Electronics Architect

npx add-mcp 'https://electronics-architect.com/mcp'

Let your assistant do the power tree

Connect Claude or any MCP client to the same solver the editor uses. The assistant proposes the tree and solves it; the design lands in your account, on the canvas, for you to check.

There is one address to remember:

https://electronics-architect.com/mcp

Paste it into your assistant as an MCP server — MCP is the standard way an AI client talks to a tool — and it gets twenty-one tools: read the design contract, list the shipped templates, validate and solve a design at every tolerance corner, run a Monte Carlo (as a job it polls when the run is too long for one call), grade every part against NASA GSFC or IPC-9592 derating, look up a part model, save, share, compare and export. Everything it saves appears in My Designs, and the link it hands you opens that design on the canvas with a note saying which key made it and what changed. Solving is unmetered. Pro and Team plans — $10 a month or $100 a year for Pro.

Three ways to connect

Which one depends on the client you use. All three end up at the same server with the same tools.

1. By address — the client signs you in

In Claude Desktop or the claude.ai connector settings, add a custom connector with the address above. In Claude Code:

claude mcp add --transport http electronics-architect https://electronics-architect.com/mcp

On the first call the client opens a sign-in page; you tick the permissions you want the assistant to have. That creates an ordinary API key labelled with the app's name under Account → Developer, and revoking it there disconnects the app at once.

2. By key — a bearer header

Some clients connect to a remote server but do not run the sign-in flow; their MCP settings take a server address and a set of headers. For those, create a key in the editor under Account → Developer and send it as a bearer header — the same header the REST API uses:

{
  "mcpServers": {
    "electronics-architect": {
      "serverUrl": "https://electronics-architect.com/mcp",
      "headers": { "Authorization": "Bearer ea_live_…" }
    }
  }
}

Clients name the address field differently: some read serverUrl, others url. The key is shown once, when you create it. Give it a label you will recognise — that label is what the canvas shows beside a design the assistant made.

3. By bridge — for clients that only run local servers

Older clients only start a server on your own machine. For those, the community package mcp-remote runs locally and carries the connection to our address, sign-in included:

{
  "mcpServers": {
    "electronics-architect": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://electronics-architect.com/mcp"]
    }
  }
}

mcp-remote is a community tool, not ours. It keeps the sign-in it performs on your disk. What it holds is an ordinary key from your account — if you stop trusting it, revoke that key under Account → Developer and the connection is dead. We do not publish a local server of our own: the solver runs only on our side.

Or skip the assistant: REST

Everything above is also a plain HTTP API for scripts and CI. The reference — keys, scopes, quotas, every endpoint and a worked create → solve → Monte Carlo → export flow in curl, Python and TypeScript — is at /developers/api. The machine-readable contract is /v1/openapi.json (OpenAPI 3.1); the design document schema is /schema/design-v1.json. Agents can read the whole of this site in one file at /llms-full.txt.

Three workflows

These are the sequences the assistant actually runs, as recorded in our acceptance runs against the live server — Claude Desktop on 10 September 2026 and Claude Code on 11 September 2026. Both finished with a saved, solved design and zero warnings at all three tolerance corners; both also found a solver defect that was fixed the same day, which is the honest reason the runs are worth reading.

A. Design from a brief, hand it back

"12 V in. I need 5 V at 1.5 A, 3.3 V at 2 A and 1.0 V at 6 A for an FPGA. Verify it and give me the PDF."

1. get_design_schema — the units rule and the node types, before writing anything.
2. get_component_model for each rail — real regulators with their datasheet values filled in.
3. validate_design, then solve_design at nom, min and max.
4. Read the warnings — each carries a code and a remedy — fix, and solve again until all three corners are clean.
5. create_design, and hand back the editorUrl it returns. The PDF is drawn in the editor; the link opens the design there.

The Claude Desktop run took 20 tool calls, chose three synchronous bucks off the 12 V bus, and noted that a linear regulator on the 1.0 V rail would have dissipated 66 W. Nominal efficiency came out at 88.7 %.

B. Review a design you already have

In the editor, open the design and use Share → Copy MCP snippet. Paste it to the assistant.

1. get_design with the id from the snippet — the assistant already holds your key, so the id is all it needs.
2. solve_design with compact: true for the numbers, then review_design on the design document for the plain-English read and the concrete changes. The reviewer solves the document itself and checks every part it suggests against that part's worst-case input voltage.
3. update_design with a one-line message saying what changed.
4. When you next open the design, the canvas says the assistant changed it; See what changed shows the previous state with Restore one click away.

C. Change something and see what it did

1. compare_designs — the saved design as a, the edited document as b: what changed, and what it did to loss, efficiency and warnings.
2. Add monteCarlo: { trials, seed } to see the yield move with the same seed on both sides, so the movement is the change's.
3. export_design as report_json — the summary, the grouped warnings with their fixes, and the BOM, with the attribution line that names the solver release and the design fingerprint.

The Claude Code run built a 34-node, seven-rail power tree for a 64-element X-band tile this way — 350 W in at 86 % nominal — and kept its own record of where the effort went, which is how the compact solve result and get_template came to exist.

What it costs

Free accounts use the editor; connecting an assistant needs Pro, and a Free account that has not had a trial can start Pro's 7-day free trial on the sign-in page and connect straight after. Pro is $10 a month or $100 a year — the annual plan is ten months' price — and includes API and MCP access, multi-sheet designs, clean PDF and BOM export and 250 Monte Carlo runs a month. Team is $15 a seat a month or $150 a seat a year, with shared designs and unlimited Monte Carlo. Solving is never metered; Monte Carlo and AI review count against the monthly allowance, from the editor or from an assistant alike. Plans.

Questions and bug reports: the contact page. The terms of service and the privacy policy apply to an assistant's use as they do to the editor.