Who handles your data#

Applies to v0.11Last reviewed

In short

The companies that process your content on Zymcos's behalf, what each one sees, and what each has committed to. Nothing you make is used to train anyone's models.

Zymcos does not make pictures, clips, speech or CODA's replies by itself: it sends the request to a specialist company and brings the result back. This page says which companies, what each one sees, and what each has committed to about keeping and training on it. The commitments are from each company's own published terms, and the date each was read is recorded in our records; where a company does not say, this page says "not stated" rather than guessing.

The companies#

Anthropic (Claude) — CODA's conversation, script drafting and the documentation's question box. When you talk to CODA, your message, the relevant parts of your production (the script, the Bible, project notes CODA has kept) and your plan's limits go to Anthropic's Claude model, which writes CODA's reply. When you ask CODA to draft a script or a scene, the request goes to Claude too. When you ask the documentation a question, the question and the few documentation sections that best match it go to Claude, which writes the answer from those sections alone. Anthropic's commercial terms say it may not train models on customer content, and that it deletes inputs and outputs from its systems within 30 days; content its safety systems flag can be kept for up to two years. It indemnifies customers against third-party intellectual-property claims over paid use and its outputs. It does not commit to a storage region.

Cohere — finding the right documentation, and the backup. When you ask the documentation a question, Cohere turns the question into numbers to find the closest documentation sections and ranks them; it does not write the answer. Cohere is also the backup: if Claude cannot answer for a moment, Cohere's Command model writes the reply, the draft or the documentation answer instead, from the same material, and CODA says so under a reply it did not write with Claude. Cohere's logged prompts and generations are deleted after 30 days. Cohere trains on customer data unless an account setting turns that off; Zymcos's account has it turned off. Cohere indemnifies customers against third-party intellectual-property claims. It does not commit to a storage region.

fal — pictures, clips and speech. Every take is made by a model running on fal, which receives the prompt and any reference pictures, and returns the result. fal stores a request's inputs and outputs for 30 days by default; its privacy policy does not state whether it trains on them. fal does not indemnify customers for the output, and its terms do not promise that an output is original or non-infringing, which is why Zymcos records where every take came from and who affirmed the rights on what it was made from.

DigitalOcean — where your files live. Your stored files (the copies Zymcos keeps of every take, your uploads, your references) are in DigitalOcean's object storage in Singapore. DigitalOcean hosts the files and does not read them.

What never leaves Zymcos#

Your sign-in, your workspace's members and roles, your spending records and your audit trail are held by Zymcos itself. A reply CODA gives is checked against what CODA actually did before you see it (Welcome to Zymcos), and nothing is generated or paid for without your click (Spending limits).

When this changes#

Adding a company to this list is a decision the owner makes on the record, with the company's terms read first, and the list is updated in the same release that first sends anything to it. If you need a copy of the terms we relied on, or have a question this page does not answer, write to support.