How to Extend Reachy Mini Capabilities With Remote MCP Tools

Hugging Face recently enabled the Reachy Mini open-source robot to call remote APIs directly from Hugging Face Spaces. The official June 3 update utilizes the Model Context Protocol (MCP) to connect the robot’s local inference engine with cloud-based capabilities. This allows you to add dynamic features like web search or weather checking without flashing new firmware or altering the local application code. You can treat the Model Hub as a remote library for your physical hardware.

This architecture represents a fundamental shift in embodied AI. Instead of hard-coding every behavior into the robot’s onboard computer, you can route non-physical tasks to remote environments. The setup minimizes the processing burden on the $300 open-source humanoid robot while granting it access to complex external systems.

System Architecture

The Reachy Mini operates on a split-execution model. The core Reachy Mini conversation app runs a fully local stack for Speech-to-Text (STT), Text-to-Speech (TTS), and base LLM inference. This local pipeline guarantees that the 11-inch, 3.3-pound chassis processes spatial awareness and user commands with minimal latency.

When the local LLM needs external data, it relies on the MCP standard to communicate with remote servers. Hugging Face implements this by turning Gradio Spaces into MCP servers. A Gradio Space exposes its functions, complete with parameter descriptions and expected arguments, as discrete tools. The robot’s local LLM discovers these tools at runtime, formulates the correct JSON payload, and executes the call across the network.

Built-in vs. Remote Tools

Routing tasks correctly requires understanding the distinction between local physical controls and remote digital extensions.

Built-in tools manage the physical chassis. These functions execute entirely within the local conversation stack to ensure safety and low latency. Remote tools handle data retrieval or heavy compute tasks that would otherwise overwhelm the onboard hardware.

Tool Category	Execution Environment	Example Functions	Primary Use Case
Built-in	Local hardware	move_head, dance, play_emotion, head_tracking, camera	Physical motor control and sensor reading
Remote	Hugging Face Spaces (Cloud)	Internet search, weather APIs, computation	External data access and complex logic

The local LLM acts as a router. When prompted to look at a user, it triggers the built-in head_tracking tool. When asked about the weather, it queries the remote Hugging Face Space.

Connecting a Remote MCP Tool

Adding a remote capability requires a single CLI command on the machine running the Reachy Mini conversation app. You register the remote tool by referencing the Hugging Face Space repository.

Execute the command within the active environment running your robot’s software stack:

reachy-mini-conversation-app tool-spaces add pollen-robotics/reachy-mini-weather-tool

The application reaches out to the specified Gradio Space, authenticates the connection, and downloads the tool schema. The local LLM immediately incorporates the new schema into its context window. You do not need to restart the core conversation service or recompile any local dependencies. The integration functions exactly like equipping agent skills in software-only environments.

Integrating with Development Environments

The modular nature of MCP allows the Reachy Mini to interact with complex local software systems alongside remote cloud tools. A community-led project recently released a dedicated MCP server that integrates the robot with Claude Code.

By running this specific MCP server locally, you expose your terminal events to the Reachy Mini’s LLM. The robot monitors the build processes or test suites running on your machine. When a test suite passes or a build fails, the LLM intercepts the event and triggers built-in physical responses. The robot can perform celebratory dances for successful deployments or display specific emotional states for compilation errors, effectively serving as an embodied coding companion.

Limitations and Tradeoffs

Delegating capabilities to remote Spaces introduces network dependency. While the core motor functions remain responsive, invoking an MCP tool over the internet incurs latency. Operations requiring immediate physical feedback must remain confined to the built-in local toolset.

Network timeouts will interrupt the execution of remote tools. The Reachy Mini conversation app handles these timeouts gracefully, typically returning a verbal error via the local TTS engine, but you must account for variable execution times when designing complex behaviors. Remote tools are also subject to the compute availability of the hosting Hugging Face Space. Heavily trafficked public Spaces may queue requests, delaying the robot’s verbal or physical response.

Review the Hugging Face MCP documentation to verify the schema requirements before building custom Gradio Spaces for your Reachy Mini. Start by provisioning a basic remote tool like a static data lookup to baseline the network latency in your specific environment.