Let your agent do the power tree

Ask an assistant for a power tree today and you get a plausible answer: three bucks, an LDO for the noisy rail, a paragraph of reasons. What you do not get is the input current at low line, the dissipation in the part that turns out to be hottest, or the rail that only fails at one tolerance corner. Those numbers come from a cascade that has to be iterated to convergence, and a language model does not iterate to convergence. It guesses.

So we gave it the solver. Electronics Architect is now an MCP server — the standard way an AI client calls a tool — at one address, https://electronics-architect.com/mcp. Connect it and the assistant gets twenty-one tools that run on our side: read the design contract, validate a design, solve it at the nominal, minimum and maximum corners, run a Monte Carlo (as a job it polls when the run is long), grade the parts against a derating standard, look up a part model, save, share, compare and export. Every number in its answer is then a solved number, not a guessed one. Solving is unmetered.

What a run looks like

On 10 September we connected Claude Desktop to the live server, signed in through the consent screen, and gave it the brief we use as our acceptance test: 12 V in; 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. Twenty tool calls later:

It read the schema before writing a line, noted that a linear regulator on the 1.0 V rail would have burned 66 W, saved the design to the account, and handed back a link. It declined to draw the PDF, correctly — that is rendered in the editor — and gave the link instead, which is exactly the hand-off we want: the assistant produces, a person opens it on the canvas and decides.

The next day a Claude Code session built something we had not scripted — a seven-rail, 34-node power tree for a 64-element X-band transmit/receive tile with gallium-nitride amplifiers on a 48 V bus. All three corners clean; 350 W in at 86 % nominal. It also wrote down, unprompted, where its effort had gone: every solve returned the whole document, the library had nothing at that power class, and the port conventions were not documented. Three of those became changes the same week.

What the runs found

Both runs found a defect in the solver. The first: a regulator entered with a no-load efficiency of exactly 70 % came back as 75 %, because a migration written for very old saved files treated exactly 70 as the old default and bumped it — 69 and 71 survived, 70 did not. The second: an RF amplifier's output power above about 50 W was silently divided by a thousand by a similar guess. Both were fixed the same day, on a failing test first, and both are the kind of thing a person clicking through a canvas would take months to hit and an assistant hits in an afternoon, because it types numbers a person would not think to. We consider that the second-best reason to connect one.

The hand-off, both ways

An agent's design is only useful if a person can look at it. When the assistant saves or changes a design, the link it gives you opens that design on the canvas — sign in if you need to — with a quiet banner saying which key made it and when. See what changed shows the previous state with Restore one click away. Nothing the assistant does is invisible, and nothing it does is final until you say so.

The other direction is one menu item. In the editor, Share → Copy MCP snippet puts a short instruction on your clipboard that any assistant connected to your account can use to fetch the design by id. Paste it, ask "why is U2 so hot", and the assistant reads the same design you are looking at. The warnings panel shows the same codes the API returns, so "clear the two thermal warnings on L1 and L2" means the same thing to both of you.

Connecting

Three ways, depending on the client. Claude Desktop and Claude Code take the address and sign you in. Any MCP client with a headers field takes the address plus a key from Account → Developer as a bearer header. Clients that can only run a local server use a community bridge, mcp-remote, which we describe but do not publish — our solver runs only on our side. All three are spelled out, with the config blocks, at /developers, along with the plain REST API for scripts and CI.

Pro and Team plans: $10 a month or $100 a year for Pro, $15 a seat a month or $150 a seat a year for Team. The annual plan costs the same as ten months.