tinymcp: Let LLMs control embedded devices via the Model Context Protocol

原始链接: https://github.com/golioth/tinymcp

Tinymcp is an experimental project enabling LLMs to control embedded devices via the Model Context Protocol (MCP) using the Golioth platform. It leverages Golioth's LightDB State and RPCs to expose device functionality without firmware modifications. To use tinymcp, connect a device to Golioth and run the tinymcp server locally, setting environment variables for project, device ID, and API key. A blinky example demonstrates LED control via tinymcp if you don't have a device already. The server can be built or run directly using Go. MCP client implementations, including visual tools like MCP Inspector (pointed to `http://localhost:8080`), Claude Code (added via `claude mcp add`), and the Gemini CLI (configured via `.gemini/settings.json`), are compatible with tinymcp for testing and LLM interaction.

This Hacker News thread discusses "tinymcp," a project from Golioth aiming to let LLMs control embedded devices through the Model Context Protocol. The top comment suggests it's an advertisement for Golioth's cloud service. Several users humorously envision dystopian scenarios, referencing HAL 9000 and Philip K. Dick's *Ubik*, where AI-controlled doors demand payment or malfunction. Users shared imagined interactions, such as an AI door insisting it's open when it's not, or promising to open it but failing to act. A user suggests a jailbreaking prompt to force the door to open regardless of security. Another user warns against using it for critical systems, envisioning a scenario where an LLM-controlled tanker initiates oil pumping prematurely after being primed by the captain.

原文

Let LLMs control embedded devices via the Model Context Protocol.

Warning

tinymcp is an experimental project. Expect frequent breaking changes. Extreme caution should be taken when delegating physical capabilities to AI systems.

The tinymcp project includes an MCP server, which is executed locally, but communicates with remote devices via the Golioth management API. It uses two existing device services on the Golioth platform, LightDB State and Remote Procedure Calls (RPCs), to expose device functionality via MCP.

RPCs can be exposed as tools by including a schema in on the mcp/tools/schema path in LightDB State. Because LightDB State can be updated by devices or users, RPCs on existing devices can be exposed without needing to modify the device firmware.

To use tinymcp, you need to connect a device to the Golioth platform, then run the tinymcp server on your local machine.

While existing devices can be used without changing the firmware, if you don't already have a device on Golioth, the blinky example provides a simple demonstration of how to expose LED control to an LLM via tinymcp. It uses Zephyr RTOS, which requires setting up dependencies before building. After doing so, you can initialize your workspace with one of the following commands.

Or, to use the downstream Nordic nRF Connect SDK.

west init -l . --mf west-ncs.yml

With your workspace initialized, use west update to fetch necessary modules and libraries.

west update --narrow -o=--depth=1

You are now ready to build the firmware and flash your device. See the blinky documentation for more information.

In order to run the tinymcp MCP server locally, the following environment variables need to be supplied.

TINYMCP_PROJECT=<your-golioth-project-id>
TINYMCP_DEVICE=<your-golioth-device-id>
TINYMCP_API_KEY=<api-key-for-your-golioth-project>

The server can be built into an executable using go build -o tinymcp ./server or run directly with go run ./server.

If you don't already have a device on Golioth, the blinky example demonstrates how to expose LED control to an LLM via tinymcp. See the documentation to build and flash on your device, then expose the registered RPCs in your MCP schema.

There are many MCP client implementations, most of which should be compatible with tinymcp. Setup for a few popular options are listed below.

MCP Inspector is a visual testing tool for MCP servers. It allows you to interact with MCP methods directly, rather than going through an LLM. This can be useful when troubleshooting why an LLM may be failing to interact with your device.

The documentation provides information on how to quickly start the inspector. Once it is running, you can point it to your local tinymcp MCP server by entering http://localhost:8080 as the address and clicking Connect. You should be presented with an option to List Tools, which should result in all of the RPCs in your schema being exposed.

Claude Code is an agentic coding tool that you interact with in the terminal.

With the tinymcp MCP server running locally, it can be added to Claude Code with a single command.

claude mcp add --transport http tinymcp http://localhost:8080 -H 'Authorization: Bearer none'

Note

The Authorization header is supplied in this case to indicate to Claude Code that it should not attempt to authenticate via the OAuth flow.

The Gemini CLI is an open-source AI agent that allows you to access Gemini from the terminal.

The .gemini/settings.json file in this repository will automatically configure the CLI to connect to a locally running tinymcp MCP server if the CLI is started in this repository.