FORGE-fastmcp

FastMCP — Build & Deploy MCP Servers in Python

Current stable: fastmcp[tasks]==3.4.2 (PyPI latest as of 2026-06-12). arifOS federation target: all Python MCP organs run FastMCP 3.4.2. arifOS convention: streamable_http transport, Pydantic v2 output schemas, fastmcp[tasks] extra.

When to Use

• Create a new MCP server in Python.
• Wrap an API, database, or CLI as MCP tools.
• Add tools, resources, prompts, or tasks to an existing MCP server.
• Deploy an MCP server as an HTTP endpoint behind Caddy/Cloudflare.
• Test an MCP server locally before wiring it into Codex, Claude Code, Cursor, or another client.

Quick Start

1. Install

# Runtime / production (use uv in federation repos)
uv add "fastmcp[tasks]==3.4.2"

# System CLI (af-forge)
pip install --break-system-packages "fastmcp[tasks]==3.4.2"
fastmcp --version   # should print 3.4.2

2. Scaffold

fastmcp scaffold \
  --template api_wrapper \
  --name "My API" \
  --output ./my_server.py

Templates: api_wrapper, database_server, file_processor.

3. Implement Tools

from fastmcp import FastMCP

mcp = FastMCP("my-server")

@mcp.tool()
def search_products(query: str, limit: int = 10) -> list[dict]:
    """Search product catalog."""
    ...

4. Test Locally

fastmcp inspect my_server.py:mcp
fastmcp list my_server.py:mcp --json
fastmcp call my_server.py:mcp search_products query=oil --json

5. Deploy

# Local HTTP test
fastmcp run my_server.py:mcp --transport streamable_http --host 127.0.0.1 --port 8000

# Production: bind to 127.0.0.1 and expose via Caddy/Cloudflare Tunnel
python my_server.py  # if the script calls mcp.run(...)

6. Wire into a Client

# Codex / Claude Code / Cursor / Gemini CLI — use the public HTTPS endpoint
codex mcp add my-server \
  --type http \
  --url https://my-server.arif-fazil.com/mcp \
  --description "Product search API"

arifOS Federation Tool Naming

Output Schema Convention (arifOS)

All tools return Pydantic v2 BaseModel instances or plain dicts:

from pydantic import BaseModel

class CustomerOutput(BaseModel):
    id: str
    name: str
    segment: str
    confidence: float

@mcp.tool()
def get_customer(customer_id: str) -> CustomerOutput:
    """Fetch customer by ID with confidence score."""
    ...

HTTP Transport (arifOS Standard)

All federation Python MCP servers use the Streamable HTTP transport:

mcp.run(transport="streamable_http", host="127.0.0.1", port=8000)

• Bind to 127.0.0.1 internally.
• Let Caddy 2 terminate TLS; use Cloudflare Tunnel for public MCP subdomains.
• Do not expose the raw backend port publicly.

Background Tasks (fastmcp[tasks])

Use the tasks extra for long-running operations with progress reporting:

from fastmcp import Context

@mcp.tool()
async def ingest_large_file(path: str, ctx: Context) -> dict:
    """Ingest a large file with progress updates."""
    await ctx.report_progress(0, 100)
    ...
    await ctx.report_progress(100, 100)
    return {"status": "done"}

FastMCP Apps (3.4+)

FastMCP 3.4 introduces Apps — interactive UIs rendered in the conversation:

• FastMCPApp / @mcp.app() for managed UI wiring.
• Prefab providers: approval, choice, file_upload, form.
• Generative UI: let the LLM build custom Prefab UIs on the fly.

Use these for human-in-the-loop gates inside MCP clients that support them. Federation organs remain headless by default; Apps are optional UI sugar.

MCP Apps Hardening (SEP-1865 + ChatGPT 2026)

Authoritative contract: /root/forge_work/2026-07-20/GEOX-CHATGPT-MCP-GUI-BLUEPRINT.md (F13 sovereign directive, 2026-07-20). Applies to any federation server that exposes tools to ChatGPT / MCP Apps hosts. GEOX is the first consolidation target.

1. Explicit 4-annotation taxonomy on every public tool

Set ALL FOUR annotations explicitly on every public tool — wire format is camelCase. Annotation silence inherits risky defaults: readOnlyHint=false, destructiveHint=true, idempotentHint=false, openWorldHint=true. A tool that forgets annotations looks dangerous-by-default to the host. Annotations are hints, not enforcement — server-side auth is still required.

Display-name precedence: title → annotations.title → name.

2. outputSchema families

Treat outputSchema as mandatory for every tool returning structuredContent:

• Declare per tool, grouped into named families (GEOX families: GeoxStatusResult, GeoxArtifactResult, GeoxWellResult, GeoxSeismicResult, GeoxMapResult, GeoxProspectResult, GeoxClaimResult, GeoxEvidenceResult).
• additionalProperties: false, required fields pinned.
• Never put secrets, trace IDs, or dense arrays in structuredContent — it is model-visible. Dense data goes into a ui:// resource or artifact reference.

Reference shape (condensed from the blueprint):

{
  "type": "object",
  "title": "GeoxWellResult",
  "additionalProperties": false,
  "required": ["status", "well_ref", "summary", "artifacts"],
  "properties": {
    "status": { "type": "string", "enum": ["ok", "partial", "failed"] },
    "well_ref": { "type": "string" },
    "summary": { "type": "string", "description": "≤3 sentence model-facing summary" },
    "artifacts": {
      "type": "array",
      "items": {
        "type": "object",
        "additionalProperties": false,
        "required": ["kind", "ref"],
        "properties": {
          "kind": { "type": "string", "enum": ["log_panel", "segy", "table"] },
          "ref": { "type": "string", "description": "ui:// or artifact URI" }
        }
      }
    }
  }
}

3. Tool _meta wiring for UI tools

Every UI-bound tool needs, on the tool definition:

• _meta.ui.resourceUri — primary binding, e.g. ui://geox/workbench-v1.html (version URIs; they are host cache keys).
• _meta["openai/outputTemplate"] — the ChatGPT alias, emitted alongside (not instead of) the primary key, with the same URI value.
• Optional _meta["openai/toolInvocation/invoking"] and _meta["openai/toolInvocation/invoked"] status strings.
• The served resource MIME MUST be text/html;profile=mcp-app (hosts only enable the app bridge for that profile).

In FastMCP, pass these via the @mcp.tool(meta={...}) / annotations kwargs if your pinned version supports arbitrary tool meta. If FastMCP cannot emit arbitrary _meta on tool definitions, the sanctioned federation workaround is bridge-side injection — see GEOX at src/geox_mcp/tools/mcp_apps_bridge.py, whose enrich_response() injects _meta.ui.resourceUri into tool responses at the MCP boundary (24 tools covered at recon time). Mirror that pattern rather than forking FastMCP internals.

4. Per-tool securitySchemes

Declare per tool for ChatGPT: noauth for public read-only tools, oauth2 + scopes for anything else. GEOX scope map (blueprint): geox.read, geox.compute, geox.ingest, geox.export, geox.claim.write, geox.seal — withhold geox.seal from v1. OAuth 2.1 resource-server rules (RFC 9728 PRM, WWW-Authenticate on 401, one canonical resource identity across endpoint/PRM/OAuth param/audience) belong to the server auth layer, but the per-tool scheme declaration lives next to the tool definition so the host can plan step-up auth.

5. Tool-originated errors

Tool failures return a CallToolResult with isError: true — never protocol-level crashes. A structured error result lets the model read the failure and self-correct; a transport crash kills the whole conversation turn. In FastMCP: catch domain exceptions inside the tool body and raise/return an error result with a concise, actionable message instead of letting the exception escape.

CLI Cheat Sheet

fastmcp --version
fastmcp inspect server.py:mcp
fastmcp list server.py:mcp --json
fastmcp call server.py:mcp tool_name arg1=val1 --json
fastmcp run server.py:mcp --transport streamable_http --host 127.0.0.1 --port 8000
fastmcp generate-cli server.py:mcp -o ./mycli

Quality Checklist

• fastmcp inspect <server.py:mcp> succeeds.
• fastmcp list <server> --json returns all expected tools.
• At least one fastmcp call succeeds against a representative tool.
• Tool names follow {service}_{action}_{resource} pattern.
• Output schemas are Pydantic v2 models or typed dicts.
• No hardcoded secrets (use env vars / SOPS).
• Health endpoint at /health when using HTTP transport.
• Bound to 127.0.0.1 in production.
• (ChatGPT/MCP Apps) All 4 annotations explicit on every public tool — no annotation silence.
• (ChatGPT/MCP Apps) outputSchema declared for every tool returning structuredContent, additionalProperties: false.
• (ChatGPT/MCP Apps) UI tools carry _meta.ui.resourceUri + openai/outputTemplate alias (bridge injection if FastMCP can't emit _meta).

Troubleshooting

# Wrong / missing fastmcp
pip install --break-system-packages "fastmcp[tasks]==3.4.2"

# Import error
python3 -c "from fastmcp import FastMCP; import fastmcp; print(fastmcp.__version__)"

# Inspect fails
fastmcp inspect my_server.py:mcp --verbose

# HTTP transport not working — check transport spelling
fastmcp run my_server.py:mcp --transport streamable_http --host 127.0.0.1 --port 8000

# FastMCP 2 → 3 migration
# See https://gofastmcp.com/getting-started/upgrading/from-fastmcp-2.md

Federation Version Alignment

If a repo lock drifts below 3.4.2, run uv lock --upgrade-package fastmcp and re-run the test suite.