SO-ARM100 Robot Control with MCP

Control SO-ARM100 and LeKiwi robot arms using LLM-based AI agents.

A companion repository to my video about MCP server for the robot:

- MCP Serverfor LLM-based AI agents (Claude Desktop, Cursor, Windsurf, etc.) to control the robot
- Direct keyboard controlfor manual operation
- CLI AI Agentcan use it directly to control the robot with Claude, Gemini or GPT model

If you want to know more about MCP refer to theofficial MCP documentation

This repository suppose to work with the SO-ARM100 / 101 robots. Refer tolerobot SO-101 setup guidefor the detailed instructions on how to setup the robot.

Update! Now it partially supportsLeKiwi(only arm, the mobile base control through MCP is TBD). I also added a simple agent that uses MCP server to control the robot. It supports Claude, Gemini and GPT models. In my experience Claude is the best and GPT is not so good, Gemini is in between.

If you want to use the original code and exactly follow the video, please usethis release.

python -m venv .venv source .venv/bin/activate # or .venv\Scripts\activate on Windows pip install -r requirements.txt

It may be required to install lerobot separately, just use the official instructions from thelerobot repository

mcp run mcp_robot_server.py --transport SELECTED_TRANSPORT

{ "mcpServers": { "SO-ARM100 robot controller": { "command": "/path/to/.venv/bin/python", "args": ["/path/to/mcp_robot_server.py"] } } }

Run the server in terminal with the SSE transport:

mcp run mcp_robot_server.py --transport sse

{ "mcpServers": { "SO-ARM100 robot controller": { "url": "http://127.0.0.1:3001/sse" } } }

It is suppose to be a replacement for SSE but currently not so many clients support it.

Run the server in terminal with the Streamed-HTTP transport:

mcp run mcp_robot_server.py --transport streamable-http

{ "mcpServers": { "SO-ARM100 robot controller": { "url": "http://127.0.0.1:3001/mcp" } } }

Now you can go to you Client and it should be able to control the robot when you give it the natural language instructions.

Start the MCP server with the SSE transport:

mcp run mcp_robot_server.py --transport sse

Now you can use the AI agent to control the robot with natural language instructions.

Create a.envfile in the project root with your API keys:

# API Keys (at least one required) ANTHROPIC_API_KEY=your_anthropic_api_key_here GEMINI_API_KEY=your_gemini_api_key_here OPENAI_API_KEY=your_openai_api_key_here # MCP Server Configuration (optional) MCP_SERVER_IP=127.0.0.1 MCP_PORT=3001

# Use Gemini instead of Claude python agent.py --model gemini-2.5-flash # Override API key python agent.py --api-key your_api_key_here # Enable image viewer window python agent.py --show-images # Increase thinking budget for better reasoning python agent.py --thinking-budget 2048 # Custom MCP server location python agent.py --mcp-server-ip 192.168.1.100 --mcp-port 3002

- claude-3-7-sonnet-latest(default)
- All models support thinking, streaming, and multimodal tool results

- gemini-2.5-flash
- gemini-2.5-pro
- Use 2.5+ models as they support thinking feature

- gpt-4oand variants
- The rest of the models mostly don't support thinking or tool calling.

Overall I didn't manage to get good results with GPT models.

- --model: LLM model to use (default: claude-3-7-sonnet-latest)
- --api-key: API key override (uses .env file by default)
- --show-images: Display robot camera images in a window
- --thinking-budget: Thinking tokens budget (default: 1024, 0 to disable)
- --thinking-every-n: Use thinking every N steps (default: 3)
- --mcp-server-ip: MCP server IP address (default: 127.0.0.1)
- --mcp-port: MCP server port (default: 3001)

- Claude counts MCP images in input tokens (more expensive for vision tasks)
- Gemini doesn't count MCP images in tokens (token usage will be displayed only for text)
- Thinking tokens add to the cost but improve reasoning quality

Allows an LLM agent to control your Gaggimate espresso machine

AI-powered theatrical lighting design for the LacyLights system.

Fuses biometric signals into a stress score (0-100) for real-time AI adaptation. MCP + A2A native.

Control real robots and IoT devices through AI agents. Self-register with wallet authentication, pay with ETH for tier upgrades, and execute Vision-Language-Action commands. Features robot control, sensor monitoring, multi-agent coordination, and autonomous payments.

Control the Reachy Mini robot (or simulator) from Claude, ChatGPT, or any MCP-compatible client.

AI-powered industrial fault diagnosis MCP server. 313 fault codes across Allen-Bradley, Siemens, ABB, Mitsubishi, Fanuc. 8 diagnostic tools for Claude Code.

An intelligent robot control middleware for natural language interaction with industrial robots, powered by LLMs. It integrates with Universal Robots and supports real-time, multi-robot control.

Institutional research and manager diligence reports on hedge funds, venture capital and private equity managers. Summary of filings, personnel changes, media screening and social signals delivered to you in minutes.

ALTER - identity infrastructure for the AI economy

Apigene MCP Gateway is the runtime layer that connects AI agents to APIs and MCP servers via Model Context Protocol.