Chiplab

Simulate & validate firmware on virtual instances of real chips, straight from your AI coding agent — no hardware.

Chiplab is a hosted MCP service. Your AI coding agent connects to it, uploads compiled firmware, runs it on a virtual instance of a real microcontroller, and reads back the UART output. You don't drive the toolchain — your agent does.

We're building the next way to develop chips, in the open and early. Today it runs Hello world! on STM32 and Nordic boards; a lot more is on the way — see veecle.ai. Try it, poke at it, and tell us what you'd like to see.

Quickstart

1. Create a free account at veecle.ai (no credit card).

2. Connect your agent to the Chiplab MCP server.

3. git clone https://github.com/veecle/chiplab && cd chiplab, then tell your agent:

   "Set up Chiplab and run the stm32f4-discovery example on it."

Your agent reads AGENTS.md, installs what it needs, builds the firmware, uploads it, runs it on the virtual board, and reports Hello world! back from the chip's UART. That's the whole loop.

Connect your agent

Chiplab is an HTTP MCP server at https://chiplab.veecle.ai/mcp. Most clients take the standard mcpServers schema:

{
  "mcpServers": {
    "chiplab": {
      "type": "http",
      "url": "https://chiplab.veecle.ai/mcp"
    }
  }
}

On first use your client opens a browser to sign in. To verify, ask your agent to call Chiplab's discovery/help tool with no arguments.

<details> <summary><b>Claude Code</b></summary>

If you cloned this repo, the server is already configured via .mcp.json — Claude Code will prompt you to trust it. Otherwise:

claude mcp add --transport http chiplab https://chiplab.veecle.ai/mcp

</details>

<details> <summary><b>Claude Desktop / Cursor</b></summary>

Use the mcpServers block above in claude_desktop_config.json or .cursor/mcp.json. </details>

<details> <summary><b>VS Code</b></summary>

Same block in .vscode/mcp.json, but under a servers key instead of mcpServers. </details>

<details> <summary><b>Codex</b></summary>

TOML in ~/.codex/config.toml:

[mcp_servers.chiplab]
url = "https://chiplab.veecle.ai/mcp"

Then sign in with codex mcp login chiplab. </details>

How it works

The contract is framework-agnostic: your agent builds an ELF → uploads it → runs it on the target board → reads the captured UART output. Runs return synchronously and are bounded to a fixed amount of virtual time.

This repo ships a ready-to-run example for every supported board, grouped by framework — bare-metal (Rust, vendor HAL), embassy-rust (Embassy async), zephyr-os (Zephyr RTOS, C), freertos (FreeRTOS, C), and threadx (Eclipse ThreadX, C). The full board × framework matrix is in supported-boards.md.

Toolchain and build details live in each framework's directory (examples/<framework>/README.md for humans, AGENTS.md for agents) — your agent finds them on its own. Prefer building by hand? Each framework README has the exact commands.

To run your own firmware, build an ELF for a supported board and ask your agent to upload and run it the same way.

Contributing

Adding a board example is the easiest way in: mirror an existing example for the same framework and add a row to supported-boards.md. All conventions live in AGENTS.md and each framework's AGENTS.md. New chip, OS, or peripheral support is server-side — open a request instead of adding an example for an unsupported board.

Found a bug? Missing a chip? Tell us here.

This is Chiplab's public home — bugs, feature ideas, and chip/OS/peripheral requests all belong in this repo's issues. Requests really do shape the roadmap — see veecle.ai/roadmap for what's planned, or come say hi on Discord.

License

MIT © 2026 Veecle GmbH.