Braintrust integration

Last updated date:

How Seedling connects to Braintrust, which read-only information it uses, and how access and deletion are controlled.

1. Overview

The Braintrust integration reads authorized evaluation assets, observability records, and access context from one organization through an API key or service token.

Seedling preserves project and ACL provenance, removes secrets from staged content, and exposes no Braintrust write operation.

2. Connection and scope

Seedling verifies the provider account before storing the connection. A personal connection remains actor-owned; a workspace connection is administered for the selected Seedling workspace. Credentials are never inferred from an email address or silently promoted between scopes.

Braintrust API key or service token
Prefer a service token assigned to a read-only permission group. Use the organization data-plane URL: US organizations normally use api.braintrust.dev and EU organizations use api-eu.braintrust.dev.
Connection ownership
Seedling supports actor-owned personal access and administrator-managed workspace access for this connector. The selected scope and exact connection remain fixed through synchronization, rotation, disconnect, and purge.

3. Braintrust data Seedling accesses

Seedling uses synchronized information for source-backed search, operating context, verification, and customer-approved workflows. Availability depends on the connected account, selected resources, provider plan, and native permissions.

Evaluation assets
Projects, experiments, evaluations, datasets, prompts, and functions visible to the authorized organization.
Observability
Logs, scores, feedback, timing, and supported trace and project relationships.
Access context
Organization users, groups, roles, ACLs, and project relationships returned by the token.

4. Read-only permission contract

Assign the token to a permission group that can read only the released Braintrust population and its access context.

  • Organization and project read — enumerate the intended organization and projects.
  • Project-object read — experiments, evaluations, datasets, prompts, functions, logs, scores, and feedback.
  • Read ACLs — stage access-control records for conservative authorization diagnostics.
  • Read users, groups, and roles — attribution and ACL principal context.
  • No project, experiment, dataset, prompt, score, feedback, ACL, user, or role mutation is requested.

5. Security and access

  • Provider credentials and other connection secrets are encrypted in the workspace tenant data plane, never returned to the browser, and excluded from logs and queue payloads.
  • Every source, synchronization, secret lookup, audience contribution, disconnect, and purge remains bound to the exact connection that created it.
  • The provider does not expose enough complete object-level audience evidence for Seedling to reproduce native permissions safely. Workspace content therefore uses the connection's restricted-by-default workspace-or-no-human-access policy; administrator and steward status grant no raw-content bypass.
  • For a personal connection, raw content remains visible only to the connecting member in human-facing Seedling surfaces. Personal content may contribute to trusted system exploration under workspace policy, but that authority never transfers to another member.
  • Generated output cannot bypass Seedling's human review, verification, and approval gates.

6. Retention and deletion

Connected data
Synchronized Braintrust records remain while the connection is active and are removed through a durable integration purge after disconnect.
Credentials
Stored Braintrust credentials or connection secrets and connection-owned synchronization state are deleted from active Seedling systems as part of the purge.
Backups
Deleted information may remain in encrypted, access-restricted backups until expiry, ordinarily within 35 days after active-system deletion.