A takeoff engine built for AI agents first, estimators second

What it does A takeoff is measuring quantities off a construction drawing — square feet of flooring, wall lengths, fixture counts — and it’s the input to every construction bid. OpenTakeoff does it with two front ends over one engine: a browser canvas for a human estimator (no backend, no account, nothing uploads) and a stdio MCP server exposing 53 tools for an AI agent. Both share the same geometry and quantity modules, so an agent’s takeoff and a person’s corrections live in one document.

The interesting bit Provenance is the load-bearing part: every shape records the scale it was measured at, the method that produced it (flood, raster trace, hand-drawn, agent-proposed), and whether a human corrected it — with the machine’s original boundary frozen beside the correction. Downstream that’s an audit trail a PM can read; upstream it’s labeled geometry-to-finish pairs, the training signal takeoff models have never had at scale. The README claims this is the first open-source takeoff engine, period, and nothing here contradicts that.

Key highlights

• Runs entirely in the browser — client-only, works on Windows/macOS/Linux in current Chrome, Edge, Firefox, Safari; nothing installs, nothing uploads
• 53 MCP tools on the official MCP registry; an agent opens a plan, reads the title block, sets the scale, traces rooms to wall faces, and checks its own work on a rendered overlay
• Agent measurements carry an explicit review status; a human can import, correct a ring, and export with review state intact
• Stitched sheets: a floor split across a match line becomes one working surface, and a room crossing the seam traces as one shape
• Tracing conventions (innermost wall face, door openings on the wall centerline) ship inside the server itself, so agents follow estimator rules without prompting

Caveats

• One-Click room detection is temporarily gated while the flood engine is re-validated against a wider plan corpus — the one_click and detect_rooms MCP verbs aren’t registered in a default build; you trace rooms manually instead
• Browser and MCP room-detection paths currently differ, so shared code doesn’t guarantee identical boundaries on every plan
• The reference plan set’s estimator traces are v2 and still unreviewed by a person

Verdict If you’re in preconstruction, flooring, or quantity surveying — or building AI agents for AEC — this is worth a long look, especially since the field apparently had zero open-source engines before it. If you never touch a blueprint, move along.