attach_agent_mcp_server means one agent can extend another's reach without a human editing a config file. That is genuinely powerful and worth thinking about before granting: an agent with this tool can grant capabilities, which is a different kind of permission from calling one. cancel_session is the practical safety valve, and App Activity logging is what makes any of it auditable afterwards.
Gumloop's first-party endpoint — 40 tools, documented and enumerated exactly. What it manages is other agents: their configuration, their skills, and which MCP connectors they in turn can reach.
- list_agents, get_agent, create_agent, update_agent — agents with a name, model and configuration
- attach_agent_skills and detach_agent_skills — compose an agent's capabilities from stored skills
- attach_agent_mcp_server and detach_agent_mcp_server — give an agent its own MCP connectors, or take them away
- start_agent and get_agent_status — asynchronous interaction, polled
- create_agent_session, get_session, send_session_message, cancel_session — a conversation with an agent, including cancellation
- list_skills, create_skill, update_skill — skills built from files in your workspace
- list_models — the model groups available
OAuth 2.0 with PKCE and dynamic client registration, scope gumstack, or a Gumloop API key. Enterprise adds incognito mode, custom retention and model access control, with all tool use recorded in App Activity.
Paste a URL, then authorize — add the endpoint to your client, then approve the OAuth consent screen