Connect an AI assistant to Arkim

Arkim speaks MCP (Model Context Protocol) — the open standard AI assistants use to connect to apps. Connect one to your account and it works on your real projects: author a part as a document, assemble it with real joints, run a simulation, read a cut list, cut drawing views, export machinable STEP. The same validators that stand behind the UI stand between it and a bad part, and it reads their refusals verbatim.

The server address is the whole setup:

Access is per person. You sign in with your own Arkim account and approve the connection before anything is shared — there is no API key to mint and no client id to copy. An assistant acts as you, and it can never reach a project in an organization you are not a member of.

Pick your assistant — every client follows the same add server → sign in → approve shape.

Heads up: desktop and command-line apps finish sign-in by sending your browser to a http://localhost:…/callback address. That is normal — it is the app on your own machine collecting its sign-in code, the same way the GitHub CLI logs in. Nothing is exposed to the internet.

Which account the assistant acts as is settled at sign-in: whoever you sign in as. Which project it works in is settled per call, not per connection — an Arkim connection carries no “active project” the way the app’s rail does. Every tool takes an explicit project id, and membership is re-checked on every single call, so the assistant reaches exactly the projects in the organizations you belong to and nothing else. Revoke someone’s membership and their assistant is refused on its next call, not at its next sign-in.

So open with “list my Arkim projects”. One call comes back with every project in every organization you are in, and the rest of the conversation names one.

You can also connect more than one account at once — your own and a client’s, say. Each is a separate server on the assistant’s side, with its own name and its own sign-in (arkim, arkim-acme). The tools are named identically across connections, so the server name is the only thing in a conversation that says which account is about to act — name them for the identity, not the product.

Read — your projects and their variables; parts of every kind with their source and parameters; assemblies, drawings, drafts, terrains, folders, saved configurations, and the full revision history of any of them. It can render a part or a posed assembly to an image, cut third-angle drawing views, and — more to the point — measure: mass properties (volume, area, bounding box, centre of mass, solid count, mass), a weldment’s trimmed cut list, a circuit’s netlist and BOM. It can export machinable STEP and a board’s Gerber fab zip, and run rigid-body and SPICE simulations.

Change — create and update parts, assemblies, drawings, drafts and terrains; save configurations; file things into folders; name a checkpoint in history; release, un-release or restore a revision; save a simulation run. Every write runs the platform’s validators first: a part that does not render, an assembly that fails its document check, a draft that does not resolve is refused with the reason, in words, and nothing is written.

Deletes, and only lossless ones — a draft, a terrain, or a folder. A deleted draft or terrain goes to the trash, restorable for 30 days. A folder holds no design at all: deleting one re-parents its contents one level up instead of taking them with it, and a folder with anything in it is refused outright unless the call says so explicitly.

Never — parts, assemblies and drawings cannot be deleted over MCP at all, and neither can a project. Nothing about people or access is exposed either: members, invitations, client-portal publications, organizations, your profile, your password, your two-factor. Those stay in the app, done by a person.

Arkim files a full snapshot on every save — from the UI and from an assistant alike — with who saved it and, optionally, a message. So an assistant’s work is never a mystery you have to reconstruct: open History on the part, assembly, drawing, draft or terrain, read a side-by-side diff against what is there now, and restore if you would rather it hadn’t. Restoring is itself a save, so it is never a one-way door: the version you restored over is still in the history behind it.

You can also just ask — “what did you change, and what did it look like before?” — and the assistant will read the same history back to you.

One thing worth knowing before you turn an assistant loose on geometry: a picture cannot settle a geometry question. Every wrong part in this system’s history rendered as a perfectly plausible one — a slot whose caps bulged inward, a mitre that silently did nothing, a bend that folded down instead of up. All of them were caught by a number, none by looking.

The server tells every assistant exactly that, and hands it the means to check itself: mass properties to hold a volume against a closed form, a cut list whose wrong length means a wrong joint, drawing views with drafting-true extents. Ask for the number alongside the render and you will catch what the render cannot tell you.

The authoritative switch is in your account: Settings → Connections lists every assistant you have approved and cuts one off on the spot. Revoking destroys its tokens rather than leaving them to expire, so it loses access on its very next call — a conversation working in Arkim right now stops there. Nothing it already created is affected.

Also remove the server on the assistant’s side, or it will keep trying:

Reconnecting always goes back through sign-in and approval. Removing it on the assistant’s side alone is weaker but not nothing: an access token lasts an hour and renews against a refresh token good for a week, so a client you have removed ages out on its own — the panel is what makes it immediate.

Something looks wrong, or a client will not connect? Email support@arkim.app. The rest of the platform is described on the home page.