ScreenshotCrew library — Agents · Models · Skills · Tools · Environments.
Crew is the workspace-scoped catalog of reusable building blocks. It has five sections that mirror the workspace navigation:
- Agents — saved agent definitions (system prompt + model endpoint + named environment + skills + MCP servers).
- Models — model endpoints. A provider, an API key, and an optional base URL + model name. Agents and the Inference block both pick from this list.
- Skills — reusable prompt bundles in
SKILL.mdformat. Attach them to agents to extend behavior. - Tools — Custom Tools (user-defined JSON-schema functions) and MCP Servers attached to agents.
- Environments — named environments holding variables and an optional setup script that the sandbox runs before an agent starts.
Workflows and the Agents block consume these directly: the Agents block takes an agentId and the agent itself references a model endpoint, a named environment, a list of skills, and a list of MCP servers.
Crew Agent vs. Agents block vs. Inference block
These three terms are not the same. The full disambiguation table lives on the Inference block page.
- A Crew Agent is the saved definition documented here, in
crew/agents. - The Agents block is the workflow block that runs one inside a Sandbox.
- The Inference block is the atomic single-shot LLM call. It does not load a Crew Agent.
Where things live
| Surface | Code | Notes |
|---|---|---|
| Crew app | app/workspace/[workspaceId]/w/crew/ | Five sub-routes match the docs. |
| Database | packages/db/schema.ts | Tables: agent, agent_skill, agent_mcp_server, skill, skill_file, workspace_model_endpoints, workspace_named_environment, mcp_servers, crew_agent_otp. |
| Agent dispatch | executor/handlers/agents/agents-handler.ts + lib/crew/ | Mints OTP tokens, dispatches to the sandbox. |
Everything in Crew is workspace-scoped. Agents in workspace A cannot see models, skills, environments, or MCP servers in workspace B.