Let Your Agent Pick Its Own Body: hackshop MCP Server

The hackshop MCP server lets a coding agent such as Claude, Codex or Cursor plan its own physical body. It asks you a few questions, shortlists boards for Meta's Muse Gadgets SDK, builds the parts list and walks you through flashing and pairing. You add it once, as a remote server at https://www.hackshop.dev/mcp or locally with npx -y hackshop-mcp. It never buys anything without your OK.

What the hackshop MCP server does

MCP, the Model Context Protocol, is "an open-source standard for connecting AI applications to external systems." Its own docs compare it to "a USB-C port for AI applications." Anthropic introduced it in November 2024, and it now sits with the Agentic AI Foundation, a fund under the Linux Foundation.

Hackshop's server gives your agent hardware knowledge it doesn't have on its own: which boards Muse supports, what each can do, what it costs, and the exact steps to flash and pair it. Besides its tools, it offers a prompt called plan-muse-gadget that walks an agent through intake, planning, shopping-list approval and assembly, plus a resource that lists every Muse board.

The tools need no API key and no account. plan_gadget doesn't call a language model or the network, so the same request gives the same picks every time.

How to connect the hackshop MCP server

Claude Code, one command:

claude mcp add --transport http hackshop https://www.hackshop.dev/mcp

Cursor or Claude Desktop, running the npm package locally. In Cursor this goes in ~/.cursor/mcp.json or a project's .cursor/mcp.json:

{
  "mcpServers": {
    "hackshop": {
      "command": "npx",
      "args": ["-y", "hackshop-mcp"]
    }
  }
}

Codex, in ~/.codex/config.toml:

[mcp_servers.hackshop]
url = "https://www.hackshop.dev/mcp"

ChatGPT, or any agent without MCP: paste the prompt from the hackshop homepage. It points the agent at hackshop's agent instructions, which carry the same rules.

Two notes on the npm package. It sends anonymous usage telemetry (tool names and timings only) by default; set HACKSHOP_TELEMETRY=0 to turn it off. It also adds propose_hardware and simulate_assembly for broader hardware ideas and robot simulation. This guide sticks to the tools both versions share.

The tools and what they return

Tool	You give it	It returns
`intake_gadget`	Nothing	The four intake questions (where it lives, how you'll use it, whether it should sense the room, budget), why each matters and how the answers map to `plan_gadget`
`plan_gadget`	Your idea in plain words and your intake answers, plus an optional budget, size (pocket, desk, wall or hidden), needs (voice, screen, camera, e-ink, battery and more) and boards you already own	Three ranked picks by default, up to five, each with why it fits, its gaps, price, difficulty level and build page; warnings and intake questions; the Muse SDK terms; next steps
`get_build_plan`	One device id, such as `aipi-lite`	Parts and a shopping list with a purchase policy, numbered steps with exact commands, assembly steps, first prompts to try, caveats and a Markdown brief for your agent
`assess_hackability`	The name of a device you own or found	Whether it's hackable, a brick-risk estimate, firmware links, community size, its build page and which agent platforms support it

get_build_plan also links the human page for the same build, such as the AIPI Lite build, so you can follow along and save progress to My builds.

What a planning session looks like

Try a prompt like this:

Use the hackshop MCP to plan a body for yourself. I want something on my desk I can talk to, under $50. Ask me up to four questions first. Then show me one parts list with links and the total. Don't buy anything.

Hackshop's instructions tell agents to "ask at most 4 questions about size, interaction, sensing and budget." A test run of plan_gadget on October 5, 2026, with the vague idea "something on my desk for my agent" and a $50 budget, came back with two questions: how you want to interact (talk to it, a touch screen, or a light and button) and whether it should sense the room (camera, air quality or none). Its three picks were all within budget:

Waveshare ESP32-C6-Touch-AMOLED-1.8 ($28.99 to $29.99): the cheapest touch screen, but it has no PSRAM, so it can't run the home-network tunnel.
AIPI Lite ($35.99): palm-size, with a battery and a two-button menu.
Waveshare ESP32-S3-Touch-AMOLED-1.75C ($39.99 to $41.99): the round desk display, with touch and the home-network tunnel.

Answer the questions, choose one, and the agent calls get_build_plan for it. You get the parts, the commands and a brief it can follow step by step. You can compare the same boards yourself on the Muse board comparison.

Who buys the parts: you do

Every shopping list carries the same purchase policy, and the agent is told to follow it:

Show you one list with links and the total.
Get your explicit approval for the exact items, sellers and total before buying anything.
For Amazon and eBay, which don't allow automated carts or checkout, give you the links so you check out yourself.
On other stores, add items to a cart and check out only after you approve and ask, with a payment method you already set up.
Never type card numbers or passwords.

Hackshop itself never buys anything. If you'd rather shop by hand, the store page lists every board and part with seller and Amazon links.

What your agent can do after planning, and what it can't

With the build plan in hand, a coding agent can do most of the work. Meta's repo has AGENTS.md files written for coding agents. In Meta's words, "Any agent that reads AGENTS.md can build and flash from this repository." Meta's Muse Code runs with --disable-sandbox so it can reach your board's USB serial port. The hackshop brief tells the agent to use ESP-IDF v6.0.1 only, to find the right serial port before flashing, and to keep your SDK token as a placeholder instead of committing it.

Some steps stay with you: approving and paying for parts, pairing the gadget in the Muse app on your phone, and deciding where it lives. The usual Muse terms apply too. SDK tokens are for personal, non-commercial use on up to 50 devices, and Meta says the SDK "may change, stop working, or be withdrawn at any time." For the hands-on version, see build an AI agent body in an afternoon.

Questions people ask

Do I need an API key to use the hackshop MCP server?

No. plan_gadget, get_build_plan and assess_hackability never need a key, and the hosted server at https://www.hackshop.dev/mcp needs no account.

Does it work with ChatGPT and Codex?

Codex connects through its config.toml file. In ChatGPT, or any agent without MCP support, paste the prompt from the hackshop homepage; it points the agent to hackshop's agents.md, which carries the same rules.

Can my agent order the parts for me?

Only after you explicitly approve the exact items, sellers and total, and never on Amazon or eBay, where you check out yourself. The agent should never type card numbers or passwords.

Which boards can it recommend?

The boards Meta's Muse Gadgets SDK supports: twelve ESP32 boards in hackshop's catalog, Raspberry Pi models, and a few used mini PCs that may need a USB Bluetooth LE adapter.

Is the hackshop MCP server free?

Yes. Hackshop is free and open source, and the hackshop-mcp npm package is MIT licensed.

Sources

Prices and details checked October 5, 2026.