Is Your MCP Server Good?

These are the examples from Is Your MCP Server Good? Building MCP Servers for Agents. They use the Prefect API to explore how tool design affects the information an agent has to read and the work it can accomplish.

You can compare an MCP generated from OpenAPI with a smaller, hand-written server, try code mode, and see what happens when middleware removes empty fields from responses. The examples use FastMCP. For background, see Stop converting your REST APIs to MCP.

Try the examples

Install uv, then start a local Prefect server:

uvx prefect server start

Leave it running. In another terminal, compare the tool listings:

uvx --from git+https://github.com/PrefectHQ/is-your-mcp-server-good context-tax

To compare a response before and after trimming nulls and empty collections:

uvx --from git+https://github.com/PrefectHQ/is-your-mcp-server-good middleware-trim

The examples connect to http://localhost:4200/api. Run a few Prefect flows against that server first so the response comparison has data to work with.

What the examples compare

server	approach
one-to-one	Generates a tool for each operation in the Prefect OpenAPI spec.
trimmed	Uses the generated tools, with middleware that removes null values and empty collections from responses.
code-mode	Wraps the generated server in FastMCP's CodeMode transform, exposing three tools for discovering and calling its operations.
prefect-designed	Runs the separately maintained Prefect MCP server, whose tools are written for common Prefect workflows.

context-tax reports the number of tools and the size of their serialized listing for the generated, hand-written, and code-mode servers. middleware-trim compares the serialized response to a request for five flow runs.

Here is output recorded on September 3, 2026:

one-to-one      187 tools  1,650,912 bytes  ~412,728 tokens
designed         14 tools     37,008 bytes  ~  9,252 tokens
code mode         3 tools      2,362 bytes  ~    590 tokens

raw            10,312 bytes  ~2,578 tokens  for 5 flow runs
trimmed         6,444 bytes  ~1,611 tokens  for 5 flow runs

The scripts estimate tokens by dividing the serialized length by four. These numbers describe the listings and responses in that run; they aren't a bill for an agent session. Results vary with package versions, the API schema, and your flow data. With code mode, the agent retrieves further schemas as it works, so the initial listing doesn't include everything it may read later.

Try them with an agent

The .mcp.json file configures all four servers as stdio commands. Clone this repository and open it in Claude Code, or adapt those entries for your MCP client. Enable one server at a time to compare how it handles the same request, such as “Why did my last flow run fail?” You'll need a failed run in the local Prefect server for that example.

Read the agent's tool calls and answer alongside the size measurements. Check whether it found the relevant run, obtained enough evidence to explain the failure, and gave you an answer you could use. A smaller listing can help, but it doesn't establish that the tools are useful or the answer is correct.

The generated servers cache the OpenAPI spec under ~/.cache/mcp-server-demos. After a successful fetch, they can start from that copy if the API is unavailable. Tool calls still need the API running.

License

Apache 2.0