Bring Your Own Models: Organization AI Providers
Connect your own model accounts, including Azure OpenAI deployments and AWS Bedrock GovCloud, so the Organization's agents use models your organization controls.
Open Settings → AI providers to connect provider accounts and choose the models your organization uses. Owners, Admins, and Security Admins manage connections. Agent Admins and Auditors can inspect sanitized settings; License Admins can inspect usage. Existing membership, SSO, access policy, and recent-authentication requirements apply.
Organizations continue using platform providers until an administrator activates a tested draft. After activation, organization AI uses only its configured routes. Provider failures do not switch to platform keys.
Connect and activate
- Add a named connection. Select a provider and enter its authentication settings. Secrets are encrypted, write-only, and omitted from settings responses and audit exports.
- Select Add model to add a model or deployment. Expand a saved model to edit its settings. Choose its connection and credential version, input/output token limits, and supported capabilities. Reasoning requires tool calling. Browser AI requires tool calling and structured output; enable vision for tasks that need screenshots. AI file interpretation requires vision.
- Assign primary models and ordered alternatives to each intelligence level. The first model can fill every level initially. Existing intelligence entitlements still apply. Summaries and lightweight tasks reuse the text route unless explicitly assigned under Supporting text tasks.
- Expand AI features. Enable and assign models for browser AI, file interpretation, embeddings, images and avatars as needed; leave unused features disabled. The activation section lists unavailable features. Embedding duplicate detection retains deterministic matching when embeddings are disabled.
- Save the draft. Run each assigned route's synthetic test. Tests use sample inputs and may incur provider charges, including image generation tests.
- Activate after all tests pass. Activation publishes the exact tested configuration atomically. If another administrator changes settings, reload and review the current draft.
For models with extended thinking, enable the reasoning capability and choose an effort level. The output limit must leave room for both thinking and the answer. Tests use the configured reasoning budget; if it cannot fit, increase Maximum output tokens or reduce Reasoning effort before testing again. Neither testing nor production calls increase the organization's output limit.
Providers
| Provider | Connection and model settings |
|---|---|
| OpenAI, Anthropic, Gemini, OpenRouter | Organization API key and provider model identifier |
| Azure OpenAI v1 | Resource URL ending in /openai/v1/, API key, deployment name as the model |
| Custom OpenAI-compatible | HTTPS base URL, API key or explicit unauthenticated mode, model identifiers |
| Bedrock-compatible | Supported regional compatible endpoint, manually supplied Bedrock API key, supported models. For Claude on the Bedrock Mantle endpoint, use base URL https://bedrock-mantle.<region>.api.aws/anthropic (Anthropic Messages API); Mantle does not serve Claude through /v1 chat completions. |
| Native Bedrock | Partition, region, credential source, optional role, model ID or inference-profile ID/ARN |
Azure uses the OpenAI v1 API with API-key authentication. Entra ID and legacy Azure API versions are outside this release. See Microsoft's API lifecycle documentation.
Native Bedrock uses Converse/ConverseStream for text, tools, browser AI, and inline vision. Amazon Titan text embeddings use InvokeModel. Model and feature availability varies by region; a successful capability test is required. Native Bedrock media generation is unavailable in this release; assign another organization connection or disable those features. Other provider features are enabled only after their assigned tests pass. See AWS Converse documentation.
Bedrock in AWS GovCloud
Select AWS GovCloud (US) and us-gov-west-1 or us-gov-east-1. Use a model available to the
account in that region. Inference-profile ARNs must match the connection's partition and region.
Choose one authentication method:
- Organization access credentials: provide a GovCloud access key and secret. Temporary credentials also require a session token and expiration with timezone. A target role is optional. Expired source credentials require replacement.
- Operator-authorized workload source: your deployment operator explicitly authorizes an AWS credential source for your organization and partition. Select that source and a target role. The application supplies the connection's unique external ID when assuming the role.
Save the connection to obtain its external ID. Add that value to the target role's trust policy alongside the authorized source principal. Allow the necessary Bedrock model invocations. Source identity, partition, and role assumption are checked during tests and credential refresh. Commercial AWS identities cannot directly assume GovCloud roles. See AWS cross-account role guidance.
The connection can be used by GovCloud-hosted workers or workers elsewhere that have authorized GovCloud credentials and network connectivity. Connecting does not move workers, storage, logs, or customer data into GovCloud.
Private endpoints and certificates
Private destinations must be reachable from the workers and the egress proxy. Ask the deployment operator to approve the connection's CIDRs and configure a platform-network denylist. Both policies must permit a destination; deny rules take precedence. DNS is checked again when each proxy connection is established, and the proxy connects to the validated address.
Optional custom PEM CA certificates replace system trust for the connection. TLS and hostname verification cannot be disabled. Redirects are rejected. Native AWS authentication requests use regional STS with the same guarded transport. FIPS selection uses AWS regional endpoint metadata; GovCloud's regional STS endpoints already serve FIPS requests. PrivateLink provisioning and routing remain deployment responsibilities. See AWS PrivateLink documentation.
Rotation, repair, and usage
Stage a new credential version, select it on affected models, save the draft, test, then activate. For connections with supplied credentials, adding, replacing, or removing custom CA certificates requires re-entering the API key or AWS access credentials, including a retained session token. Changing an endpoint also requires re-entering supplied credentials. Existing requests keep the original revision. Emergency revocation immediately rejects new credential use.
Missing, expired, or revoked credentials produce an organization configuration error. After a successful repair and activation, pending agent work blocked by provider configuration is queued again. Return to platform providers is an explicit administrator action.
Usage shows the model, purpose, outcome, token totals, and available provider cost estimates for the last 30 days. Unknown pricing, including Azure deployment aliases and custom servers, appears as unavailable. This view shares the inference ledger with Usage and agent profiles. Customer-rate estimates name their revision; default public estimates may be stale and exclude negotiated terms, commitments and discounts. Estimates remain separate from provider-reported amounts and invoices.