Every AI feature in Trace.Space, from the Space Agent chat to quality checks and the AI reviewer, runs on a model from an external AI provider. This article explains what goes to that provider, when, and what stays inside Trace.Space.
Your organization chooses the provider
An org admin decides on the AI Providers page which providers and models the organization may use, and which model is the default. The providers Trace.Space can talk to are OpenAI, Anthropic, Google Vertex, Bedrock, Mistral, GroqCloud, xAI and any OpenAI Compatible endpoint, including a model you host yourself.
Two arrangements are possible, and both can be active at once:
Built-in providers come with your deployment and use keys held by the platform operator.
Your own providers use an API key from your organization's own account with the provider (bring your own key). Content then goes directly to your account, under your contract with that provider. The key is encrypted at rest and never shown again.
If no provider and default model are enabled, AI features are switched off for the organization. See AI providers.
What is sent, and when
Content is sent to a provider only when someone uses an AI feature. Nothing is sent in the background on its own.
Feature | What goes to the provider |
Space Agent chat | Your messages, the items you add as context, the items and traces the agent reads while answering, and any images you attach. |
Generate with AI, Edit with AI, attribute generation | The text of the item or selection you are working on, plus the item type's AI instructions. |
Quality checks | The item text being checked and the quality rules in force. |
AI reviewer | The items in the review and the review instructions. |
Suggested traces | The items being compared. |
Image attachments in the chat can be JPEG, PNG or WebP files, usually up to 5 MB each; the exact limit depends on the model. Whether a given model accepts images depends on the provider.
The agent acts with your permissions. It can only read items you can read, and every change it proposes is shown to you for approval before it is saved.
What is kept inside Trace.Space
Chat history is stored per user, per organization. Only you can see your chats. You can rename, bookmark and delete them; the AI reviewer's reasoning is stored as read-only chats. See AI chat history.
Background work. A chat reply or an AI review keeps running on the server after you close the tab. Its state is stored in the platform's database until it finishes or you stop it. See Background processes.
Attachments you add to a chat are stored with that chat.
What a provider does with content it receives is governed by that provider's terms, or by your own agreement with it when you bring your own key.
Usage telemetry
For sign-in, sessions, secret storage and where the platform itself is hosted, see the Security overview.
Trace.Space records how AI features are used: which model was called, token counts, cost and latency. Prompt and response content is stripped before these records leave the platform. Telemetry is sent only when the deployment has a telemetry token configured, which is the case for the hosted service; self-hosted deployments without one send nothing. The telemetry endpoint is in the EU.
Connecting an external assistant (MCP)
When you connect Claude, Cursor, VS Code or another assistant through the Trace.Space MCP server, the path is different. Trace.Space serves your data to that assistant through tools; the assistant's own model, run by its vendor, does the reasoning. Your organization's AI providers are not involved, and what happens to the content is governed by the assistant vendor's terms. The assistant signs in as you and acts with your permissions. See MCP server setup and the MCP tools reference.