Create an interpreter with AI

How it works

1. In the hub, open My interpreters and follow AI draft (from a specification) (hub.trace-analyzer.net/interpreters/ki-entwurf).
2. Upload the specification: a PDF (up to 600 pages) or a text file such as .txt, .md, .dbc, .h or .xml, at most 20 MB. Scanned PDFs without a text layer are harder for the AI to read.
3. Check the protocol name. If you have one, paste a capture from the real device as hex, and say in the notes what matters to you (for example “only the status messages”).
4. Choose the cost limit, confirm that you may pass the document on, and start. You may close the page meanwhile; a draft usually takes a few minutes.
5. Review the result: the report (what is covered, what is missing, which assumptions the AI had to make), the check per example and the XML. If you like it, create it as an interpreter in your library — under an existing name as a new version.

What happens to your document

Only after you confirm it for the individual job, the document, the protocol name, your notes and your example bytes go to Anthropic, the provider of the AI model, and are processed there in the USA. Account data such as your name or e-mail address are not sent.

Whether you may pass a document on is your decision alone. This applies in particular to specifications under a non-disclosure agreement (NDA). Trace does not check this and takes no responsibility for it; if in doubt, ask the publisher of the specification.

On the hub, the document is kept for seven days after the job ends and then deleted — immediately if you remove the job or delete your account. Details are in the privacy policy.

What it costs

On trace-analyzer.net every AI draft costs credits, billed by actual usage: by the amount of text the AI reads and writes, across all check rounds. Before you start, the page shows an estimate; a large document costs more than a small one.

• For each draft you set a limit (500 credits suggested, at most 2,000). You are never charged more than this limit; when it no longer covers the next check round, the work ends with the best version so far.
• Each account has a monthly limit of 3,000 credits for AI drafts; the page shows how much of it is used.
• One draft runs per account at a time. Guest accounts cannot create drafts.
• If you cancel a running draft, it stops after the check round in progress; that round and all before it are charged.

How good is the result?

The AI checks every version itself: the hub loads the definition like any other, runs it over the examples from the specification and over your capture, and reports per example whether every byte is in a frame and no frame counts as faulty. If the protocol has a checksum whose parameters the document leaves open, the AI identifies it from example frames using a catalog of common algorithms (sums, XOR, LRC, CRC-8/16/32).

It is still a draft. Examples from a document only show that the definition matches the AI’s reading of it; real bytes are the stronger proof. Review the report — especially the assumptions — and test the interpreter on a real recording before you rely on it.

On your own hub

The feature is available on a self-hosted hub as well. It is switched on with your own Anthropic API key in the environment variable Hub__AiDraft__ApiKey — never in a file. Anthropic then bills the operator directly; no Trace credits are charged. Model, effort and limits can be set under Hub:AiDraft.

Other ways to an interpreter

• The interpreter assistant in the hub builds a definition without AI from marked example bytes; it identifies checksums too.
• Your own AI: through the hub’s MCP access (page AI access), an AI program on your computer — such as Claude Desktop — can read the specification and work with the hub’s tools check_interpreter and identify_checksum. The document then goes to your program’s AI provider, not to the hub’s.
• By hand with the TrIML reference from the documentation.