Providers and AI systems
Provider types
| Type | Typical use |
|---|---|
openai_compatible | OpenAI and compatible chat-completions endpoints, including compatible Gemini endpoints |
anthropic_compatible | Claude Messages-compatible endpoints |
perplexity_compatible | Search-grounded answers and citation monitoring |
custom_agent | Customer or partner agent endpoint with a compatible contract |
Platform profiles are available for OpenAI, Anthropic Claude, Perplexity and Google Gemini. A workspace can also create a custom profile. This list does not restrict AI component modeling: engines, models, custom agents, MCP servers/tools, RAG sources, routers and workflows can be represented in the component registry.
Workspace provider configuration
| Field | Meaning |
|---|---|
| Name and type | Workspace display name and runtime adapter |
| Endpoint URL | HTTPS provider endpoint |
| Model | Default model used when an operation does not override it |
| Credential | Encrypted API key or environment/vault-style secret reference |
| Capabilities | Text, vision, JSON mode and tool support |
| Web search | Optional provider-native search/grounding for an eligible direct provider profile |
| Usage pricing | Input/output price per million tokens, web-search price per request and currency |
| Configuration | Provider-specific JSON such as API version or request options |
| Status | Draft, active, degraded or disabled |
| Default | Provider used by workflows that do not make an explicit selection |
Only active workspace-owned providers can execute. Selecting a new default clears the previous default. If the default is deleted, the oldest remaining active provider becomes the replacement.
Workspace pricing is operational configuration, not a global list-price catalog. Use the rates from the workspace owner's provider contract. A provider-reported total cost takes precedence over the configured estimate; otherwise input, output, and search rates are applied. Missing prices remain explicitly unpriced.
Provider-native web search
Provider-native search is available only for direct OpenAI, Anthropic Claude, Perplexity, and Google Gemini profiles. Perplexity search is always part of the supported direct profile; the other eligible profiles expose an explicit setting. Generic OpenAI-compatible and custom endpoints cannot enable the feature.
Search-enabled requests use the provider's native search or grounding contract and preserve structured citations. Search can carry additional provider charges. An API observation remains distinct from a consumer AI product surface such as ChatGPT, Claude.ai, Gemini, or Google AI Overview.
Runtime selection rules
- Default platform workflows use the workspace default provider.
- Operations that expose provider selection may choose one or multiple providers.
- Multi-provider monitoring stores independent answer snapshots; it does not merge raw responses.
- Provider credentials remain server-side and are never returned to the browser.
- Private-network and insecure HTTP endpoints are rejected unless deployment policy explicitly enables them.
Every direct provider call also produces a workspace-scoped usage event with status, latency, normalized tokens, provider-reported searches, and cost provenance. See Provider Usage and Cost Management.
Components and evaluations
Register components with version, dependency, capability and configuration metadata. Evaluations can compare engine/model/custom-agent components and tool/router/workflow components. Use dependency relationships to explain how a provider, model, RAG source, MCP tool and workflow jointly affect an answer.
Each evaluation records component and dependency versions. Changing a model, prompt, MCP tool or RAG source creates a new comparison context rather than silently changing historical results.