Agents
Agent overview
How agents access the platform — principals, scoped credentials, and permissions.
Agents are first-class users of the platform: every console operation is available programmatically under the same rules — the same ownership checks, the same immutable revision service, the same validation.
Identity and scopes
Programmatic callers act as a principal derived from your account, limited by a scoped credential:
- Scopes distinguish read, write, render, share, publish, and delete. Grant the minimum set an integration needs.
- Artifact restriction optionally pins a credential to specific artifacts.
- Ownership always applies on top of scopes: a credential can never touch another user's resources, and one agent cannot mint another credential.
See API keys and scopes for the full credential model.
Transport options
| Transport | Use case | Notes |
|---|---|---|
| MCP over stdio | Local agent hosts | Launches a local server process; credentials via environment |
| MCP over Streamable HTTP | Remote agents and services | OAuth with PKCE or a manually provisioned scoped credential |
| REST + SDK | Direct integrations | Versioned endpoints, typed contracts |
Both MCP transports use the same validated commands and structured outputs.
Safety rules the platform enforces
- Idempotency — retried operations with the same key and identical body return the original result; the same key with a different body fails.
- Revisions — concurrent writes conflict explicitly; nothing is silently overwritten.
- Structured errors — failures carry a code, message, and validation issues; MP4 binaries are never embedded in protocol responses.
- Bounded retries — SDK clients retry only transient conditions with a stable key.