Toolbartender MCP

🚀 ToolBartender 🍸

ToolBartender is a Planner MCP that takes a user's natural - language request as a single goal and generates an execution plan (plan) for which MCP tools to call and in what order. It doesn't actually execute the tools but returns a JSON containing steps / assumptions / required_confirmations / execution_hint to enable the LLM/execution agent to execute safely and predictably.

One - line summary: ToolBartender = MCP that creates “tool combination plans” rather than “tool executions”

🚀 Quick Start

1) Local run

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python src/main.py
# Open http://localhost:3333/mcp in your browser

2) Local testing with Inspector

npx @modelcontextprotocol/inspector
# http://localhost:6274

Connect the server URL to http://localhost:3333/mcp in the Inspector UI.

3) PlayMCP integration test

• A public URL is required (ngrok/deployment).
• Go to PlayMCP → “Register a new MCP server” → Enter https://<public - host>/mcp → Temporary registration → Apply to AI chat. The flow is already in the existing README.

✨ Features

• Goal - based Planning: Receives complex requests as a single goal and generates a plan, including requests for travel planning, schedule inquiry/adjustment, information exploration and summarization, and result delivery.
• Safe Execution: Returns a JSON with steps, assumptions, required_confirmations, and execution_hint to ensure safe and predictable execution by the LLM/execution agent.

📦 Installation

The installation steps mainly involve setting up a virtual environment and installing dependencies. Refer to the "Quick Start" section for detailed commands.

📚 Documentation

• docs/00_quickstart.md: 5 - minute setup.
• docs/01_concept.md: Planner MCP concept, philosophy, and roadmap.
• docs/02_tools.md: Input/output/examples for each tool + naming rules.
• docs/03_playmcp.md: Tips for PlayMCP registration and review checklist.
• docs/04_deploy.md: Deployment (including nginx/SSE) + domain/Cloudflare points.
• docs/05_troubleshooting.md: Collection of issues related to 405/handshake/SSE.

🔧 Technical Details

PoC Endpoint

http://mcp.toolbartender.dev/mcp (Streamable HTTP / SSE)

Keywords

planner, safe execution, tool orchestration, PlayMCP

What it does

It takes complex requests as a single goal and generates a plan, such as:

• Travel planning (e.g., “Travel from Pangyo to Gangnam at 8 PM today”).
• Schedule inquiry/adjustment.
• Information exploration and summarization.
• Result delivery (email/messenger, etc.).

Output Plan includes

• steps: MCP tools to use + execution order.
• assumptions: Pre - conditions.
• required_confirmations: Tasks requiring user confirmation (especially write - related).
• execution_hint: Guide for the execution agent.

Exposed MCP tools (this server)

Due to PlayMCP registration restrictions (regular expressions), it is recommended that the tool names exposed by ToolBartender be fixed in ASCII.

• plan_create: Generate a plan from a goal + available tools.
• plan_validate: Verify if a plan is executable in the current context (available tools).
• plan_render_prompt: Generate a prompt for the LLM execution agent to call steps “in order”.
• plan_explain: Explain the plan to the user easily.

Note that the internal plan.steps[*].tool_name refers to the tool names of other MCPs (e.g., calendar.read, map.route), so it is natural to keep them as they are. For tool I/O, schema, and examples, refer to docs/02_tools.md.

📄 License

This project is licensed under the MIT license.