Tabnine Protected Chat is a model and governance choice inside a broader developer assistant, not a blanket promise that every Tabnine workflow uses the same model or retention path. The private-installation notes document the Protected Chat Model Version 2 becoming the default on May 23, 2024, and a later enterprise change in March 2025.
What the evidence actually establishes
The March 19, 2025 release notes state that enterprises could elect to use third-party AI models exclusively and disable the requirement to keep Tabnine Protected enabled. This is a meaningful version boundary: a description written for mid-2024 cannot be transplanted into all 2026 installations. Tabnine also documents editor plug-ins, contextual code sent for generation and differences between Tabnine-developed inference and third-party endpoints. Model selection, code indexing and agent tools are separate configuration surfaces.
How a development team should use or migrate it
An enterprise security review should record the deployment type, plug-in/engine versions, enabled model providers, RAG index location and administrator settings in one evidence sheet. Run two non-sensitive scenarios: a short fix on a file with unrelated secrets stored elsewhere, and a repository question involving two folders. Capture which snippets are submitted, whether any third-party endpoint is invoked, whether a user can change models and whether an administrator can restrict those choices. Verify via local network and audit telemetry when authorized; this article does not claim those tests were run.
A realistic engineering example
Suppose two teams share a company but use different deployment boundaries. One uses a company-operated Tabnine cluster and only a Tabnine-provided chat model. Another permits a third-party endpoint through a managed model selector. A generic “private AI assistant” statement conceals exactly the difference procurement needs to know. Track the network destination, administrator toggle, model route, repository-index location, software version and legal agreement as separate fields. Reevaluate whenever a deployment upgrades, because the release notes show that model selection policy changed over time. A valid report should distinguish claimed deletion immediately after inference from independently observed network traffic, logs and retention policies.
Security, licensing and production limits
Tabnine’s official privacy page describes a no-train/no-retain policy for its models and states that context is discarded after inference, while discussing metrics, logs and self-hosted deployments separately. That first-party policy should not be silently broadened to all third-party models, every enterprise agreement or every retention mechanism. Access rights for repository indexing, operational telemetry and incident-response exports need their own contract review. Exact costs, seats, geography and private-deployment entitlements are deliberately omitted.
Who should choose it and what to compare
Treat Protected Chat as one of several acceptable model-governance arrangements if its own inference, licensing and network boundaries match policy. For teams needing a complete on-premises design, compare it with Tabby or a separately self-hosted assistant using equivalent tasks and legal checks. A vendor policy is evidence of an asserted practice; it is not an independent audit of customer traffic.
Decision checklist
- Which engine and IDE plug-in versions are deployed?
- Can the administrator disable every external model path?
- Where are repository indexes, logs and prompts processed?
- Does the enterprise contract match the published no-retain statement?
Sources & Further Reading
- https://docs.tabnine.com/main/administering-tabnine/release-notes — official product, publisher or project material (checked October 25, 2026).
- https://docs.tabnine.com/main/welcome/readme/privacy — official product, publisher or project material (checked October 25, 2026).