Providers
You can add multiple connectors per organization foranthropic, openai,
bedrock, vertex, openrouter, xai, or google. Provider identity and API
format are separate: an xAI or Google connector remains labeled as that
provider even when it uses an OpenAI-compatible API.
Choose a provider connector that your supported harness
can use. Readiness-gated versions of Claude Code, OpenCode, DeepAgents, Goose,
and Codex CLI share the same maintained model-routing contract. Each can use the
current Anthropic, OpenAI, Google, and xAI frontier routes, plus a custom
OpenAI-compatible endpoint and model ID. The selected harness version publishes
its model-readiness status and capabilities in the harness catalog.
Custom endpoints
Set an HTTPS endpoint and explicit API format when your provider exposes an OpenAI-compatible API or you operate a compatible gateway:Choosing models
Each key offers a set of models. Pick an org default - a key plus a model from it - on the dashboard. Every agent uses that default unless it overrides to a different key and model, so one org can route different agents to different keys. Custom model IDs are passed through as provider-owned opaque strings; they do not need to appear in Karta’s managed model catalog.Model parameters
Organization defaults and agent overrides support the same portable controls:
An agent can override one field while inheriting the rest from the organization.
Unsupported route or parameter combinations fail before the harness launches;
Karta does not silently drop parameters or switch providers.
Anthropic subscription
Instead of an API key, you can connect an Anthropic subscription. Use it for your own coding work running on Karta - not to serve requests on behalf of your end-users, which Anthropic’s terms do not permit and which can lead to your subscription being suspended. For any agent serving users, virtual employees, backend jobs, or customers, use an API key.Store a key
Set one from the dashboard, CLI, or management SDK:How it is used at request time
Karta resolves a connector and model once for each turn:
If a provider rejects a BYOK key, the turn surfaces a
byok_key_rejected error
422for unary requests, or anerrorevent mid-stream for streaming requests.
model_route_invalid,
model_route_unsupported, or model_parameter_unsupported errors before a
provider request is made.

