MCP Creator

MCP Creator — FastMCP v3

Build production-ready MCP servers with FastMCP v3 (3.0.0rc2). This skill guides through research, scaffolding, implementation, testing, and deployment. All first-party output follows this repo's conventions: mcp/<name>/ directory, fastmcp.json config, exact uv workspace member, imperative voice, kebab-case naming. Reserve mcp/servers/ for machine-local third-party MCP installs; it is gitignored.

Target: FastMCP v3 rc2 — Provider/Transform architecture, 14 built-in middleware, OAuth 2.1, server composition, component versioning, structured output, background tasks, elicitation, sampling.

Input: $ARGUMENTS — the service, API, or capability to wrap as an MCP server.

Dispatch

Route $ARGUMENTS to the appropriate mode:

Consult Live Documentation

Before implementation, fetch current FastMCP v3 docs. Bundled references capture rc2 — live docs may be newer.

1. Context7 — resolve-library-id for "fastmcp", then query-docs for the topic.
2. WebFetch — https://gofastmcp.com/llms-full.txt — comprehensive LLM-optimized docs.
3. WebSearch fallback — site:gofastmcp.com <topic> for specific topics.

If live docs contradict bundled references, live docs win. Always fetch live docs first — API details shift between rc2 and stable.

Phase 1: Research & Plan

Goal: Understand the target service and design the MCP server's tool/resource/prompt inventory.

Load: references/tool-design.md (naming, descriptions, parameter design, 8 tool patterns)

1.1 Understand the Target

• Read any documentation, SDK, or API reference for the target service.
• Identify the core operations users need (CRUD, search, status, config).
• Note authentication requirements (API keys, OAuth, tokens).
• Check for existing MCP servers for this service — avoid duplicating work.

1.2 Architecture Checklist

Answer before proceeding:

• Auth needed? (API key → env var, OAuth → HTTP transport, none → stdio)
• Transport? (stdio for local, Streamable HTTP for remote/multi-client)
• Background tasks? (long-running operations → task=True, requires fastmcp[tasks])
• OpenAPI spec available? (→ OpenAPIProvider or FastMCP.from_openapi())
• Multiple domains? (→ composed servers with mount() + namespaces)

1.3 Design Tool Inventory

Plan 5-15 tools per server. For each tool, define:

1.4 Design Resources and Prompts

• Resources for static/slow-changing data (config, schemas, status). URI-addressed.
• Prompts for reusable message templates that guide LLM behavior.
• See references/fastmcp-v3-api.md §6-7 for URI patterns and prompt design.

1.5 Plan Architecture

Decide on composition strategy:

• Single server — Most cases. One FastMCP instance with all tools.
• Composed servers — Large APIs. Domain servers mounted via mount() with namespaces.
• Provider-based — Dynamic tool registration. FileSystemProvider or custom Provider.
• OpenAPI conversion — Auto-generate tools from OpenAPI spec. OpenAPIProvider or FastMCP.from_openapi().

See references/server-composition.md for patterns.

1.6 Deliverable

Produce a tool/resource/prompt inventory table before proceeding:

| Component | Type | Name | Description (brief) |
|-----------|------|------|---------------------|
| Tool | tool | search_issues | Search GitHub issues by query |
| Resource | resource | config://settings | Current server configuration |
| Prompt | prompt | summarize_pr | Summarize a pull request for review |

Phase 2: Scaffold

Goal: Create the project directory, tests, and configure dependencies.

2.1 Create Project Structure

Run python scripts/scaffold_mcp.py <name> to scaffold:

mcp/<name>/
├── server.py          # FastMCP entry point
├── pyproject.toml     # uv project config
├── fastmcp.json       # FastMCP CLI config
└── tests/
    ├── conftest.py    # Client fixture, mock Context
    └── test_server.py # Automated test suite

Customize the scaffold:

• Set server name in server.py: mcp = FastMCP("name").
• Set package name in pyproject.toml: name = "mcp-<name>".
• Update description in pyproject.toml.
• Add service-specific dependencies to both pyproject.toml and fastmcp.json.

2.2 Add Test Configuration

Add to pyproject.toml:

[tool.pytest.ini_options]
asyncio_mode = "auto"

[dependency-groups]
dev = ["pytest>=8", "pytest-asyncio>=0.25"]

Create tests/conftest.py — see references/testing.md §4 for the complete template.

2.3 Configure and Verify

• Ensure root pyproject.toml has [tool.uv.workspace] members = ["mcp/<name>"].
• Run cd mcp/<name> && uv sync to install dependencies.
• Verify: uv run python -c "from server import mcp; print(mcp.name)".

Note: python scripts/check.py validates this skill only, NOT MCP servers. Use the import check above for MCP server validation.

Phase 3: Implement

Goal: Build all tools, resources, and prompts from the Phase 1 inventory.

Load: references/fastmcp-v3-api.md (full API surface), references/tool-design.md (patterns)

3.1 Server Setup

from fastmcp import FastMCP, Context

mcp = FastMCP(
    "server-name",
    instructions="Description of what this server provides and when to use it.",
)

For shared resources (HTTP clients, DB connections), add a composable lifespan:

from fastmcp.server.lifespan import lifespan

@lifespan
async def http_lifespan(server):
    import httpx
    async with httpx.AsyncClient() as client:
        yield {"http": client}

mcp = FastMCP("server-name", lifespan=http_lifespan)
# Access in tools: ctx.lifespan_context["http"]

Combine lifespans with |: mcp = FastMCP("name", lifespan=db_lifespan | cache_lifespan)

3.2 Implement Tools

For each tool in the inventory, follow this pattern:

from typing import Annotated
from pydantic import Field
from fastmcp import Context
from fastmcp.exceptions import ToolError

@mcp.tool(
    annotations={
        "readOnlyHint": True,
        "openWorldHint": True,
    },
)
async def search_items(
    query: Annotated[str, Field(description="Search term to find items.", min_length=1)],
    limit: Annotated[int, Field(description="Max results to return.", ge=1, le=100)] = 10,
    ctx: Context | None = None,
) -> dict:
    """Search for items matching a query.

    Use this tool when you need to find items by keyword. Returns a list of
    matching items with their IDs and titles. Use the limit parameter to
    control result count. Does not search archived items.

    Returns a dictionary with 'items' list and 'total' count.
    """
    if ctx:
        await ctx.info(f"Searching for: {query}")
    try:
        results = await do_search(query, limit)
        return {"items": results, "total": len(results)}
    except ServiceError as e:
        raise ToolError(f"Search failed: {e}")

Key rules for every tool:

• Annotated[type, Field(description=...)] on EVERY parameter.
• Verbose docstring: WHAT, WHEN to use, WHEN NOT, WHAT it returns.
• ToolError for expected failures (always visible to client).
• annotations dict on every tool — at minimum readOnlyHint.
• ctx: Context | None = None for testability without MCP runtime.

See references/tool-design.md §9 for 8 complete tool patterns (sync, async+Context, stateful, external API, data processing, dependency-injected, sampling, elicitation).

3.3 Implement Resources

import json

@mcp.resource("config://settings", mime_type="application/json")
async def get_settings() -> str:
    """Current server configuration."""
    return json.dumps(settings)

@mcp.resource("users://{user_id}/profile")
async def get_user_profile(user_id: str) -> str:
    """User profile by ID."""
    return json.dumps(await fetch_profile(user_id))

See references/fastmcp-v3-api.md §6 for URI templates, query params, wildcards, and class-based resources.

3.4 Implement Prompts

from fastmcp.prompts import Message

@mcp.prompt
def summarize_pr(pr_number: int, detail_level: str = "brief") -> list[Message]:
    """Generate a prompt to summarize a pull request."""
    return [Message(
        role="user",
        content=f"Summarize PR #{pr_number} at {detail_level} detail level.",
    )]

3.5 Composition (if applicable)

For large servers, split into domain modules and compose. See references/server-composition.md.

from fastmcp import FastMCP
from .issues import issues_server
from .repos import repos_server

mcp = FastMCP("github")
mcp.mount(issues_server, namespace="issues")
mcp.mount(repos_server, namespace="repos")

3.6 Auth (if applicable)

Load references/auth-and-security.md when implementing authentication.

• stdio transport: No MCP-level auth. Use env vars for backend API keys.
• HTTP transport: OAuth 2.1 via JWTVerifier, GitHubProvider, or RemoteAuthProvider.
• Per-tool auth: @mcp.tool(auth=require_scopes("admin")).
• Dual-mode pattern: common.py with shared tools, separate auth/no-auth entry points.

Phase 4: Test

Goal: Verify all components work correctly with deterministic tests.

Load: references/testing.md (patterns, 18-item checklist)

4.1 Write Tests

Use the in-memory Client — no network, no subprocess:

import pytest
from fastmcp import Client
from server import mcp

@pytest.fixture
async def client():
    async with Client(mcp) as c:
        yield c

async def test_search_items(client):
    result = await client.call_tool("search_items", {"query": "test"})
    assert result.data is not None
    assert not result.is_error

async def test_list_tools(client):
    tools = await client.list_tools()
    names = [t.name for t in tools]
    assert "search_items" in names

4.2 Test Categories

Cover all 8 categories from references/testing.md:

1. Discovery — list_tools(), list_resources(), list_prompts() return expected names.
2. Happy path — Each tool with valid input returns expected output.
3. Error handling — Invalid input produces ToolError, not crashes.
4. Edge cases — Empty strings, boundary values, Unicode, large inputs.
5. Resources — read_resource(uri) returns correct content and MIME type.
6. Prompts — get_prompt(name, args) returns expected messages.
7. Integration — Tool chains, lifespan setup/teardown.
8. Concurrent — Multiple simultaneous calls don't interfere.

4.3 Interactive Testing

# MCP Inspector (browser-based)
fastmcp dev inspector mcp/<name>/server.py

# CLI testing
fastmcp list mcp/<name>/server.py
fastmcp call mcp/<name>/server.py search_items '{"query": "test"}'

4.4 Run Tests

cd mcp/<name> && uv run pytest -v

Phase 5: Deploy & Configure

Goal: Make the server available to MCP clients.

Load: references/deployment.md (transports, client configs, Docker)

5.1 Select Transport

5.2 Generate Client Config

Add to client config (Claude Desktop, Claude Code, Cursor):

{
  "mcpServers": {
    "server-name": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/mcp/<name>", "fastmcp", "run", "server.py"]
    }
  }
}

See references/deployment.md §7 for complete configs per client.

5.3 Validate

MCP server validation (skill check.py does NOT check MCP servers):

# Import check
uv run python -c "from server import mcp; print(mcp.name)"

# List registered components
fastmcp list mcp/<name>/server.py

# Interactive inspection
fastmcp dev inspector mcp/<name>/server.py

5.4 Quality Checklist

Before declaring the server complete:

• Every tool has Annotated + Field(description=...) on all parameters
• Every tool has a verbose docstring (WHAT, WHEN, WHEN NOT, RETURNS)
• Every tool has annotations (at minimum readOnlyHint)
• Every tool uses ToolError for expected failures
• No print() or stdout writes in any tool
• Resources have correct URI schemes and MIME types
• Tests pass: uv run pytest -v
• MCP Inspector shows all components correctly
• fastmcp.json lists all required dependencies
• pyproject.toml has correct metadata and dependencies
• No deprecated constructor kwargs (removed in rc1)
• Custom routes have manual auth if sensitive
• mask_error_details=True set for production deployment

Reference File Index

Load these files on demand during the relevant phase. Do NOT load all at once.

Critical Rules

These are non-negotiable. Violating any of these produces broken MCP servers.

1. No stdout. Never use print() or write to stdout in tools/resources/prompts. Stdout is the MCP transport. Use ctx.info(), ctx.warning(), ctx.error() for logging.

2. ToolError for expected failures. Always raise ToolError("message") for user-facing errors. Standard exceptions are masked by mask_error_details in production.

3. Verbose descriptions. Every tool needs a 3-5 sentence docstring. Every parameter needs Field(description=...). LLMs cannot use tools they don't understand.

4. Annotations on every tool. Set readOnlyHint, destructiveHint, idempotentHint, openWorldHint. Clients use these for confirmation flows and retry logic.

5. No *args or **kwargs. MCP requires a fixed JSON schema for tool inputs. Dynamic signatures break schema generation.

6. Async state access. In v3, ctx.get_state() and ctx.set_state() are async — always await them.

7. URI schemes required. Every resource URI must have a scheme (data://, config://, users://). Bare paths fail.

8. Test deterministically. Use in-memory Client(mcp), not manual prompting. Tests must be repeatable and automated.

9. Module-level mcp variable. The FastMCP instance must be importable at module level. fastmcp run imports server:mcp by default.

10. Secrets in env vars only. Never hardcode API keys. Never accept tokens as tool parameters. Load from environment, validate on startup.

Quick Reference

Load references/quick-reference.md for the complete quick reference with minimal examples for server, tool, resource, prompt, lifespan, test, and run commands.

Canonical Vocabulary

Use these terms consistently. Do not invent synonyms.

Validation Contract

Before declaring changes to this skill complete, run:

python skills/mcp-creator/scripts/check.py

Completion criteria:

1. scripts/check.py exits 0.
2. MCP scaffold smoke-check: python scripts/scaffold_mcp.py --dry-run example-server.