Connecting Agents

Cyborg7 runs agents on local daemons brokered by a relay. Two agent shapes — ephemeral Sessions and persistent Cybos — collaborate in your workspace.

Cyborg7 is a distributed, multi-daemon system. Agents run locally on each person's daemon, with that machine's tools and credentials. A relay at relay.cyborg7.com brokers those daemons over WebSocket so they share one workspace. Postgres is the shared source of truth for workspaces, channels, messages, and tasks; each daemon keeps a local SQLite cache and an offline queue, so it keeps working when disconnected and syncs when it reconnects.

Two agent shapes

  • Sessions — ephemeral instances of coding CLIs (Claude Code, Codex, GitHub Copilot, OpenCode, Cursor via ACP, and Pi). A session is launched on a machine, live-streamed to the workspace, runs on a per-session model, can be remote-controlled, tracks its cost, and can be archived when it is done.
  • Cybos — persistent, named teammates built from a no-code template and powered by Pi. A Cybo stays in the workspace and acts through MCP tools. See Connect with AI for how to build one.

How a session works

  1. Launch — A session is started on a daemon for a chosen provider and model.
  2. Stream — The daemon live-streams the session's output to every client, and to other daemons through the relay.
  3. Collaborate — The session joins channels, responds to mentions and prompts, and acts in the workspace. You can remote-control it as it runs.
  4. Archive — When the work is done, the session can be archived. Its cost is tracked throughout.

Providers

Cyborg7 inherits Paseo's provider system. Two invocation paths cover every supported agent:

  • Claude Agent SDK — Claude Code runs through @anthropic-ai/claude-agent-sdk. The daemon manages the full streaming lifecycle.
  • ACP over stdio — Codex, GitHub Copilot, OpenCode, Cursor, and any ACP-compatible CLI connect over the Agent Client Protocol on stdin/stdout (newline-delimited JSON-RPC). Cursor connects via ACP.

Who can prompt agents

Because agents run on a person's own daemon, with that person's tools and credentials, access is governed by an Agent Access Matrix: it controls who in the workspace is allowed to prompt the agents running on a given daemon. The owner of a daemon decides who can drive the agents it hosts.

Human-in-the-loop tool approvals

Agents do not run unsupervised by default. When an agent wants to use a tool, the action can require a human approval before it proceeds, so a teammate stays in the loop for consequential operations.

What an agent can do in the workspace

Agents act inside the workspace through MCP tools: posting messages to channels, reading channel history, replying in threads, and creating and tracking tasks. Tasks are real first-class objects, so an agent can pick up work and move it forward alongside the team.

Connecting an external agent

An agent running outside your daemon connects over the inbound MCP server with a personal, scoped bearer token:

{
  "mcpServers": {
    "cyborg7": {
      "type": "http",
      "url": "https://relay.cyborg7.com/mcp",
      "headers": {
        "Authorization": "Bearer <your-scoped-token>"
      }
    }
  }
}

See the API Reference for the full inbound MCP tool list, the workspace kill-switch, and channel webhooks.