What is OpenTakeoff?
OpenTakeoff measures construction drawings for bids: floor and wall areas, lengths, counts and the material quantities that follow, at the scale of each sheet. It has two front ends on one engine. People drag a PDF, image or zipped plan set into a browser canvas, set the scale, trace rooms and export a report and a marked-up planset; AI agents drive the same geometry through an MCP server with 58 tools over stdio.
Agent work arrives in the canvas as dashed proposals that a person accepts or rejects, and every shape records its scale, the method that drew it and whether a person or an agent drew it. Exports include CSV, JSON, Excel, DXF, a marked-up PDF, an estimate and a bid proposal PDF. Everything runs in the browser and is stored in IndexedDB and localStorage; nothing is uploaded.
It is written in TypeScript by Kentucky AI, whose founder runs estimating for a commercial flooring company. The code was Apache-2.0 up to the tag apache-2.0-final and is now AGPL-3.0. The MCP package, opentakeoff-mcp, reached 0.9.100 on 25 October 2026. Kentucky AI also sells a hosted Premium workspace with its own vision and language models.
Is OpenTakeoff worth it?
OpenTakeoff is a serious attempt at agent-driven quantity takeoff, and the design shows someone who bids work for a living. Measuring an unscaled sheet is refused rather than defaulted; rooms are traced on the inner wall faces; when a symbol sweep or derivation is unsure it says why and returns coordinates instead of a plausible number. An agent can sign its own work, but only a person at the canvas can apply the APPROVED seal.
The agent side is large. The MCP server exposes 58 tools grouped by stage, from load_plan and set_scale through measure_polygon, symbol_sweep and sweep_schedule_row to create_rfi, set_pricing and export_marked_pdf. The project's demo shows an agent taking off three patient rooms on a federal finish plan, 743.64 square feet in all, each export landing in the canvas as a proposal for the operator to accept. That many tools costs context on every session, and prices always come from the estimator, never from the agent.
It is a young project with one feature switched off. One-Click Area, the flood-fill room detector, is temporarily gated while it is re-checked on more plans, so rooms are traced by hand or with measure_polygon for now. The repository was created in June 2026, has 754 stars and 16 contributors, and ships MCP releases often; October alone added estimating, a sheet tree and a faster room engine.
The licence changed. Code up to the tag apache-2.0-final stays Apache-2.0, but current versions are AGPL-3.0 with a contributor agreement and a section 7(b) attribution term, so a firm that runs a modified copy for clients must publish its changes. Against paid desktop takeoff software, its case is narrower and stronger: it costs nothing, keeps plans on your machine, and was built from the start for an agent to do the measuring while a person keeps the final say.
Where it earns its place
An estimator at a five-person flooring subcontractor
Measure finish areas on a plan set without paying for takeoff software
Open the web app, run the guided first takeoff, set the scale per sheet and trace rooms with Area; conditions carry waste percentages and export to Excel and a marked set PDF. The software costs nothing and plans stay in the browser.
A developer at a construction-tech startup
Have Claude measure rooms on a finish plan and hand the result to an estimator
claude mcp add opentakeoff -- npx -y opentakeoff-mcp, then ask the agent to load the plan, read the scale note, set_scale and trace rooms with measure_polygon. Its shapes arrive in the canvas as proposals the estimator accepts; model costs are on your own API key.
A general contractor's preconstruction team
Run a private copy for staff without sending drawings to a vendor
Fork the repository and deploy it to Netlify or any static host; nothing phones home. Internal use carries no obligation, but under AGPL-3.0 a modified copy offered to clients must publish its changes.
Before you commit
- Setup
- No install for the canvas: open opentakeoff.kentucky-ai.com. The MCP server needs Node 20 or later and runs with npx -y opentakeoff-mcp; documented for Claude Code, Codex CLI, Cursor, VS Code, Claude Desktop (an .mcpb file) and Docker.
- Integrations
- 58 MCP tools over stdio plus sheet resources and a takeoff://wiki resource; imports PDFs, images and zipped plan sets; exports CSV, JSON, Excel, DXF, marked PDF, estimate and proposal PDF; listed on the official MCP registry.
- Data and privacy
- Client-only: work is kept in IndexedDB and localStorage and never uploaded. Opt-in Contribute sends derived geometry only to a capture server you run yourself. AGPL-3.0 for current code, Apache-2.0 up to the tag apache-2.0-final.
- Cost at scale
- The open-source app and MCP server are free for commercial work; agent sessions use your own model provider. Kentucky AI's hosted Premium is $99 a month for one seat with 300 AI credits, $299 for three seats with 800, and $89 per extra seat.
- Limits
- One-Click Area is gated pending re-validation; prices are never filled in by the agent; 22 open issues and pull requests; a large tool list adds context on every agent session.
What’s good
- Free, runs in the browser, plans never uploaded
- 58 MCP tools sharing one engine and file format with the canvas
- Scale gate and person-only approval make agent work reviewable
- Every shape records scale, method and author
- Exports marked-up PDF, Excel, DXF, CSV and bid proposals
What’s not
- One-Click Area room detection temporarily gated
- Relicensed from Apache-2.0 to AGPL-3.0 with a CLA
- 58 tools take a lot of agent context
- Young project; most features date from 2026
Best for
Flooring and finish estimators who want a free takeoff tool, and developers building AI agents that need to measure construction drawings under human review.
OpenTakeoff: common questions
- Is OpenTakeoff worth it?
- OpenTakeoff is a credible free takeoff tool built by a working estimator, and its MCP server is designed so an agent can measure but only a person can approve. We rate it 3.9 out of 5.
- How much does OpenTakeoff cost?
- Free under AGPL-3.0, including for commercial bids; bring your own model API key for agent work. Kentucky AI's hosted Premium is $99/month for one seat or $299/month for three.
- What are the downsides of OpenTakeoff?
- One-Click Area room detection temporarily gated; relicensed from Apache-2.0 to AGPL-3.0 with a CLA; 58 tools take a lot of agent context; young project; most features date from 2026.
- Who is OpenTakeoff best for?
- Flooring and finish estimators who want a free takeoff tool, and developers building AI agents that need to measure construction drawings under human review.
- How much does OpenTakeoff cost for a team or in production?
- The open-source app and MCP server are free for commercial work; agent sessions use your own model provider. Kentucky AI's hosted Premium is $99 a month for one seat with 300 AI credits, $299 for three seats with 800, and $89 per extra seat.
- Is OpenTakeoff safe for company data?
- Client-only: work is kept in IndexedDB and localStorage and never uploaded. Opt-in Contribute sends derived geometry only to a capture server you run yourself. AGPL-3.0 for current code, Apache-2.0 up to the tag apache-2.0-final.
Answers checked Nov 5, 2026.