Bring Your Own Key (BYOK) lets a workspace or organization supply its own API keys for the providers SteelEngine calls on your behalf. When a key is present, SteelEngine uses it for that provider's hosted models and knowledge base jobs instead of its own key. Agent blocks, Chat, and document processing all pick it up.
BYOK keys live in Settings → BYOK, not in Secrets or Integrations. A secret is a value you reference yourself with {{KEY}}. A BYOK key is applied automatically whenever a block calls that provider.
Workspace Settings → BYOK is available on hosted and self-hosted installations. Authorized environment readers can inspect masked key metadata; changing keys requires workspace BYOK management permission. Installation-wide environment variables and organization keys under Organization Settings → Models remain available.
Supported Providers
| Provider | Used for |
|---|---|
| OpenAI | LLM calls, speech-to-text, and knowledge base embeddings |
| Anthropic | LLM calls |
| LLM calls | |
| Mistral | LLM calls and knowledge base OCR |
| Kimi, DeepSeek, Cerebras, xAI, Fireworks | LLM calls |
| ElevenLabs, Deepgram, AssemblyAI | Speech-to-text and text-to-speech |
| Firecrawl, Exa, Serper, Linkup, Perplexity, Jina AI, Parallel AI | Web search, scraping, and research |
| Google Cloud | Translate, Maps, PageSpeed, and Books APIs |
| Brandfetch | Brand assets and company information |
| TypeSafe | Judgment block and Guardrails hallucination check (Jev). BYOK only: there is no hosted key |
The BYOK page lists every provider with a short description of what the key is used for.
TypeSafe has no hosted key, so removing it breaks every deployed workflow that uses the Judgment block or the hallucination check at its next run. Judgment state and instructions, and knowledge-base excerpts used by the hallucination check, are sent to TypeSafe (US-hosted); an organization-level key enables this for every workspace in the organization.
Self-hosting note: setting TYPESAFE_API_KEY in the host or worker environment switches these tools to a platform-key path. That path is not supported or documented in the environment contract; use BYOK.
Adding a Workspace Key
- Open workspace Settings → BYOK.
- Find the provider, or search for it with Search providers....
- Click Add Key.
- Paste your API key and click Save.
SteelEngine encrypts the key and shows only a masked version afterwards. Click Update to replace a key or Delete to remove it. Adding, updating, and deleting keys requires the workspace:manage_byok permission. Members with environment read access can see which providers have a key but cannot change them. Direct API requests enforce the same permissions. Replacement forms start empty; stored keys cannot be revealed.
Use Refresh keys to reload keys created or rotated through the API while this page is open. Reopening the page also reloads the current workspace. Workspace rows show a masked value and a Workspace key badge; an inherited organization key is labeled separately. A failed read shows an access/load error, and failed saves or deletes can be retried.
Deployment Pipeline Configuration
An approved workload identity can manage workspace provider configuration through separate workspace:read_byok, workspace:set_byok, and workspace:delete_byok grants. This machine contract returns fixed masked metadata, accepts keys only as writes, and requires exact state/revision checks, confirmation and stable retry keys. Deleting a local key preserves organization keys and may select the existing organization fallback; managed release bindings can block cleanup. Existing dashboard permissions remain unchanged.
Organization Keys
An organization admin can add keys for the whole organization under Organization Settings → Models. An organization key applies to every workspace in the organization that has no key of its own for that provider. The workspace BYOK page shows such a key as Inherited with an Org key badge.
The same Models page holds the organization's model restrictions, described below.
How Keys Are Resolved
For each provider call, SteelEngine looks for a key in this order:
- The workspace key
- The organization key
- The installation/platform key or a key entered on the block, according to that provider's existing credential policy
This precedence applies to direct supported providers on self-hosted installations too, including OpenAI, Anthropic, and Google. A stored key that cannot be decrypted fails the request; it does not silently select an organization or platform key instead. Revoked or rejected keys must be replaced. Removing a workspace key enables the existing organization fallback; without it, a provider needs installation or block credentials.
Promotion-bound credentials and cloud-specific credential paths (such as Vertex, Bedrock, and Azure) keep their own configuration and precedence. An OpenAI BYOK key does not configure an Azure endpoint, deployment, region, or API variant. Existing behavior treats a BYOK database lookup failure as fallback-eligible; this is distinct from finding a stored key that cannot decrypt.
Calls made with a BYOK key are not billed to your SteelEngine usage and carry no hosted-model multiplier. You pay the provider directly at its own rates. See Costs for how hosted usage is charged.
Knowledge base processing follows the same rule. An OpenAI key is used for embeddings and a Mistral key for OCR, as described in Knowledge Base.
Scheduled Sub Agents use the same model and key resolution as Chat. A task saves its model choice, not a copy of the provider key; current keys and model restrictions are checked when it runs.
Provider Availability on Self-Hosted Installations
A configured workspace or organization key makes the existing direct provider's models available in the editor, subject to model and provider restrictions. Adding or deleting a key refreshes model availability for that workspace. The provider catalog describes implemented adapters; it does not verify credentials, network access, or every API variant.
| Credential path | Existing providers | Configuration |
|---|---|---|
| Direct model API with workspace/organization BYOK | OpenAI, Anthropic, Google Gemini (google), Mistral, Kimi (moonshot), DeepSeek, Cerebras, xAI, Fireworks | Add the vendor's direct API key in BYOK; select a model the existing adapter supports. |
| Cloud-hosted model API | Azure OpenAI, Azure Anthropic, Vertex, Bedrock | Use the existing service-specific endpoint, identity, region, deployment, and credential configuration. A direct-provider key does not enable the cloud variant. |
| Other BYOK services | Search, speech, scraping, OCR, and other rows above | A service key does not necessarily add an LLM model to the editor. Verify its actual tool or ingestion path separately. |
Before rollout, record the exact provider ID, model/deployment, API variant, credential scope, required features, and execution host. Run a representative request from both interactive and worker/scheduled execution where used. Verify workspace-key precedence and organization fallback, safe failure for revoked keys, and provider/model denials. Test streaming, tool calling, structured output, and embeddings independently when required. Confirm network/TLS access from each host and check vendor usage without recording key material.
If a provider is missing, check the current workspace, its organization fallback, the canonical provider ID, server configuration, and organization/workspace restrictions. Refresh the settings and model selector after API changes. Successful text generation does not establish tool-calling, streaming, embedding, or customer deployment qualification. New adapters require separate scope.
Model Restrictions
Below the provider list, Settings → BYOK shows every model SteelEngine offers, grouped by provider. Turn a model off to hide it from the workspace. A model turned off at the organization level shows Disabled by org and cannot be turned back on in the workspace. Click Save to apply your changes.
Disabled models disappear from model selectors in the workspace. A run that still names a disabled model is rejected. Changing workspace model restrictions requires the workspace:manage_settings permission. Organization restrictions are managed on Organization Settings → Models.
Where the Laya Decision block is available, the list also includes Laya (Laya Decision block). Turning it off hides the block from the workspace, and runs that reach an existing Laya Decision block fail at that block.