AgentCM

For construction teams who want Claude Code to read their drawings

Makes a construction project folder as navigable to AI as a codebase: drawing sets, specifications and schedules turned into an index and a database that Claude Code can find its way around.

  • TypeScript
  • React
  • PostgreSQL
  • Poppler
  • Claude Code

The problem

Construction projects live in hundreds of PDFs, spreadsheets and markups spread across nested folders. The information is spatial, cross-referenced and locked inside pixels. Claude Code navigates a repository with speed and precision, but it cannot read a 96-page bid set the same way: it does not know what sheet A-2.01 contains, where detail 5/A7.10 points, or which rooms a door schedule covers.

What it does

AgentCM reads the project folder, extracts the text of every drawing sheet, helps you check it and teach it how your project's tags look, then writes an index next to your files and opens a read-only database, so the agent can find its way around.

$your project, after AgentCM
your-project/ ├── Drawings/ your files, where they were ├── Specifications/ └── .construction/ written by AgentCM ├── CLAUDE.md where the agent starts ├── INDEX.md every sheet and spec section ├── sheets/ each sheet's text, in reading order ├── schedules/ one table per schedule ├── specs/ one file per spec section ├── tags.tsv every room, door and grid tag ├── xrefs.tsv every reference between sheets └── query.sh one read-only database query
  1. Open the folder. AgentCM indexes every file and classifies it: drawing sets, specifications, RFIs, submittals, schedules.
  2. Extract. Each page's text comes from the PDF's own text layer; only scanned pages go to OCR. Sheet numbers are read from the title blocks for you to confirm.
  3. Review. Fix the text on a canvas over the drawing, and apply a fix to the same spot on many sheets at once.
  4. Teach and detect. Show it a few room tags, door tags, callouts or grid bubbles on one sheet; it finds the rest across the set, and you accept or reject each find.
  5. Navigate. Accepted finds become rooms, callout links, views, notes, grid lines and schedules the agent can query.

What you can ask

You askThe answer comes from
"List all architectural sheets"the sheet index: number, title and discipline of every sheet
"Where does detail 5/A7.10 point?"the accepted callouts and the sheets they point to
"What finishes does room 101 get?"the room profile: its level and the schedule rows linked to it
"Show the door schedule"the schedule's table, read from the drawing's text
"What does spec 07 92 00 say about sealants?"the section's text, read from the project manual

How it is built

Everything runs on your machine: a local web app (Express and React), PostgreSQL in Docker, and Poppler to render pages and read their text. The only outside service is optional OCR for scanned pages. Two principles hold throughout: reference, don't copy (your files stay where they are, and change only when you ask), and meet projects where they are (a project with months of history onboards as it is).

What an agent may rely on, every file and column in .construction/ and the read-only views, is a written contract that the test suite checks against the code. The Construction Skills plug into it: in a folder AgentCM has processed, they read its index instead of splitting sheets again, and send what they find back.

Open source from October 16, under the AGPL-3.0.