TakeoffLens

AI drawing intelligence for reviewed, traceable quantity takeoff

TakeoffLens turns engineering drawings into quantities that remain connected to their source evidence. It combines PDF inspection, vector and layer analysis, symbol detection, assisted review, markup, and accuracy measurement in one local MCP-powered workflow.

Electrical and ELV are the first active disciplines. The shared Core + Discipline architecture is designed to expand to HVAC, Plumbing, Fire Protection, Architectural, Structural, and other building systems.

TakeoffLens is an assisted-review product. Automatic detections are candidates, not final quantities, until they have been reviewed.

Download v0.2.1 · Installation guide · Architecture · Development · Roadmap

Why TakeoffLens

Most PDF tools can display a drawing or extract text. TakeoffLens focuses on the engineering meaning around that content:

• identifies supported systems and symbol classes;

• uses native vector geometry and PDF/CAD layer evidence when available;

• keeps every counted item tied to page coordinates and review status;

• exports candidate crops, markup, CSV, JSON, and audit evidence;

• records source, template, and candidate hashes for reproducibility;

• measures precision, recall, F1, and where detector misses occurred;

• preserves uncertain items instead of silently converting them into quantities.

Related MCP server: PDF2MD MCP Server

Product capabilities

Supported disciplines

get_discipline_catalog is the machine-readable source of truth. Planned packs must not be presented as supported until their rules, templates, and reviewed benchmarks are available.

How it works

flowchart LR
    A["Engineering PDF"] --> B["Inspect page and layers"]
    B --> C["Apply discipline rules and templates"]
    C --> D["Generate candidates and evidence"]
    D --> E["Accept, reject, or mark uncertain"]
    E --> F["Export reviewed quantities and markup"]
    F --> G["Benchmark and audit provenance"]

The engine prefers CAD metadata when available, native vector geometry for vector PDFs, and high-resolution visual review for raster scans. OCR is text context only; it is not treated as proof that a symbol exists.

Quick start

Install as a Codex Plugin

1. Download takeoff-lens-plugin-v0.2.1.zip from the latest release.

2. Verify the accompanying .sha256 file.

3. Extract the archive.

4. Add the extracted plugin-marketplace directory to Codex.

5. Install takeoff-lens, then start a new task so the Skill and MCP tools load.

The first MCP launch creates a private per-user Python environment and installs the pinned headless dependencies. Drawing files are processed locally.

See Installation for exact commands, migration from the legacy plugin, manual MCP setup, managed Python environments, and troubleshooting.

Example prompt

Use TakeoffLens to inspect this engineering PDF.
Prepare a FIRE_ALARM audit, create markup, and review every candidate.
Keep uncertain locations unresolved and show their crop or coordinates before
reporting a final quantity.

Review contract

A result is final only when all of the following are true:

• every detector candidate has an accepted, rejected, or uncertain decision;

• no unresolved or uncertain item affects the subtotal;

• the independent wall and door sweep is confirmed where relevant;

• review_complete is true;

• clarification_required is false;

• provenance_verified is true.

Visually confirmed misses are recorded as manual_points; they are never added silently. A zero result remains provisional until the candidate set and relevant plan regions have been reviewed.

Output artifacts

TakeoffLens can produce:

• candidates.json and candidate crops;

• candidate_pool.json with shortlisted, filtered, and ranked-out records;

• detection_manifest.json with SHA-256 provenance;

• an HTML candidate review page;

• confirmed CSV and JSON quantities;

• marked-up page images with stable detection IDs;

• accuracy reports containing TP, FP, FN, precision, recall, and F1.

Accuracy status

The current benchmark foundation rejects incomplete ground truth and verifies the source PDF hash before evaluation. The first private reviewed Fire Alarm baseline covers eight symbol classes and ten confirmed locations on one vector sheet. Layer-aware filtering reduced false positives from 83 to 25 while preserving recall at 1.000; shortlist precision is 0.286 and F1 is 0.444.

This is a regression baseline for one sheet, not a general production-accuracy claim. Broader multi-project benchmark coverage is part of the roadmap.

MCP tools

Privacy and project data

• Drawings are processed locally by the bundled MCP server.

• The repository and public release do not include customer PDFs, DWG/DXF files, marked-up plans, private ground truth, test outputs, or machine-specific paths.

• Private benchmark suites should remain under ignored directories such as test-runs/.

• Network access is needed only for first-run dependency installation unless a managed Python environment is supplied.

Platform and limitations

• The packaged launcher currently targets Windows and PowerShell.

• The vector matcher is intended mainly for vector and hybrid PDFs.

• Raster scans require high-resolution tiling and a separate vision/object detection workflow; this is not yet a production feature.

• Symbol geometry varies by company and project, so project legend confirmation remains essential.

• The optional Windows Desktop App is experimental. Product development currently prioritizes the MCP + Skill workflow.

Development

TakeoffLens is implemented in Python with PyMuPDF, OpenCV, NumPy, and Pillow. Start with Development for setup, tests, validation, benchmarking, packaging, and repository safety rules. See Architecture for the Core, MCP, Discipline Pack, and host-adapter boundaries.

Release history

See CHANGELOG.md. The current public release is TakeoffLens v0.2.1.

License

MIT