Skip to main content

AI Assistant Connector

EngCanvas publishes your cloud projects to an AI assistant over the Model Context Protocol (MCP). Once connected, you can ask the assistant to read your calculations, change them, run design checks and assemble a report — working on your real projects, not a copy you paste into a chat window.

One full-access URL works for every supported client. Only the paste steps differ. For multi-agent setups, also use the separate Reader URL so only one agent can mutate projects.

Keep this page and the app in step

The in-app MCP panel shows the same setup steps next to your live connector URL (apps/app/src/cloud/ClaudePanel.tsx). This page is the canonical version — if you change setup instructions here, change them there too.

Why a connector instead of a chat window​

EngCanvas doesn't embed a chat box, and that's deliberate:

  • claude.ai cannot be embedded — it blocks framing.
  • Third-party use of assistant subscription OAuth is prohibited by the assistant vendors.

So the conversation stays in the assistant you already pay for, and EngCanvas exposes itself as a tool that assistant can call. The practical upside: every token is billed to your assistant subscription, not marked up through EngCanvas.

Get your connector URL​

  1. Open EngCanvas
  2. Open the right panel and select the MCP tab
  3. Copy the URL shown for your client

The URL is minted per account and signed on the server. It is not something you can construct by hand.

Treat this link like a password

Anyone holding it can read, change and create projects in your account. Don't paste it into a shared document, a screenshot, or a support ticket.

The connector is included with a paid EngCanvas plan — it rides the same entitlement as cloud sync.

Setup​

Claude​

  1. In Claude, open Customize → Connectors
  2. Click + → Add custom connector
  3. Name it EngCanvas and paste the URL
  4. Click Add

Claude asks permission the first time each tool runs. To stop the prompts, open Customize → Connectors → EngCanvas → Tool permissions and set the tools you trust to Always allow — permissions are grouped by category, so you can approve read-only tools in one action and leave writes on approval.

ChatGPT (OpenAI)​

  1. Open Settings → Connectors. If you don't see custom connectors, enable Developer mode under Advanced / Security — availability varies by plan
  2. Click Create / Add custom connector
  3. Name it EngCanvas and paste the URL
  4. Prefer Streamable HTTP if offered a transport choice — not legacy SSE alone
  5. If authentication is required, choose None — the signed token is already in the URL
  6. Create the connector and enable it in a chat that supports tools

Grok​

  1. Open grok.com/connectors
  2. Remove any old EngCanvas connector first
  3. Click New Connector → Custom
  4. Name it something new (e.g. EngCanvas-v6) and paste the Grok (fresh tool list) URL from the MCP panel — it uses /api/mcp/v6/u/…?surface= so Grok/Cursor re-lists tools after surface upgrades
  5. If asked for authentication, choose None and leave credentials empty — the link already identifies your account
  6. Save, then start a brand-new chat and enable this connector

Multi-agent (one Writer + N Readers)​

When several agents share one EngCanvas account:

  1. Give exactly one agent the full Claude or Grok URL (the Writer — only this agent may create/edit projects)
  2. Give every other agent the Reader connector URL from the MCP panel (Multi-agent section). Its tool list has no mutators, which prevents sticky multi-writer thrash
  3. Optional: use Copy multi-agent harness for a paste block with both URLs and tool allowlists

Readers can run catalogs, check_expression, and reads in parallel. They cannot mutate projects.

Microsoft Copilot Studio​

  1. Open your agent and go to Tools
  2. Click Add a tool → New tool → Model Context Protocol
  3. Name it EngCanvas and paste the URL
  4. Set Authentication to None, then Create
  5. Click Create a new connection → Add to agent

Copilot Studio requires Streamable HTTP (it dropped SSE in August 2025), which this endpoint speaks. Authentication: None is correct rather than a shortcut — the signed token in the URL path is the credential, so there's no separate key for Copilot to send.

What a connected assistant can do​

  • Read your projects — variables, calculations, sheets, drawings, documents
  • Report which design checks pass or fail
  • Create new projects from a conversation
  • Insert a calculation template and fill in your parameters
  • Edit calculations and sheet cells
  • Add design checks and results tables to a calculation
  • Write documents and draw simple geometry
  • Solve for a target value with the goal seeker
  • Assemble multi-section report PDFs
  • Add, rename and delete tabs, and set the title block

It cannot delete a project.

Things to ask it​

Open my Level 3 framing project and tell me which checks are failing.

Insert the CMU shear wall template and set f'm to 2000 psi and the wall height to 18 ft.

In the Loads sheet, add a load combination row for 1.2D + 1.6W and reference it from the report.

What beam depth gets the stress ratio to 0.90?

Build me a PDF with a cover page, the calculations and the drawing tabs.

Security model​

  • Your account id is bound to the connector on the server and is never something the assistant sends. A project id belonging to another account reads as "not found" — the scoping happens in the database query, not in the tool.
  • The assistant sees only what a tool returns. Reads return outlines and the assistant drills down, so a single question doesn't pull your whole project into its context.
  • Revoke access by removing the connector in your assistant's settings.