The question most teams ask before approving an AI coding tool is whether the provider keeps or trains on their code, and the answer almost always depends on the plan and the sign-in route rather than the product. This tracker sets out, for eight AI coding tools, what each provider's own documentation says about model training on individual and business plans, how to opt out, how long data is kept, and whether zero data retention (ZDR) is offered. Every cell comes from the provider's privacy, data use or documentation page, and where a provider's page does not state something, the table says so. It is written for security, privacy and procurement teams and for the engineering leads who answer their questionnaires. Tools are listed alphabetically, and the order is not a ranking. Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Policies change, and contracts (such as a data processing agreement) can override public pages, so confirm against each provider's current terms.
Key takeaways
- Business plans are opted out of training by default in every tool listed. Each of the eight providers documents that business, enterprise or commercial customer content is not used for model training by default, or is covered by a business agreement.
- Individual and free plans usually carry a training setting. Anthropic, Cognition, GitHub, Kiro, OpenAI and AWS (Amazon Q Developer Free tier) each describe individual-plan content that may be used for training or service improvement unless the user opts out or, for Anthropic, unless the setting is off.
- Retention periods are published unevenly. Anthropic and Kiro publish day counts for coding tool data; others point to account-level settings, abuse monitoring windows or the length of the customer relationship.
- ZDR is a separate, approved arrangement. Where offered, ZDR is enabled per organisation after an eligibility check, or is tied to a privacy setting, not switched on by default.
- The sign-in route decides the policy. A developer using a personal account on a work repository is under the individual terms, so enforcing the corporate sign-in is the control that makes the business policy apply.
Why one product can have several data policies
AI coding tools sit on top of a subscription or a cloud account, and the data terms attach to that account, not to the editor or CLI. The same binary can therefore run under consumer terms on one laptop and commercial terms on the next. Three routes recur across providers:
- Individual subscription: free and paid personal plans. Training use is typically governed by a user setting.
- Business or enterprise subscription: team and enterprise plans under commercial terms. Training is off by default, and retention controls, audit logs and ZDR sit here.
- API or cloud account: the tool calls a model API with an API key or through a cloud provider. The API organisation's data settings apply.
This is why the tables below split every field by plan type. For the same question about model providers rather than coding tools, see the LLM Data Retention and Training Policies Compared page.
Does it train on my code? Training use by plan
As of 15 November 2026, from each provider's own pages (listed in the source column of the reference table further down).
| Tool (provider) | Individual or free plans | Business or enterprise plans |
|---|---|---|
| Amazon Q Developer (AWS) | Free tier: AWS may use questions, responses and generated code for service improvement, including model training | Pro: content not used for service improvement |
| Claude Code (Anthropic) | Free, Pro, Max: user choice; data used to train new models when the setting is on | Team, Enterprise, API, third-party platforms, Claude Gov: not used to train generative models under commercial terms unless the customer opts in |
| Cursor (Anysphere) | Privacy Mode on: not used for training by Cursor. Privacy Mode off: Cursor may use and store codebase data, prompts, editor actions and code snippets to train its models | Same Privacy Mode rules; Privacy Mode can be enforced organisation-wide on Teams and Enterprise |
| Devin (Cognition) | By default Cognition may use data for model training; paid plans can opt out | Teams: admin can opt out. Enterprise: no training without express prior written consent |
| Gemini Code Assist (Google) | Individual tier stopped serving requests on 18 June 2026 (see Limits below) | Gemini for Google Cloud: Google states Gemini does not use prompts or responses as data to train its models |
| GitHub Copilot (GitHub) | Free, Pro, Pro+: GitHub may use prompts, suggestions and code snippets to train and improve AI models; users can opt out | Business, Enterprise: data uses are set out in GitHub's Data Protection Agreement, per the Copilot FAQ |
| Kiro (AWS) | Free tier and individual subscribers: content may be used for service improvement, which can include model training | Enterprise and IAM Identity Center users: content not used for service improvement |
| OpenAI Codex (OpenAI) | Services for individuals: OpenAI may use content to train models; opt out available | Business: not used for training by default; business users and API opted out unless they opt in |
How do I opt out? Settings by tool
As of 15 November 2026. Where the provider names the setting, the name is given exactly.
| Tool | Opt-out or control | Who can change it |
|---|---|---|
| Amazon Q Developer | IDE data-sharing settings; for console, mobile app, websites and chat apps, an AI services opt-out policy in AWS Organizations | User in the IDE; organisation administrator for the opt-out policy |
| Claude Code | Privacy settings at claude.ai/settings/data-privacy-controls (consumer accounts) | User, at any time |
| Cursor | Privacy Mode | User; enforced organisation-wide by admins on Teams and Enterprise |
| Devin | Settings > General > Data controls & privacy > "Disable training" | Paid plan users; on Teams, administrators only |
| Gemini Code Assist | Trusted Tester Program is optional data sharing for product improvement, not model training | Organisation |
| GitHub Copilot | Copilot settings: turn off use of data for AI model training | Individual subscriber, at any time |
| Kiro | Settings > User > Application > Telemetry and Content > uncheck "Content Collection for Service Improvement" | Free and individual users; enterprise users are opted out automatically and admins control telemetry |
| OpenAI Codex | OpenAI privacy portal "do not train on my content", or ChatGPT Data Controls; applies to new conversations | Individual user |
How long is data kept? Retention and ZDR
| Tool | Stated retention | Zero data retention |
|---|---|---|
| Amazon Q Developer | Not stated by AWS on the service improvement page | Not stated by AWS on the service improvement page |
| Claude Code | Consumer: 5 years if training is on, 30 days if off. Commercial: 30 days standard. Local transcripts: 30 days by default under ~/.claude/projects/. /feedback transcripts: 5 years | Qualified Claude for Enterprise accounts, enabled per organisation by the account team; not part of the standard Enterprise plan |
| Cursor | File contents cached temporarily and encrypted with client-generated keys; never permanently stored. Day counts not stated by Cursor on the data use page | With Privacy Mode on, Cursor states it has ZDR agreements with all model providers; abuse-flagged data may be stored for investigation |
| Devin | For the duration of the customer relationship unless specified otherwise; feedback and interaction data as Cognition determines | On paid plans, opting out of training also enables ZDR with Cognition's model providers |
| Gemini Code Assist | Not stated by Google on the data governance page | Not stated by Google on the data governance page |
| GitHub Copilot | Not stated by GitHub on the Copilot features page; refers to the GitHub Privacy Statement | Not stated by GitHub on the Copilot features page |
| Kiro | Free tier: inputs up to 60 days for abuse detection. Model-specific: classifier-flagged OpenAI GPT traffic up to 30 days; Claude Fable 5.1 traffic up to 30 days. Enterprise prompt logging to a customer S3 bucket | Not stated by AWS on the Kiro data protection page |
| OpenAI Codex | ChatGPT sign-in: workspace retention and residency settings. API key: API organisation settings; abuse monitoring logs up to 30 days by default | API: ZDR and Modified Abuse Monitoring with prior approval from OpenAI |
Retention of the local copy matters as much as the provider copy. Claude Code keeps plaintext transcripts on the developer's machine, and Codex keeps access tokens in ~/.codex/auth.json unless the OS keychain is selected. See the AI Coding Tools and Secrets Sprawl guide for what ends up in those files.
Where is the data stored and processed?
- Claude Code: prompts and outputs travel over TLS 1.2 or later. Encryption at rest depends on the provider: AES-256 on the Anthropic API; AWS-managed keys with optional KMS customer-managed keys on Amazon Bedrock; Google-managed keys with optional CMEK on Google Cloud's Agent Platform. Cloud sessions clone the repository into an isolated VM, and GitHub credentials stay encrypted on Anthropic's servers and are attached by a proxy.
- Cursor: requests, including those using the customer's own API key, pass through Cursor's backend, where final prompt building happens.
- Gemini Code Assist: prompts are handled under the Google Cloud terms of service and Cloud Data Processing Addendum; code customisation stores the private codebase to deliver the service.
- Kiro: content for Free tier and individual subscribers is stored in US East (N. Virginia); enterprise content in the region where the profile is configured. Inference may run in other regions without changing the storage location.
- OpenAI Codex: local tasks run on the developer's machine; cloud tasks run in OpenAI-hosted environments and require Sign in with ChatGPT. Enterprise and Edu plans include data retention and residency controls.
- Devin: for Enterprise VPC deployments, Cognition states that all customer data remains within the customer's tenant.
Storage location details for Amazon Q Developer and GitHub Copilot are not stated on the pages listed in this tracker.
Limits, lifecycle changes and scope notes
- Amazon Q Developer end of support: AWS blocked new Free tier accounts and new subscriptions from 15 May 2026; IDE plugins and paid subscriptions reach end of support on 30 April 2027, and AWS points users to Kiro. See the Amazon Q Developer End of Support guide.
- Gemini Code Assist for individuals: Google states that the IDE extensions and Gemini CLI stopped serving requests for the individual, Google AI Pro and Google AI Ultra tiers on 18 June 2026; Standard and Enterprise follow a separate schedule.
- Feedback channels: thumbs ratings, feedback commands and bug reports often follow different retention than normal use (Claude Code
/feedbacktranscripts are kept for 5 years). Disable them by policy where needed (DISABLE_FEEDBACK_COMMAND=1for Claude Code). - Model provider retention inside a tool: a coding tool that routes to third-party models can add the model provider's own retention, as Kiro's model-specific rows show. Check both layers.
- Scope: this tracker covers tools whose official privacy or data pages could be read on the check date. Other tools are profiled in the AI Coding Tools Hub.
How to use this tracker in a review
- Inventory the sign-in routes: for each tool in use, record plan, account type (personal, business, API or cloud) and who owns the account. The Unsanctioned AI Coding Assistants guide covers finding tools nobody approved.
- Map each route to a row: read the training, opt-out and retention cells for that plan type, then open the source page and confirm the wording on the day of review.
- Enforce the business route: use the tool's sign-in restriction (for example Codex
forced_chatgpt_workspace_id, Cursor org-wide Privacy Mode) so individual terms cannot apply to company code. - Decide on ZDR: where regulated code or data is involved, request ZDR through the documented route and record the approval and its scope.
- Close the local layer: set local transcript retention, credential storage and feedback opt-outs.
- Record and recheck: keep the date checked and the URL for each answer, and recheck at contract renewal.
The AI Coding Assistant Buyer's Guide turns these answers into procurement questions, and the AI Coding Tools Enterprise Controls Compared page covers SSO, audit and policy controls.
Measuring whether the policy holds
- Corporate sign-in coverage: share of active users of each tool signed in through the business account.
- Training opt-out coverage: share of any permitted individual accounts with the training setting off, confirmed by attestation or admin view.
- ZDR status: tools with an approved ZDR arrangement versus tools that need one under your data classification.
- Local retention compliance: share of managed devices with the configured transcript retention and keychain credential storage.
- Review currency: days since each row was last checked against the provider page.
Reference: source pages by tool
| Tool | Official page used | Checked |
|---|---|---|
| Amazon Q Developer | AWS: Amazon Q Developer service improvement; AWS DevOps blog end of support announcement | 15 November 2026 |
| Claude Code | Anthropic: Claude Code data usage | 15 November 2026 |
| Cursor | Cursor: Data use | 15 November 2026 |
| Devin | Cognition: Devin admin security; Devin enterprise security overview | 15 November 2026 |
| Gemini Code Assist | Google Cloud: Gemini data governance; Gemini Code Assist individuals notice | 15 November 2026 |
| GitHub Copilot | GitHub: Copilot features page and FAQ | 15 November 2026 |
| Kiro | Kiro: Data protection | 15 November 2026 |
| OpenAI Codex | OpenAI: Codex authentication; Codex pricing; how your data is used to improve model performance; API your data | 15 November 2026 |
Practitioner checklist
- List every AI coding tool in use with its plan, sign-in route and account owner.
- For each, record the training default, opt-out setting and retention period from the provider page, with the URL and date.
- Enforce business sign-in so consumer terms never cover company repositories.
- Turn on organisation-wide privacy controls where offered (for example Cursor Privacy Mode enforcement, Devin "Disable training" set by an admin).
- Request and record ZDR where your data classification requires it.
- Disable feedback and transcript-sharing commands by policy where code must not leave the estate.
- Set local transcript and credential retention on managed devices.
- Plan migrations for tools with announced end of support, such as Amazon Q Developer IDE plugins on 30 April 2027.
Standards and references
- AWS: Amazon Q Developer service improvement
- Anthropic: Claude Code data usage
- Cursor: Data use
- Cognition: Devin security
- Google Cloud: How Gemini for Google Cloud uses your data
- GitHub: Copilot features and FAQ
- Kiro: Data protection
- OpenAI: How your data is used to improve model performance
Related NHI Mgmt Group resources: Claude Code vs OpenAI Codex, AI Coding Tools Pricing Tracker, AI Coding Tools Hub, Vibe Coding Tools Compared, AI Agent Platforms Compared.