AI Privacy
This page explains the privacy and trust boundaries for AI features in Zed, including Zed Agent, Edit Prediction, Inline Assistant, and Git commit generation.
Zed does not retain your prompts or code context by default. For Zed-hosted models, Zed has no-training commitments from model providers, and provider agreements require zero data retention for inference requests except for provider-designated models with safety retention, such as Anthropic's Mythos-class models. Zed only retains AI data when you explicitly share feedback or opt in to training data collection.
AI Request Paths
| Path | Who handles model requests | What to know | Details |
|---|---|---|---|
| Zed-hosted models | Zed routes requests to hosted model providers | Provider agreements prohibit training on your prompts or code context and require zero data retention for inference requests, except for provider-designated models with safety retention, such as Anthropic's Mythos-class models. | Zed-hosted model commitments |
| Provider API keys | The configured provider | The provider handles requests under its own terms. Provider keys saved through Zed are stored in the system keychain, not in settings.json. | Use API Access |
| Existing subscriptions | The subscription provider | The provider handles requests under the subscription terms. | Use an Existing Subscription |
| Gateways | The configured gateway and upstream providers | The gateway and upstream providers handle requests under their own terms. | Use a Gateway |
| Local models | The local server or self-hosted endpoint | The local server handles requests according to how you configured that server. | Use a Local Model |
| External Agents | The External Agent and its configured providers | The External Agent handles model requests under its own terms. Tool and MCP behavior depends on agent and ACP configuration. | External Agents |
| Terminal Threads | The CLI or TUI running in the terminal | The CLI or TUI owns its auth, model routing, tools, instructions, MCP configuration, and data handling. | Terminal Threads |
| Edit Prediction | The selected edit prediction provider | Each keystroke can send local editing context to the selected provider. Zeta requests are processed transiently unless training data collection is enabled; third-party providers follow their own terms. | Edit Prediction, Feedback and Training Data |
| Agent tools, MCP, and integrations | Zed, configured MCP servers, and external systems | Tools can read, edit, search, run commands, fetch URLs, or call external systems depending on profile, MCP server, and tool permission settings. | Agent Profiles, Tool Permissions, MCP |
| Project trust and instructions | Zed and the trusted worktree | Project-local instructions and skills are loaded from trusted worktrees. External Agents and Terminal Threads may read their own instruction files. | Worktree Trust, Skills, Instructions |
Zed-Hosted Model Commitments
For Zed-hosted models, Zed has commitments from model providers that prohibit training on your prompts or code context and require zero data retention for inference requests, except for provider-designated models with safety retention, such as Anthropic's Mythos-class models. The public provider documents linked below describe provider programs or default API terms; Zed-hosted model requests are governed by Zed's provider agreements.
| Provider | No training reference | Zero-data-retention reference |
|---|---|---|
| Anthropic | Yes | Yes, except covered models |
| Yes | Yes, see Service Terms sections 18 and 20(h) | |
| OpenAI | Yes | Yes |
Provider Safety Retention for Designated Models
Some providers require limited data retention for specific models as a condition of offering them, on every platform where those models are available. Anthropic retains prompts and outputs for models it designates as covered models (its Mythos-class models, such as Claude Fable 5) for 30 days for trust and safety purposes. Zed cannot opt out of this retention; it applies wherever these models are served. See Anthropic's data retention practices for Mythos-class models.
For these models:
- The no-training commitment still applies. Retained data is used for safety review, not model training.
- Zed does not retain your prompts or outputs. Retention happens at the provider, under the provider's documented access controls and deletion timelines.
- All other Zed-hosted models keep zero-data-retention handling.
If you don't want provider-side retention, use a model that the provider has not designated for safety retention. Switching to your own API key or subscription does not avoid this retention for covered models, because providers apply it on every platform where those models are offered.
AI Data Retained by Zed
Zed may retain AI data only when you explicitly share it or opt in:
- Response ratings and feedback can send a conversation thread to Zed for review and improvement.
- Edit Prediction training data is collected only when you opt in, the project is open source, and the file is not excluded.
See Feedback and Training Data for the full list of what can be stored in each opt-in case.
Controls and Related Privacy Docs
- Telemetry: What telemetry Zed collects and how to control it.
- Privacy for Business: How Zed Business enforces privacy settings across an organization.
- Admin Controls: How owners and admins control Zed-hosted models, Edit Prediction, and feedback sharing.
- AI Quick Start: How to turn AI off.
- Privacy Policy: Zed's privacy policy.
- Subprocessors: Zed's subprocessors.
- Terms of Service: Zed's terms.