Skip to main content
Run this on your computer:
It adds the Opengeni MCP server (https://app.opengeni.ai/v1/mcp) to the coding agents it finds (Claude Code, Codex, Cursor, and VS Code with Copilot) and installs a short opengeni skill for each, so the agent knows when to hand work to Opengeni. Running it again changes nothing that’s already set up. Then sign in once in each agent, as the table says. Sign-in opens in your browser, where you pick the organization, what the agent may do, and which workspaces it reaches. The command never sees a password or key. Options: --client claude,codex,cursor,vscode (or all) sets up those even when they aren’t detected, --no-skill skips the skill, --no-sign-in skips starting sign-in, and --base-url <url> points at a self-hosted Opengeni.

Set up by hand

Without Node.js, add the server yourself, then sign in each agent as in the table above.
  • Claude Code: claude mcp add --transport http --scope user opengeni https://app.opengeni.ai/v1/mcp
  • Codex: add this to ~/.codex/config.toml, then run codex mcp login opengeni:
  • Cursor: in ~/.cursor/mcp.json, add "opengeni": { "url": "https://app.opengeni.ai/v1/mcp" } under mcpServers.
  • VS Code (Copilot): in your user mcp.json (MCP: Open User Configuration), add "opengeni": { "type": "http", "url": "https://app.opengeni.ai/v1/mcp" } under servers.
The Opengeni plugin also includes the skill. Self-hosted, use your own origin instead of https://app.opengeni.ai.

What your agent can do

  • Hand off work: delegate starts an Opengeni chat with a brief, a repository and branch, and optionally an agent or sandbox environment. wait_for follows it (up to 50 seconds per call) until it finishes with a final answer; session_send sends follow-ups. Stopping a wait in your agent doesn’t stop the chat.
  • Apps: apps_list, app_describe, app_query and app_write read and change your apps’ data. Agents are viewers by default: they read and write what each app lets a viewer. Set App data to Manager for an agent under Connected agents to let it manage collections too. Agents only reach apps that are members-only or unpublished.
  • Connected tools: tools_search and tool_call. A tool that needs approval returns a link; approve it in Opengeni and the agent’s retry runs it once.
  • Skills and Knowledge: skills_sync uploads your local Skills (to your own Skills by default, following your Agent learning setting), syncs every Skill folder of a public GitHub repository, or pulls the workspace’s; knowledge_search, knowledge_get and knowledge_save.
  • Everything else: opengeni_actions_search, opengeni_action_describe and opengeni_action_call reach every Opengeni API action.
With several workspaces, the agent asks once and calls workspace_use; later calls use that workspace. Retrying a call with the same input doesn’t start the same work twice.

Self-hosted

Set OPENGENI_PUBLIC_BASE_URL to your public origin; sign-in and approval links use it. Run npx opengeni connect --base-url https://opengeni.example.com. Rate limits are shared across API replicas. Not supported yet: MCP Tasks and MCP event notifications. Use wait_for to follow a chat. Prefer a plugin with more skills? See Opengeni plugin.