phospho-app/phospho-mcp-server

phospho-mcp-server

phospho's official MCP Server

This repository implements a Model Context Protocol (MCP) server for phosphobot, enabling natural language interaction and control over a physical robot. It exposes tools to execute actions (e.g. pick up an object) and stream images from cameras.

Built using mcp and tailored for Claude.

https://github.com/user-attachments/assets/2d91f716-c581-4c36-94ab-cc5a94001468

Features

• Camera stream: retrieves the current webcam frame
• Replay tool: triggers a robot action from a dataset (e.g. pick up banana)
• phosphobot wrapper: manages local API processes and communication

Getting Started

1. Install and run phosphobot

phosphobot is an open source software that lets you control robots, record data, train and use VLA (vision language action models).

Run this to install phosphobot:

#macOS
curl -fsSL https://raw.githubusercontent.com/phospho-app/phosphobot/main/install.sh | bash

#Linux
curl -fsSL https://raw.githubusercontent.com/phospho-app/phosphobot/main/install.sh | sudo bash

#Windows
powershell -ExecutionPolicy ByPass -Command "irm https://raw.githubusercontent.com/phospho-app/phosphobot/main/install.ps1 | iex"

Run the phosphobot server:

phosphobot run

2. Install the MCP server in Claude Desktop

Make sure Claude desktop is installed.

Install phospho-mcp-server this way using uv:

curl -LsSf https://astral.sh/uv/install.sh | sh # Install uv on MacOs or Linux
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex" # On Windows

# Clone repo
git clone https://github.com/phospho-app/phospho-mcp-server.git

# Install the server
cd phospho-mcp-server/phospho-mcp-server
uv run mcp install server.py

This will:

• Boot the MCP server under the name phospho
• Register all tools with

Now, open Claude desktop. You should see the phospho MCP server listed along the tools.

How it works

The server speaks to a local instance of phosphobot through its REST API (default http://localhost:80).

• /frames → image feed
• /recording/play → trigger replay from dataset

All calls are wrapped via tools/phosphobot.py. If you run phosphobot on a different port, you need to modify the base URL.

Development and testing

To test your server with the MCP inspector, run:

uv run mcp dev server.py

Camera from Claude

Ask:

“What’s on my desk?”

Claude will call:

{
  "tool": "get_camera_frame"
}

Replay from Claude

Ask:

“Pick up the banana”

Claude will call:

{
  "tool": "pickup_object",
  "args": { "name": "banana" }
}

Tools

pickup_object

Launches a replay episode to simulate object manipulation.

@mcp.tool()
def pickup_object(name: Literal["banana", "black circle", "green cross"]) -> str

• Launches a pre-recorded episode based on the object name.

get_camera_frame

Captures a JPEG image from phosphobot's camera.

@mcp.tool()
def get_camera_frame() -> Image

• Uses the local phosphobot API (/frames)
• Returns a base64-encoded JPEG image

Architecture

phospho-mcp-server/
│
├── server.py              # MCP server (FastMCP) with tools
├── tools/
│   ├── phosphobot.py      # Wrapper for phosphobot process and API
│   └── replay_api.py      # Tool to launch a replay
├── pyproject.toml         # Project configuration
└── README.md              # This file

The PhosphoClient class is used to manage the phosphobot process lifecycle, and send GET/POST requests to its local API.

References

• Model Context Protocol
• phosphobot