MCP Factory

tests

The test count is verifiable below (python -m pytest tests/ → 348 passed, 8 skipped) and enforced in CI by scripts/check_readme_counts.py, which fails the build if this README's counts drift from the live suite.

60-Second Quickstart

From PyPI (registry users):

pip install jaimenbell-mcp-factory
mcp-factory-hub --serve
# equivalent: python -m mcp_factory --serve

From a git checkout (contributors): see the python hub_server.py ... examples throughout this README — hub_server.py at the repo root is a backward-compat wrapper around the same mcp_factory.cli module the console script runs, so behavior is identical either way.

The manifest-driven engine behind the MCP Integration Sprint. Write one mcp.yaml for a bot repo and the factory generates the server stub and the ~/.claude.json entry; run the hub and it serves every bot's tools through a single MCP endpoint.

The SDK wrapper is the easy part. What makes an MCP server safe to put in front of a real internal tool — scoped auth/env, fail-soft error handling, validated manifests, a collision-safe registry, and a real test suite — is the engineering this engine is built around. That same production layer is hand-built per engagement; the factory scaffolds it, it doesn't fake it.

Browse before you reply

This repo is public so you can verify the discipline instead of taking my word for it. Every claim below maps to a file you can open:

Honesty rails: 348 is the real, reproducible count on a clean checkout — the same number the public CI run produces and gates on. mcp-factory generates the scaffold and runs the hub — it does not "generate the production server" or carry any client/CI claims. The hardened production layer (per-tool auth boundaries, the full failure set, two-axis version-pinning) is built per engagement on top of this engine. That applies to both Python scaffold styles below — see "Two Python styles" for exactly what the fastmcp variant does and doesn't add on top of that baseline.

Quick Start

Factory mode (generate config from one manifest)

# Reference an existing MCP server (no code generated — just the config entry)
python hub_server.py --manifest examples/fleet_health.yaml

# Scaffold a new MCP server from scratch
python hub_server.py --manifest my_bot/mcp.yaml --output-dir my_bot/

# Dry run — preview without writing
python hub_server.py --manifest my_bot/mcp.yaml --dry-run

# Self-verify: compare factory output to live ~/.claude.json entry
python hub_server.py --manifest examples/fleet_health.yaml --verify ~/.claude.json

Output always goes to ~/.claude.json.factory-test by default — never to the live ~/.claude.json. Copy entries manually after review.

Scan mode (batch-register all bots)

# Dry-run diff: show what would change in ~/.claude.json
python hub_server.py --scan C:\path\to\projects

# Apply: write ~/.claude.json after backing it up
python hub_server.py --scan C:\path\to\projects --apply

# Force-update entries already registered
python hub_server.py --scan C:\path\to\projects --apply --force

--scan discovers all projects/*/mcp.yaml files, validates each, and diffs them against the current ~/.claude.json. Default root is C:\path\to\projects. With --apply, a timestamped backup is created at ~/.claude.json.scan-backup-<timestamp> before writing.

Skip logic: manifests whose name already exists in ~/.claude.json are skipped unless --force is passed. This prevents accidentally overwriting manually-crafted entries.

Serve mode (runtime hub — single MCP for all bots)

# Run the hub as a live MCP server (stdio transport)
python hub_server.py --serve

# Serve with multiple scan roots (repeatable)
python hub_server.py --serve \
  --scan-root C:\path\to\projects \
  --scan-root C:\path\to\Claude

# Register+serve in one step (scan --apply then start hub)
python hub_server.py --register \
  --scan-root C:\path\to\projects \
  --scan-root C:\path\to\Claude

The hub scans all mcp.yaml manifests under each --scan-root at startup, then exposes every bot's tools under the <bot>.<tool> namespace (e.g., fleet-health.fleet_status, my-bot.run_scan). Tools are proxied to per-bot subprocess MCP servers with lazy startup.

Hub meta-tool: _hub.list_bots returns the registered bots and their subprocess status.

Hub is pre-registered in ~/.claude.json as mcp-factory-hub (see scripts/register_hub.py).

Node.js template

Factory generates Node.js stubs when runtime.type: node is set in mcp.yaml:

runtime:
  type: node
  command: "node"
  output: "path/to/server.js"

Generated stubs use @modelcontextprotocol/sdk with stdio transport and zod for argument validation. See examples/node_example.yaml for a working demo.

Two Python styles: raw SDK vs. FastMCP

For runtime.type: python, the factory can scaffold either of two styles from the exact same manifest:

runtime:
  type: python
  command: "python"
  style: raw       # default — official `mcp` SDK, hand-rolled list_tools/call_tool
  # style: fastmcp # FastMCP v3 (PrefectHQ/fastmcp), decorator-based tool registration

Both styles read the same tools: / env_required: fields and produce a server that speaks the same stdio JSON-RPC wire protocol — the runtime hub's SubprocessAdapter proxies either one without any adapter changes (see tests/test_fastmcp_template.py::TestFastmcpServeSmoke for a live generate-and-call test).

Gaps, stated honestly: neither style implements per-tool authorization, rate limiting, or the "full failure set" the hub-level subprocess_adapter.py gives you for free (typed errors, lazy start, atexit cleanup) — that's still a per-engagement build on top of either scaffold. The fastmcp template's fail-soft wrapper and env-presence check are new, real code (read mcp_factory/templates/python_fastmcp.j2), not a marketing claim about auth — they were added because FastMCP's decorator model made them cheap to include cleanly; they have not (yet) been backported to the raw template, which is why the two styles differ slightly in what ships out of the box. If your engagement needs FastMCP-specific features beyond this (resources, prompts, HTTP/SSE transport, middleware-based auth), the generated file is a normal FastMCP app — extend it directly.

See examples/fastmcp_example.yaml for a working demo manifest.

mcp.yaml Schema

name: my-bot                   # REQUIRED — unique MCP server name (key in claude.json)
description: >                 # REQUIRED — shown in Claude's tool descriptions
  What this bot does and when to use it.

runtime:                       # REQUIRED
  type: python                 # python | node | binary
  command: "C:\\Python314\\python.exe"  # full path to interpreter
  script: "path/to/server.py"  # existing server (skips scaffold generation)
  output: "path/to/out.py"     # where to write generated scaffold (omit = auto)
  style: raw                   # python only: raw (default) | fastmcp — see "Two Python styles"

tools:                         # REQUIRED — list of MCP tools to expose
  - name: tool_name            # REQUIRED — programmatic identifier
    title: "Tool Name"         # Optional — human display name (see below)
    description: >             # REQUIRED — used by Claude for routing
      What this tool does.
    annotations:               # Optional — the four MCP behaviour hints
      readOnlyHint: true       # tool does not modify anything
      destructiveHint: false   # tool does not perform destructive updates
      idempotentHint: true     # repeat calls have no additional effect
      openWorldHint: false     # tool does not touch an open external world
    args:                      # Optional list of arguments
      - name: arg_name         # REQUIRED
        type: string           # string | number | boolean | object | array
        required: true         # default: true
        description: "..."     # shown in Claude's tool schema

env_required:                  # env var names that must be set at runtime
  - MY_API_KEY

env:                           # static env vars injected into claude.json entry
  MY_API_KEY: ""               # leave value empty — fill in ~/.claude.json manually

tags: [trading, health]        # for documentation / future routing
priority: high                 # high | medium | low

Key rules

• runtime.script + existing file → factory references it, skips scaffold
• runtime.script + missing file → validation error (use runtime.output for new scaffolds)
• runtime.output → explicit path for generated stub (absolute recommended)
• Neither script nor output → error at config-write step

title and annotations

Both are optional and both are worth declaring.

title is the human display name, which lets name stay a programmatic identifier. Display precedence is title → annotations.title → name; the factory exposes only the top-level title so one manifest cannot declare two competing display names.

annotations matters more than it looks. The four hints have pessimistic defaults — destructiveHint and openWorldHint both default to true, readOnlyHint to false — so a tool that declares nothing is treated by a careful client as the most dangerous thing it could be. Declaring readOnlyHint: true on a read-only tool is how you opt out of that. Only the four hint names above are accepted, each must be a real boolean, and an unknown key is a hard error rather than a silent drop (the typo readonlyHint would otherwise leave the tool effectively unannotated on the wire).

Annotations are hints, not a security boundary — the spec is explicit that a client should never make tool-use decisions based on annotations from an untrusted server. They inform a client's UX; they do not enforce anything.

Both fields are emitted by the two Python templates. The Node template does not render them yet — the repo vendors no node_modules, so an altered call shape could not be executed in test, and shipping unverified generated code is worse than shipping the current shape. Declaring them in a node manifest is still valid and forward-compatible.

Generation-time tool-design lint

Manifest validation is a security gate: it fails closed on anything that could inject code into a generated server. It says nothing about whether the resulting tools are any good. A manifest can be perfectly valid and still produce a server that makes an agent measurably worse.

Every run of factory mode therefore also runs a tool-design lint (Step 2), covering 15 rules across three scopes:

It is advisory by default — findings print, generation proceeds, exit 0. --lint-strict makes error-severity findings fatal: it reports them, writes no scaffold, and exits 3.

mcp-factory-hub --manifest examples/fastmcp_example.yaml --lint-strict

Three deliberate properties:

• It reports INDETERMINATE, not a clean bill, when it has no discriminating power — a manifest whose runtime.script exists (the factory references that hand-written file rather than generating one, so the tool list may not match the real server), or a runtime with no template. --lint-strict never blocks on an indeterminate result: the absence of a verdict must not be converted into one.
• There is no tool-name charset rule, even though SEP-986 defines one. The manifest's own identifier validation is already stricter and rejects violations at parse time, so such a rule could never fire on any input reaching the lint — and a check that cannot fail is not a check. Only the length half of SEP-986 is unenforced upstream, so only the length half is a rule.
• Every rule ships a positive control: a test proving it fires on a known-bad tool definition and stays silent on a known-good one, plus a suite-level test asserting a well-formed manifest produces zero findings. The examples in examples/ are themselves held to the lint by tests/test_examples_lint_clean.py.

How to Add a New MCP

1. Write mcp.yaml at your bot repo root (or in examples/)
2. Run the factory:
   bash

   python hub_server.py --manifest path/to/mcp.yaml
   

3. Review ~/.claude.json.factory-test — confirm the entry looks correct
4. Copy the entry into ~/.claude.json under mcpServers
5. Restart Claude Code

If the bot has no existing server, the factory generates a stub at generated/<name>_server.py. Fill in the # TODO: implement sections and set runtime.script to the stub path for future runs.

Runtime Hub Architecture

hub_server.py --serve
  └── mcp_factory/runtime/
      ├── hub.py               async MCP server (lists + routes all tools)
      ├── registry.py          maps <bot>.<tool> → manifest + adapter
      └── subprocess_adapter.py  spawns per-bot MCP server, proxies JSON-RPC

Subprocess lifecycle:

• Adapters start lazily on first tool call (no upfront spawn)
• Keep-alive for the hub session (one process per bot)
• _hub.list_bots() reports status: idle (not yet started) or running
• All adapters stopped via atexit on hub exit; stop() kills if needed after 5 s

Tool naming: <bot-name>.<tool-name> — hyphens preserved, dots as separator. Example: fleet-health.fleet_status, my-bot.get_alerts.

Day 4 — workflow_runner.py

Standalone CLI harness for research workflows, independent of hub_server.py.

# Discover and list all SKILL.md workflows
python -m mcp_factory.workflow_runner --list

# Run a specific workflow
python -m mcp_factory.workflow_runner --run my-skill

# Validate all discovered SKILL.md files
python -m mcp_factory.workflow_runner --validate

# Write/update registry.json from discovered skills
python -m mcp_factory.workflow_runner --write-registry

# Check for drift between discovered skills and registry.json
python -m mcp_factory.workflow_runner --check

# Control cache behavior
python -m mcp_factory.workflow_runner --run my-skill --cache-policy force-refresh
python -m mcp_factory.workflow_runner --run my-skill --cache-policy read-only

How it works

workflow_runner.py scans ~/research by default (override with --scan-root) for SKILL.md files containing YAML frontmatter. Each SKILL.md defines a named workflow with metadata:

---
name: my-skill
description: What this workflow does
output_path_template: "~/vault/output/{date}/{name}.md"
---
Prompt body passed to claude -p subprocess...

• Discover: git ls-files to enumerate tracked SKILL.md files under each scan root
• Validate: checks required frontmatter fields (name, description)
• Cache: SHA-based cache keyed on prompt content; auto (default) skips re-run if output unchanged, force-refresh always re-runs, read-only never writes
• Run: invokes claude -p <prompt> as a subprocess, streams output
• Write output: expands output_path_template, writes result to vault
• Registry: --write-registry persists discovered skills to registry.json; --check detects drift between filesystem and registry without writing

Directory Layout

mcp-factory/
├── hub_server.py              # CLI entry point (factory / scan / serve)
├── mcp_factory/
│   ├── manifest.py            # Manifest dataclass + YAML loader + validation
│   ├── generator.py           # Python MCP server stub scaffolder
│   ├── config.py              # claude.json entry builder + comparator
│   ├── scan.py                # --scan mode: manifest discovery + diff/apply
│   ├── workflow_runner.py     # Day 4: standalone CLI harness for SKILL.md workflows
│   ├── templates/              # packaged as data so `pip install` ships them too
│   │   ├── python_server.py.j2    # Jinja2 template — raw mcp SDK stubs (style: raw, default)
│   │   ├── python_fastmcp.j2      # Jinja2 template — FastMCP v2 stubs (style: fastmcp)
│   │   └── node_server.js.j2      # Jinja2 template for generated Node.js stubs
│   └── runtime/
│       ├── subprocess_adapter.py  # subprocess MCP client (JSON-RPC proxy)
│       ├── registry.py            # tool registry with collision detection
│       └── hub.py                 # async hub MCP server
├── tests/
│   ├── fixtures/
│   │   ├── fleet_health.yaml   # Day 1 self-verification fixture
│   │   ├── minimal.yaml        # Minimal valid manifest
│   │   └── mock_mcp_server.py  # Stdlib-only mock MCP server for adapter tests
│   ├── test_manifest.py
│   ├── test_generator.py
│   ├── test_subprocess_adapter.py
│   ├── test_registry.py
│   ├── test_scan.py
│   ├── test_hub_cli.py
│   ├── test_mcp_pkg.py
│   ├── test_node_template.py
│   ├── test_python_template.py
│   ├── test_fastmcp_template.py  # style: fastmcp generation + import + serve-smoke tests
│   ├── test_register_flag.py
│   ├── test_registration.py
│   ├── test_smoke_hub.py
│   ├── test_watcher.py
│   ├── test_workflow_runner.py  # Day 4: workflow_runner unit + integration tests
│   └── test_integration_fleet_health.py  # live integration tests (skipped if server absent)
├── examples/
│   ├── fleet_health.yaml      # Example manifest referencing an existing server
│   ├── node_example.yaml      # Example manifest for the node template
│   └── fastmcp_example.yaml   # Example manifest for the fastmcp template
└── pyproject.toml

Self-Verification

The examples/fleet_health.yaml manifest references an example server. Running:

python hub_server.py --manifest examples/fleet_health.yaml --verify ~/.claude.json

confirms the factory produces a matching ~/.claude.json entry.

Running Tests

python -m pytest tests/ -v

On a clean checkout (Python 3.12), with pip install -e .[dev]: 348 passed, 8 skipped, 0 failed — the same numbers the public CI run produces and gates on.

The 8 skipped tests skip automatically when the resource or condition they need is absent:

• test_integration_fleet_health.py (5 tests) requires a fleet-health server.py on disk (FLEET_HEALTH_SERVER_PATH).
• test_node_template.py (1 test) requires node and @modelcontextprotocol/sdk (node_modules/) to be present.
• test_examples_lint_clean.py (2 tests) skips examples/fleet_health.yaml, which references an existing hand-written server — the tool-design lint correctly reports INDETERMINATE there rather than judging code it cannot see, and a skip is the honest way to record that.

test_smoke_hub.py (4 tests) no longer needs a live bot fleet to run for real: the hub's demo-manifest fallback (see below) gives it something to discover even against an empty scan root, so these run unconditionally on a clean checkout now.

(On the maintainer's fleet machine, where the fleet-health server and live bots exist, the remaining skipped integration tests run for real and the passed count is higher — but this README claims only what a clean checkout and public CI reproduce.)

The fastmcp-style template tests (test_fastmcp_template.py) are not in this skip list — fastmcp is installed as a [dev] extra, so they run for real on a standard dev setup.

Commercial support

Maintained by Jaimen Bell. For production MCP integrations, custom servers, or agent-reliability work, see jaimenbell.dev.

Building your own MCP server? The MCP Starter Kit has templates, a build playbook, and packaging war-stories from shipping this one.

mcp-name: io.github.jaimenbell/mcp-factory