PolyMCP

Create an MCP HTTP server from plain functions:

from polymcp import expose_tools_http


def add(a: int, b: int) -> int:
    return a + b


app = expose_tools_http(
    tools=[add],
    title="Math Server",
    description="MCP tools over HTTP",
)

Run:

uvicorn server:app --host 0.0.0.0 --port 8000

Quick Start (TypeScript)

Requirements: Node.js 18+

cd polymcp-ts
npm install
npm run build

Skills (skills.sh)

PolyMCP delegates skills management to the skills.sh CLI. PolyAgent and UnifiedPolyAgent automatically inject relevant skills into prompts to improve tool selection and planning.

npx skills --help

# Install in current project (./.agents/skills)
npx skills add vercel-labs/agent-skills

# Install globally (~/.agents/skills)
npx skills add vercel-labs/agent-skills -g

npx skills list
npx skills list -g

Python helper:

from polymcp import run_skills_cli

run_skills_cli(["add", "vercel-labs/agent-skills"])

Agents load skills from:

• ./.agents/skills
• ./.skills
• ~/.agents/skills

If Python agents find no project skills, they print:

[WARN] No project skills found in .agents/skills or .skills.
Use global skills: polymcp skills add vercel-labs/agent-skills -g
Or local skills: polymcp skills add vercel-labs/agent-skills

Skills.sh Agent Example (Python)

python examples/simple_example.py
polymcp skills add vercel-labs/agent-skills
python examples/skills_sh_agent_example.py

Minimal server:

import { tool, exposeToolsHttp } from 'polymcp-ts';
import { z } from 'zod';

const add = tool({
  name: 'add',
  description: 'Add two numbers',
  parameters: z.object({ a: z.number(), b: z.number() }),
  execute: async ({ a, b }) => a + b,
});

const app = await exposeToolsHttp([add], {
  title: 'Math Server',
  description: 'MCP tools over HTTP',
  version: '1.0.0',
});

app.listen(3000);

Inspector

The Inspector is the fastest way to test MCP servers, tools, prompts, and UI resources.

Run standalone Inspector:

cd polymcp-inspector
python -m polymcp_inspector --host 127.0.0.1 --port 6274 --no-browser

Open:

http://127.0.0.1:6274/

Features:

• chat playground with independent multi-tab sessions (per-tab provider, model, servers, and history)
• tool, resource, and prompt explorers
• MCP Apps UI preview with tool-call bridge
• LLM orchestration (Ollama, OpenAI, Anthropic)
• settings page for Inspector/OpenAI/Claude API keys
• auto-tools routing (LLM decides if tools are needed)
• secure mode with API key and rate limits

MCP Apps SDK

Use polymcp_sdk_mcp_apps/ to build UI-first MCP Apps quickly.

cd polymcp_sdk_mcp_apps
npm install
npm run example:quickstart

Connect Inspector to your app server and open the app://... resource in Resources.

Agent Orchestration Examples

Python, multiple MCP servers:

from polymcp.polyagent import UnifiedPolyAgent, OpenAIProvider

agent = UnifiedPolyAgent(
    llm_provider=OpenAIProvider(model="gpt-4o-mini"),
    mcp_servers=[
        "http://localhost:8000/mcp",
        "http://localhost:8001/mcp",
    ],
    verbose=True,
)

answer = agent.run("Read sales data, compute totals, then summarize.")
print(answer)

TypeScript, HTTP + stdio:

import { UnifiedPolyAgent, OpenAIProvider } from 'polymcp-ts';

const agent = new UnifiedPolyAgent({
  llmProvider: new OpenAIProvider({
    apiKey: process.env.OPENAI_API_KEY!,
    model: 'gpt-4o-mini',
  }),
  mcpServers: ['http://localhost:3000/mcp'],
  stdioServers: [{ command: 'npx', args: ['@playwright/mcp@latest'] }],
  verbose: true,
});

await agent.start();
const answer = await agent.run('Collect data and summarize.');
console.log(answer);

🦞 PolyClaw (Autonomous OpenClaw-Style Agent)

PolyClaw is an autonomous agent inspired by OpenClaw, designed for end-to-end execution in PolyMCP workflows.

What it does:

• Understands a goal and executes the workflow autonomously
• Runs real shell actions inside Docker with the project mounted at /workspace
• Can create, configure, register, and test MCP servers via polymcp CLI when useful
• Adapts strategy from command output and continues until completion
• Produces a final report with completed actions, failures, and next concrete step

Safety defaults:

• Delete/remove commands require confirmation by default (--confirm-delete)
• If destructive commands are denied, PolyClaw reports that no removal was executed
• Recommended usage: isolated Docker environments for autonomous runs

Python examples:

• examples/polyclaw_example.py
• examples/polyclaw_mcp_workflow_example.py

CLI

polymcp init my-project --type http-server
polymcp server add http://localhost:8000/mcp
polymcp agent run
polymcp agent run --type polyclaw --query "Create and validate a local MCP workflow"
polymcp inspector

See polymcp/cli/README.md for full commands.

Development

Python:

pip install -e .[dev]
python -m pytest

TypeScript:

cd polymcp-ts
npm run build
npm test
npm run lint

License

MIT