SpecCapture at a glance

Title block

• Part number
• Drawing number
• Revision
• Title
• Material
• Stock size
• Finish or coating, and plating specification
• Surface roughness
• Heat treatment
• Scale
• Units
• Sheet size
• Sheet number and sheet count
• Who drew, checked and approved it, with the dates
• Weight
• General tolerances for fractions, decimals (by number of places) and angles
• Projection angle
• Drawing standard
• CAGE code
• Export-control marking

Tables, every row

• Revision history: revision, description, date, approved by
• Bill of materials: item number, quantity, part number, description, material
• Hole table: tag, size, quantity, depth
• Application block: next assembly, used on
• Tabulated dash numbers: dash number, description

Notes

General notes, each in full and in order.

It reads the facts a quote starts from, such as the material, finish and general tolerances, but it does not price a job, and it does not read every dimension or balloon a drawing for inspection.

• It accepts PDF, TIFF, PNG, JPEG, BMP and WebP files; GIF files are not accepted.
• A PDF or TIFF file can hold up to 200 sheets; an image is one sheet.
• A file can be up to 50 MB, and a sheet up to 150 million pixels, which fits a D-size sheet scanned at 400 dpi, or an E-size or A0 sheet at 300 dpi.
• Each sheet is read on its own, in the background: on our test drawings a sheet took 36 to 49 seconds on average.
• Results download as Excel (XLSX), CSV or JSON, or come through a REST API, which can also send a signed webhook when a file is finished.
• Each uploaded file exports on its own, as one workbook, one set of CSV files or one JSON file, so to keep a batch together, combine its sheets into one PDF or TIFF file, or collect every file's values through the API.
• Through the API, each value comes with a confidence score from 0 to 1, a review state and, while the drawing is kept, a bounding box in page coordinates from 0 to 1, measured from the top left of the sheet.
• Paper drawings are read from scans: scan each sheet to PDF or TIFF, at up to 150 million pixels a sheet (a D-size sheet scanned at 400 dpi, or an E-size or A0 sheet at 300 dpi), and the values come back as a spreadsheet you can sort and filter, or as JSON.
• SpecCapture also works from AI assistants such as Claude, GitHub Copilot in VS Code and Cursor: its MCP server at https://api.speccapture.com/mcp lets them send drawings and read their values, with the same API keys.

• SpecCapture reads each sheet's printed text with OCR, which works on scans as well as PDFs, has an AI model pick out each field's value, and boxes each value where its text was found on the sheet.
• Every value it reads comes with a box around the place on the drawing it was read from, so checking it takes a glance.
• Rules that work the same way every time check dates, revision letters, material specifications and units.
• Each value gets a confidence level, and a value that fails a rule, was written by hand, or that SpecCapture is unsure of is flagged for a person to check.
• A new workspace starts with every value waiting for a person to confirm it; a field is accepted automatically only after people have confirmed its values 20 times in a row with no correction, and even then only when SpecCapture is confident in the value.
• Exports mark every value still waiting for review, so nothing unchecked passes as final.

• SpecCapture starts with a free trial: 25 sheets over 14 days for $0, with no card needed and every feature included.
• A sheet is one page of a drawing file, so a three-sheet PDF uses three.
• Prices are per workspace: everyone in it shares the plan, with no charge per user.
• SpecCapture is not open for sign-ups yet; the free trial and these prices are what it is planned to open with.

Credit packs

Credit packs need no plan: 60 sheets for $60, 275 for $250 or 700 for $600, each one payment. A pack's sheets count for 12 months, and a workspace on a plan uses its plan's sheets first.

What happens when the trial or a workspace's sheets run out: Pricing.

• On PDFs exported from CAD, in drawing styles it has been tuned on, SpecCapture read the critical fields (part number, revision, material and finish) right 97.4% of the time.
• On drawings it was seeing for the first time it did worse: 85.6% on 35 new sheets in styles it knew, and 88.4% on 9 sheets in three styles it had never seen. Your drawings are new to it, so the first-run figures are the better guide until it has been checked on them.
• Scans are weaker: the critical fields came out right 82.1% of the time on 1960s microfilm scans and 58.6% on drawings scanned from old printed reports. On the microfilm, most wrong values came from damage on the scan itself, which no reading can undo.
• On every test set with labeled parts lists, PDFs and scans alike, it found every row.

How it was measured

• Every test drawing has an answer key: the value printed on the sheet for each field, written down separately from SpecCapture's reading.
• SpecCapture read every sheet three times and the figures count every reading, because an AI model does not answer exactly the same way twice.
• A value counts as right only when it matches the answer key; a value it missed, or misread because the scan itself was damaged, counts as wrong.
• The test drawings are public-domain and openly licensed drawings from open-hardware projects and US government archives, including NASA; none is a customer's.
• Every percentage is rounded down.
• Measured on October 6, 2026 and October 7, 2026.

• You choose how long drawings are kept, 0, 7, 30 or 90 days; then the file, its page images and the marks showing where each value was read are deleted automatically, and the values read from it stay until you delete the drawing or your workspace. New workspaces start at 30 days.
• Your drawings are not used to train AI models that anyone else uses.
• Files and extracted data are stored in a US region.
• Each workspace's data is kept apart from every other's, enforced by the database itself.
• SpecCapture is not for export-controlled drawings, such as those covered by ITAR or the EAR.

More on the trust page.