MCP Eval Runner

npm mcp-eval-runner package

A standardized testing harness for MCP servers and agent workflows. Define test cases as YAML fixtures (steps → expected tool calls → expected outputs), run regression suites directly from your MCP client, and get pass/fail results with diffs — without leaving Claude Code or Cursor.

Tool reference | Configuration | Fixture format | Contributing | Troubleshooting | Design principles

Key features

• YAML fixtures: Test cases are plain files in version control — diffable, reviewable, and shareable.
• Two execution modes: Live mode spawns a real MCP server and calls tools via stdio; simulation mode runs assertions against expected_output without a server.
• Composable assertions: Combine output_contains, output_not_contains, output_equals, output_matches, schema_match, tool_called, and latency_under per step.
• Step output piping: Reference a previous step's output in downstream inputs via {{steps.<step_id>.output}}.
• Regression reports: Compare the current run to any past run and surface what changed.
• Watch mode: Automatically reruns the affected fixture when files change.
• CI-ready: Includes a GitHub Action for running evals on every config change.

Your first prompt

Create a file at evals/smoke.yaml. Use live mode (recommended) by including a server block:

name: smoke
description: "Verify eval runner itself is working"
server:
  command: node
  args: ["dist/index.js"]
steps:
  - id: list_check
    description: "List available test cases"
    tool: list_cases
    input: {}
    expect:
      output_contains: "smoke"

Then enter the following in your MCP client:

Run the eval suite.

Your client should return a pass/fail result for the smoke test.

Fixture format

Fixtures are YAML (or JSON) files placed in the fixtures directory. Each file defines one test case.

Top-level fields

server block (live mode)

server:
  command: node # executable to spawn
  args: ["dist/index.js"] # arguments
  env: # optional environment variables
    MY_VAR: "value"

When server is present the eval runner spawns the server as a child process, connects via MCP stdio transport, and calls each step's tool against the live server.

steps array

Each step has the following fields:

Execution modes

Live mode — fixture has a server block:

• The server is spawned and each step calls the named tool via MCP stdio.
• Assertions run against the real tool response.
• Errors from the server cause the step (and by default the case) to fail immediately.

Simulation mode — no server block:

• No server is started.
• Each step's output is taken from expected_output (or empty string if absent).
• Assertions run against that static output.
• Useful for authoring and CI dry-runs, but output_contains assertions will always fail if expected_output is not set.

Assertion types

All assertions go inside a step's expect block:

expect:
  output_contains: "substring" # output includes this text
  output_not_contains: "error" # output must NOT include this text
  output_equals: "exact string" # output exactly matches
  output_matches: "regex pattern" # output matches a regular expression
  tool_called: "tool_name" # verifies which tool was called
  latency_under: 500 # latency in ms must be below this threshold
  schema_match: # output (parsed as JSON) matches JSON Schema
    type: object
    required: [id]
    properties:
      id:
        type: number

Multiple assertions in one expect block are all evaluated; the step fails if any assertion fails.

Step output piping

Reference the output of a previous step in a downstream step's input using {{steps.<step_id>.output}}:

steps:
  - id: search_step
    tool: search
    input:
      query: "mcp eval runner"
    expected_output: "result: mcp-eval-runner v1.0"
    expect:
      output_contains: "mcp-eval-runner"

  - id: summarize_step
    tool: summarize
    input:
      text: "{{steps.search_step.output}}"
    expected_output: "Summary: mcp-eval-runner v1.0"
    expect:
      output_contains: "Summary"

Piping works in both live mode and simulation mode.

Note on create_test_case

Fixtures created with the create_test_case tool do not include a server block. They always run in simulation mode. To use live mode, add a server block manually to the generated YAML file.

Tools

Running

• run_suite — execute all fixtures in the fixtures directory; returns a pass/fail summary
• run_case — run a single named fixture by name
• list_cases — enumerate available fixtures with step counts and descriptions

Authoring

• create_test_case — create a new YAML fixture file (simulation mode; no server block)
• scaffold_fixture — generate a boilerplate fixture with placeholder steps and pre-filled assertion comments

Reporting

• regression_report — compare the current fixture state to the last run; surfaces regressions and fixes
• compare_results — diff two specific runs by run ID
• generate_html_report — generate a single-file HTML report for a completed run

Operations

• evaluate_deployment_gate — CI gate; fails if recent pass rate drops below a configurable threshold
• discover_fixtures — discover fixture files across one or more directories (respects FIXTURE_LIBRARY_DIRS)

Design principles

• No mocking: Live mode evals run against real servers. Correctness is non-negotiable.
• Fixtures are text: YAML/JSON in version control; no proprietary formats or databases.
• Dogfood-first: The eval runner's own smoke fixture tests the eval runner itself.

Verification

Before publishing a new version, verify the server with MCP Inspector to confirm all tools are exposed correctly and the protocol handshake succeeds.

Interactive UI (opens browser):

npm run build && npm run inspect

CLI mode (scripted / CI-friendly):

# List all tools
npx @modelcontextprotocol/inspector --cli node dist/index.js --method tools/list

# List resources and prompts
npx @modelcontextprotocol/inspector --cli node dist/index.js --method resources/list
npx @modelcontextprotocol/inspector --cli node dist/index.js --method prompts/list

# Call a tool (example — replace with a relevant read-only tool for this plugin)
npx @modelcontextprotocol/inspector --cli node dist/index.js \
  --method tools/call --tool-name list_cases

# Call a tool with arguments
npx @modelcontextprotocol/inspector --cli node dist/index.js \
  --method tools/call --tool-name run_case --tool-arg name=smoke

Run before publishing to catch regressions in tool registration and runtime startup.

MCP Registry & Marketplace

This plugin is available on:

• MCP Registry
• MCP Market

Search for mcp-eval-runner.

Requirements

• Node.js v22.5.0 or newer.
• npm.

Getting started

Add the following config to your MCP client:

{
  "mcpServers": {
    "eval-runner": {
      "command": "npx",
      "args": ["-y", "mcp-eval-runner@latest"]
    }
  }
}

By default, eval fixtures are loaded from ./evals/ in the current working directory. To use a different path:

{
  "mcpServers": {
    "eval-runner": {
      "command": "npx",
      "args": ["-y", "mcp-eval-runner@latest", "--fixtures=~/my-project/evals"]
    }
  }
}

MCP Client configuration

Amp · Claude Code · Cline · Cursor · VS Code · Windsurf · Zed

Configuration

--fixtures / --fixtures-dir

Directory to load YAML/JSON eval fixture files from.

Type: string Default: ./evals

--db / --db-path

Path to the SQLite database file used to store run history.

Type: string Default: ~/.mcp/evals.db

--timeout

Maximum time in milliseconds to wait for a single step before marking it as failed.

Type: number Default: 30000

--watch

Watch the fixtures directory and rerun the affected fixture automatically when files change.

Type: boolean Default: false

--format

Output format for eval results.

Type: string Choices: console, json, html Default: console

--concurrency

Number of test cases to run in parallel.

Type: number Default: 1

--http-port

Start an HTTP server on this port instead of stdio transport.

Type: number Default: disabled (uses stdio)

Pass flags via the args property in your JSON config:

{
  "mcpServers": {
    "eval-runner": {
      "command": "npx",
      "args": ["-y", "mcp-eval-runner@latest", "--watch", "--timeout=60000"]
    }
  }
}