b00b7. Keep the selected release and its documentation together during evaluation.
The September audit does not resolve October security notices
The supplied release notes say v4.1.3 updates certifi-backed dependencies following the reported eMudhra CA compromise. They also say FastMCP 4 uses system trust by default and require updates to operating-system, container, custom and installer trust stores: upgrading FastMCP alone is insufficient.
The v3.4.11 notes likewise require certifi 2026.10.19 or later for the HTTP stack and certifi-backed integrations, with separate custom, system and installer trust-store updates. These are attributed release requirements, not remediation verified by our sandbox.
The September 29 run predates these October notices and does not verify remediation. Before deployment, check the applicable release guidance and trust stores in your own environment. A successful build and an audit record with zero packages cannot establish that those requirements have been met.
Compare the APIs your application needs
The official MCP Python SDK supports MCP servers and clients. Its README also shows typed Python functions defining tools without handwritten schemas or protocol handling. Evaluate it when you want to build directly with the protocol project's Python APIs. Generated schemas alone do not distinguish FastMCP from that SDK.
The official MCP TypeScript SDK provides server and client libraries for TypeScript. Its README documents those libraries and separate server and client packages. Evaluate it when your application is written in TypeScript and you want to keep the integration in that language. We have no comparative setup or performance measurements for either SDK.
What the assessment supports
Setup receives 3/5 for a bounded successful installation and build with material verification gaps. For this category, 1 means setup could not complete; 2 means only part of installation and build completed; 3 means both completed but functional verification remains incomplete; 4 requires a verified CLI and representative client connection; 5 also requires verification in the intended deployment. This is an editorial rubric, not a performance benchmark.
Documentation, community support and production maturity remain unassessed numerically. The supplied documentation establishes described coverage, but its correctness was not exercised. Readable issue and release pages do not establish support responsiveness, and a source build does not establish production reliability.
Before adoption, pin a release, review its security requirements, and verify a representative tool call and denied-access case with your intended client. Authentication and deployment need their own acceptance checks.