List the model sources this tenant can connect (GET /v1/model-source-profiles). Every source runs on your provider key/account by construction: each entry carries its protocol/auth profile versions and endpoint policy. No credential material is involved.

Start a hosted browser credential handoff for a source (POST /v1/source-credential-handoffs). Returns a continue_url for a HUMAN to complete in a browser against the credential broker -- this tool never accepts or returns a raw key, service-account JSON, cloud secret, or token. Poll with solver_poll_source_handoff until state=completed, then pass the handoff_intent_id to solver_create_source_connection.

Poll a credential handoff intent until the human completes it in the browser (GET /v1/source-credential-handoffs/{handoff_intent_id}). Answers the intent state only.

Bind a source connection for your provider key/account from a COMPLETED handoff intent (POST /v1/source-connections). Takes only the opaque handoff_intent_id -- never credential material. A fresh connection is disabled until solver_test_source_connection passes; deployments then come from solver_sync_source_deployments.

Run the live credential test on a source connection (POST /v1/source-connections/{id}/test). A passing test activates the connection; the wire reports test_state/test_error.

List this tenant's source connections with their test/rotation state (GET /v1/source-connections). auth_binding_ref is an opaque broker handle, never a secret.

Discover and persist the connection-authorized, certification-backed deployments (POST /v1/source-connections/{id}/deployments/sync). Requires an ACTIVE (tested) connection; answers the synced deployment set. Discovery is not certification: unevidenced identities are never persisted.

List the tenant's certification-backed model deployments (GET /v1/model-deployments). Enable one as an arm with solver_enable_model_arm.

Enable a certification-backed model deployment as an arm (POST /v1/arms with kind=model and model_deployment_id). Use the catalog entry's secretless arm_registration_template values; claims wider than the certified template are rejected.

Register a model/agent/skill arm (POST /v1/arms). Returns the created arm, including its arm_id, status, and (for skill kind) artifact_hash.

List the tenant's registered arms (GET /v1/arms). Returns { arms, next_cursor }.

Submit a task for governed execution (POST /v1/executions). Requires a unique request ID via idempotency_key (sent as the Idempotency-Key header) -- an MCP client cannot construct a valid call without it.

Poll an execution's lifecycle state (GET /v1/executions/{id}).

Cancel a running execution (POST /v1/executions/{id}/cancel).

Fetch the retention-bound result of a COMPLETED execution (GET /v1/executions/{id}/result): the final text plus the winner attempt's model provenance (requested vs resolved identity) when one exists.

Fetch the receipt for an execution -- route rationale, per-slice evals, cost (GET /v1/receipts/{execution_id}). Content-free: never returns task output.

List pending repair proposals, or inspect one when proposal_id is set (GET /v1/proposals[/{id}]).

Approve or reject a repair proposal (POST /v1/proposals/{id}/approve|reject); the decision parameter selects which.