GitHub - manojbajaj95/mcp-skill: Convert an mcp server to an skill that agents can use.

6 min read Original article ↗

Turn any MCP server into a typed Python SDK.

Compile MCP tools into code.

mcp-skill introspects an MCP server and generates a Python class where each tool becomes a typed async method.

Before and After

Before

Agents call MCP tools through the model loop — one round-trip per tool call:

llm.call_tool("web_search_preview", {"query": "..."})
# → model decides next step → calls another tool → model decides again → ...

After

Agents call tools directly in code:

result = await app.web_search_preview(
    objective="find latest news",
    search_queries=["topic X 2026"]
)
# Agent processes result in code — no round-trip back to model

How It Works

┌─────────────┐      ┌───────────────┐      ┌────────────────────┐
│  MCP Server  │─────▶│  mcp-skill    │─────▶│  Generated Skill   │
│  (any URL)   │      │  CLI          │      │                    │
│              │      │               │      │  app.py            │
│  Tools:      │      │  1. Connect   │      │  ├─ Typed class    │
│  - search    │      │  2. Introspec │      │  ├─ Async methods  │
│  - fetch     │      │  3. Map types │      │  ├─ Auth + storage │
│  - ...       │      │  4. Generate  │      │  └─ JSON parsing   │
│              │      │  5. Validate  │      │                    │
└─────────────┘      └───────────────┘      │  SKILL.md          │
                                             │  └─ Agent docs     │
                                             └────────────────────┘
  1. Connects to the MCP server using fastmcp
  2. Introspects all available tools via list_tools()
  3. Converts each tool's JSON Schema into Python type annotations
  4. Generates a typed App class where each MCP tool becomes an async method
  5. Validates the output with ast.parse → ruff → ty
  6. Generates SKILL.md with tool documentation and usage examples for agents

Motivation

MCP servers give agents access to tools, but every tool call round-trips through the model — request tool, execute, full result back into context, decide next step. For large payloads or sequential calls, this burns tokens and adds latency.

Programmatic Tool Calling fixes this: the agent writes code that calls tools directly, without model round-trips per invocation. Fetch, filter, aggregate — all in one code block.

mcp-skill makes this possible by compiling any MCP server into a plain Python class. Each tool becomes a typed async method. The agent just writes Python.

For a deeper explanation of why MCP vs. CLI is the wrong framing, see MCP vs CLI is the Wrong Question.

from parallel_search.app import ParallelApp

app = ParallelApp(auth="sk-...")
result = await app.web_search_preview(
    objective="find latest news on topic X",
    search_queries=["topic X 2026"]
)

Setup

# Install with uv
uv pip install -e .

# Or use directly without installing globally
uvx --from . mcp-skill create --url https://your-mcp-server.com/mcp --auth api-key

Requires uv and Python 3.10+.

Release Workflow

Releases are managed by release-please, which only considers commits on main that follow the Conventional Commits format.

Use commit subjects like:

feat: add CLI command to inspect generated apps
fix: handle auth recovery when cached credentials are stale
chore: update release workflow docs

Version bump behavior:

  • feat: creates a minor release
  • fix: creates a patch release
  • feat!: or a BREAKING CHANGE: footer creates a major release

Best practices:

  • Prefer squash merges so the PR title becomes the commit on main
  • Write PR titles in Conventional Commits format
  • Keep the subject line short and imperative
  • Add a short body when the change needs extra context

Optional local enforcement with pre-commit:

uv tool install pre-commit
pre-commit install
pre-commit install --hook-type commit-msg

This repo includes a commit-msg hook in .pre-commit-config.yaml that rejects non-conventional commit subjects before they are created.

If release-please says a commit "could not be parsed", push a new conventional commit to main and rerun the workflow.

Usage

# Interactive mode — prompts for URL, auth type, etc.
uvx --from . mcp-skill create

# Non-interactive mode
uvx --from . mcp-skill create \
  --url https://search-mcp.parallel.ai/mcp \
  --auth api-key \
  --api-key YOUR_KEY \
  --name parallel-search \
  --non-interactive

Discover Existing Apps

Use the CLI as a local catalog for both version-controlled skills/ apps and generated .agents/skills/ apps.

# List every discovered app
uvx --from . mcp-skill list-apps

# List callable functions for a specific app
uvx --from . mcp-skill list-functions notion

# Inspect a specific function's signature and docstring
uvx --from . mcp-skill inspect notion notion_search

Use an App in Python

Once you find the app and method you need, call it from async Python:

import asyncio
from sentry.app import SentryApp


async def main():
    sentry = SentryApp()
    user = await sentry.whoami()
    print(user)


asyncio.run(main())

All generated tool wrappers are async. Use them carefully with await inside an async function. If you skip await, you will get a coroutine object instead of the actual result.

Generated Output

The skill lands in .agents/skills/<name>/ as a Python package:

.agents/skills/parallel_search/
├── __init__.py
├── app.py          # Typed Python class wrapping the MCP server
└── SKILL.md        # Agent-facing docs, dependencies, and usage

Here's what the generated app.py looks like:

class ParallelApp:

    def __init__(self, url: str = "https://...", auth=None) -> None:
        ...

    async def web_search_preview(
        self,
        objective: str,
        search_queries: list[str],
    ) -> dict[str, Any]:
        """Search the web with multiple queries in parallel."""
        ...

    async def fetch_url(
        self,
        url: str,
        max_length: int = None,
    ) -> dict[str, Any]:
        """Fetch and extract content from a URL."""
        ...

    def list_tools(self):
        return [self.web_search_preview, self.fetch_url]

Each method connects to the MCP server, calls the underlying tool, and returns parsed JSON. Auth credentials are persisted to ~/.mcp-skill/auth/ after first use — keyed by server URL, so credentials persist across restarts automatically.

Who Is This For?

Developers building:

  • MCP-based agents that need direct tool access without model round-trips
  • Automation systems using MCP tools as programmatic building blocks
  • Code-execution agents using Programmatic Tool Calling

Current Limitations

  • Auth: Supports API key (Bearer or custom header), OAuth (including PKCE with either dynamic client registration or pre-registered client_id), and none — no mTLS
  • Runtime dependency: Generated code depends on fastmcp for MCP client connections
  • Connection per call: Each method creates a new MCP client connection (no pooling)
  • Tools only: MCP resources and prompts not yet supported

Task List

Tracked improvements based on real-world usage:

  • Fix output directory path — Changed from .agents/skill/<name> to .agents/skills/<name>
  • Add dependency info to SKILL.md — Dependencies listed with uv and pip install commands
  • Generate __init__.py — Skill directory is a proper Python package
  • Post-generation validation — ast.parse → ruff check → ty check with uvx fallback
  • Package-style imports — Moved app.py to skill root; import via from <skill>.app import <Class>
  • Persistent token storage — FileTree-backed credential storage at ~/.mcp-skill/auth/, keyed by server URL
  • Unified auth signature — All auth types use auth=None in __init__
  • Sanitize skill names — Hyphens/dots converted to underscores for valid Python identifiers
  • Async CLI — CLI commands are fully async via asyncclick
  • Local app discovery — mcp-skill list-apps, mcp-skill list-functions <app>, and mcp-skill inspect <app> <function>
  • Support MCP resources and prompts — Currently only tools are introspected and generated