Blog · Agents
What leaves your computer when an agent edits your video
· Updated · 6 min read
Short answer
Your source files stay put unless the agent is allowed to send them. What leaves is what it reads: your prompts, transcripts, project details and every frame it looks at. Whether that can train a model, and how long it's kept, depends on your account. Personal plans differ from business plans and API keys.
On this page
A personal plan, even an expensive one, is a consumer plan.
We checked each provider’s own pages on November 4, 2026. They change, so re-read them at the start of a job. This isn’t legal advice, and your client contract decides what you’re allowed to do.
What actually travels
Anything the model reads travels, and that’s more than footage. Claude Code’s data usage page says it plainly: it runs locally but sends all prompts and model outputs over the network, encrypted with TLS 1.2 or later.
| What | Does it leave? | Watch for |
|---|---|---|
| Your messages | Yes | Client names and plot details you type |
| Frames the agent looks at | Yes, as images | Size varies by tool. Each look is a new upload |
| Transcripts, file names, timeline details | Yes, once the agent reads them | File names can carry a project’s code name |
| Source media files | Not by themselves | An agent with a shell and the web can send a file if it’s allowed to |
| Feedback you submit | Yes, with the conversation | See the note under the next table |
| Local session logs | No, but they persist | Claude Code keeps plaintext transcripts in ~/.claude/projects/ for 30 days by default. Codex saves session transcripts under CODEX_HOME unless you change its history settings |
We found no separate rule that shields images from these terms. One image-specific caveat cuts the other way: OpenAI’s API data page says image inputs flagged by its child-safety scan are kept for manual review even under zero data retention.
What each provider does, by account
The account type decides. Consumer accounts may train on what you send, and commercial ones don’t by default. As of November 2026:
| Agent and account | Trains on it? | How long it’s kept |
|---|---|---|
| Claude Code: Free, Pro, Max, Ultra | Only if the setting is on. Even with it off, feedback and content flagged for safety review can be used | 5 years with training on, 30 days with it off |
| Claude Code: Team, Enterprise, API | No, unless you opt in through a partner program | 30 days standard. Zero retention for qualified Enterprise accounts. Flagged content up to 2 years |
| Codex: ChatGPT Free, Plus, Pro | Yes by default. Opting out covers Codex tasks, except Codex’s own Include environments setting | The pages we read give no period for local Codex tasks. Ask |
| Codex: ChatGPT Business, Enterprise, Edu | No by default | Admins set it. Deleted items go within 30 days, with legal and safety exceptions |
| Codex: API key | No, unless you opt in | Abuse logs up to 30 days. Zero retention by approval |
| Gemini CLI: personal Google sign-in | Ended June 18, 2026 for the free individual tier, AI Pro and Ultra. Google moved them to Antigravity CLI, a different product we didn’t review | n/a |
| Gemini CLI: unpaid API key | Yes, and human reviewers may read it. In the EEA, Switzerland and UK the paid terms apply | 55 days for abuse checks |
| Gemini CLI: paid API key | No | Logged for abuse checks. The policy page’s 55 days doesn’t separate paid from unpaid |
| Gemini CLI: Code Assist Standard, Enterprise | No | The page we read gives no period. Ask |
| Gemini CLI: Vertex AI key | No, without your permission | In-memory cache up to 24 hours. Prompts may be logged for abuse. Zero retention takes specific settings |
Which terms apply is set by how you sign in. Codex’s docs say a ChatGPT sign-in follows your workspace’s settings and an API key follows your API organization’s. Gemini CLI’s page maps each sign-in to its own terms.
Google’s warning is worth taking literally: “Do not submit sensitive, confidential, or personal information to the Unpaid Services.” And commercial doesn’t mean unseen: OpenAI says authorized staff and bound contractors may review ChatGPT Business and API data for abuse.
Feedback is the common trap. It overrides the setting you thought protected you. ChatGPT’s thumbs on an individual plan can put the whole conversation into training, even after opting out. Claude Code’s /feedback, /bug and /share send the conversation to Anthropic, kept for 5 years.
What a client may require
A personal account with training off can still fail a client’s rule. Netflix’s guidance for its owned scripted live action says proprietary and pre-release material, including footage, must go through a secure path: an enterprise tool, local or offline processing, or another tool it has vetted. Never a free, personal or unvetted one.
Other clients have their own language. If your NDA limits third parties, a model provider is one. A studio may also ask about the Trusted Partner Network, whose assessments, it says, produce a report rather than a certification. Ask what yours requires.
The checklist
Before the job
- Read the contract for AI and third-party language. If it’s silent, ask, and get the answer in writing with the agent and account type named.
- Check which account the agent is signed in to on this computer, today.
- Use a commercial account or an API key for client work where you can.
- On a personal plan, turn training off first: Claude’s privacy controls, or ChatGPT’s Settings > Data controls > Improve the model for everyone, plus Codex’s Include environments. OpenAI’s opt-out applies to new conversations.
- Never use an unpaid Gemini API key on client work.
During the job
- Don’t send feedback on a client conversation: no thumbs, no
/feedback, and answer No if Claude Code asks to look at your session transcript.DISABLE_FEEDBACK_COMMAND=1turns the command off. - Keep approvals on, so you see any step that reaches for the shell, files or web.
- Rename files that carry code names, and bring in only the scenes the job needs.
- Treat every plug-in or connected server as one more recipient with its own terms.
After the job
- Clear the local copies: agent session logs, and any frames saved in the project.
- Keep a note of which agent, which account and which dates. A client’s security questionnaire will ask.
Our own answer
nolanlabs is an editor on your computer that runs the agent you choose: Nolan Edit, which comes with it, or one of your own. Import is in place and the service is local, so footage stays on your disk. Frames the agent looks at go to the agent you chose. With Nolan Edit they go through nolanlabs.io, which keeps only counts of use, to the company that runs NolanX, its model, under that company’s terms. With your own agent they go to its provider, under the terms above.
What that means in the app. The agent’s look tool sends JPEGs, 640 px wide by default and up to the timeline’s full width. It shows sound as a waveform picture and loudness figures. A chip on your message shows what goes along (timeline, playhead, selection, marks), and you can click it to leave them out. Speech to text uses your computer’s own recognition, though the transcript’s text goes to the agent when it reads it. Runs are kept in the project’s agent/ folder, including the frames, so that’s on your clean-up list.
The honest limit: with Nolan Edit or Claude Code, the agent is a full Claude Code session with its own shell, files and web. Those need your answer in the chat unless you’ve set Project settings › Agent may to “Use everything without asking,” and we can’t make an agent’s provider promise anything. Approving a step approves what that step sends. Usage stats are off until you switch them on in Settings › Privacy; with them on, the app records its own window, chat and file names included and nothing masked, and sends it from your computer straight to PostHog. The app itself is covered in our Privacy Policy.