Contracts, readverbatim by any agent.
Verbatim turns a Word agreement into structured data that software and AI agents can rely on: the clauses and their numbering, tracked changes and who made them, comments, defined terms and cross-references. Text, changes and comments are checked against the source on every call.
- Word .docx and .docm
- REST API and MCP
- No AI model
- Free during the public test
Why it exists
Plain text loses what lawyers read for.
Most tools turn a Word agreement into plain text before software or an AI assistant reads it. The tracked changes disappear, or run together with the text they replace. Comments are dropped. Clause numbers, which Word calculates rather than stores, are lost.
An assistant reading that text can't tell what was proposed from what was agreed, and can't cite clause 12.3(b) because it never saw the number.
Verbatim keeps all of it, in a form software can use.
What you get back
The agreement's structure, not just its words.
Clauses and numbering
The clause tree with the number each clause shows, like 12.3(b), worked out separately for the accepted and the rejected version of a redline.
Tracked changes
Every insertion, deletion and substitution, with its author and date, placed exactly in the text.
Comments
Comment threads with their replies and the text each one is attached to.
Defined terms and cross-references
Each defined term with its definition, and each reference to another clause with the clause it points to.
Three readings of every redline.
Each one has its own text and numbering, and every position in the result points back to the exact place in the source.
The agreement as it reads if every tracked change is accepted.
The agreement as it read before the changes.
Both at once, in order, with insertions and deletions marked so a reader can see what was proposed.
Three ways to use it
In a browser, from code, or from an assistant.
- 01
In the browser
Open the public test, choose a Word file and see its outline, changes, comments and defined terms.
Open the public test - 02
From your software
Send the file to the REST API. No key is needed in the public test.
curl --fail-with-body \ -H 'Content-Type: application/octet-stream' \ --data-binary @agreement.docx \ https://verbatim.agmt.legal/v1/parse - 03
From an AI assistant
Connect Claude Code, or any assistant that supports MCP, and it can read the agreement before it answers.
claude mcp add --transport http agmt-verbatim \ https://verbatim.agmt.legal/mcp
The full request and response formats, limits and error codes are in the API documentation, with setup for Claude, Cursor and other assistants in the agent guide.
Your documents
Unlike Execute, Verbatim runs on our servers.
An API has to receive the file to read it, so Verbatim processes documents on Agmt's servers. It's built to keep as little as possible, and we'll say so plainly only once it's been verified. Until then, test it with agreements that aren't confidential.
Processed in memory
The service is designed to read the file in memory and not write its text to application logs or databases.
No AI model
The parser is deterministic and makes no calls to an AI model. The same file and options give the same result every time.
Checked on every call
Text, tracked changes and comments are checked against the source each time. A failed check is reported, never hidden.
Still being verified
Production settings, monitoring and deletion are under review. Until they're verified, don't send confidential documents.
Verbatim has its own draft privacy notice and security page.
In the public test
What it doesn't do yet.
You should know a tool's limits before you rely on it. These are Verbatim's today. The benchmark page shows what has been measured and what hasn't.
- Word files only: .docx and .docm. No PDFs, old .doc files or password-protected files.
- Without a key: 10 documents a day, each up to 5 MiB and 100 page-equivalents of 500 words.
- Numbering is checked against LibreOffice, not Microsoft Word. Matching Word exactly isn't verified yet.
- Accuracy for defined terms and cross-references on real agreements hasn't been independently reviewed yet. Test results so far come from prepared documents.
- The JavaScript and Python libraries and the command-line tool are built but not published yet.
- Production hosting, monitoring and deletion are still being verified. Use agreements that aren't confidential.
Questions
Asked before trying it.
Who is Verbatim for?
Is it free to try?
Where are my documents processed?
Does it use AI?
Is Execute connected to Verbatim?
How do I report a problem?
Try it on a redline.
Pick a Word agreement with tracked changes and comments, one that isn't confidential, and see what Verbatim reads from it.