Skip to content

Chat Tab

The dashboard ships a Chat tab that talks to the same multi-agent runtime as bubbaloop agent chat, but over HTTP+SSE instead of Zenoh. The browser is a thin client — every LLM call runs daemon-side. No SDK or API key lives in the browser.

Browser              Daemon (agent runtime)            LLM provider
 Chat tab ── HTTP ──> /api/agent/* ── SSE ── AgentEvent stream
   │                       │
   │                       ├── agent loop (per-agent tokio task)
   │                       ├── tool dispatch (47 MCP + 43 internal tools)
   │                       └── ContentBlock::Image for vision turns
   │
   └── POST /cancel ─> abort the in-flight turn (same semantics as Ctrl-C)

What you can do

  • Pick an agent. The picker is populated from the daemon's per-agent manifest queryables. Unknown agent IDs surface a 404.
  • Pick a provider. The status indicator shows which providers are configured. Gemini is the recommended default (free tier, strong vision); Claude works once bubbaloop login is done; Ollama works if ollama serve is running locally. If you choose a Claude OAuth subscription, the UI shows a risk warning — see the bubbaloop login --status docs.
  • Use tools. Tool calls show as collapsible chips inline; the result follows. Tool-call traces help you debug what the agent actually did vs. what it said it did.
  • See images. When the agent calls grab_frame, the daemon decodes the JPEG payload + JSON metadata attachment from the camera node and emits a ContentBlock::Image for the next LLM turn. The Chat tab renders the image inline so you can compare what the model saw vs. what it said.
  • Cancel a turn. Press the cancel button (or just close the tab) — the dashboard posts to the per-agent cancel endpoint, which publishes to bubbaloop/global/{machine_id}/agent/{agent_id}/cancel. The daemon aborts the in-flight LLM call.

Transport

  • HTTP for request/control, SSE for the response stream. Connects to the daemon's MCP HTTP port (8088 by default — override with BUBBALOOP_MCP_PORT).
  • Bearer token auth: the token is read from ~/.bubbaloop/mcp-token on the daemon side and is picked up automatically when the dashboard is served from the same host.
  • The wire shapes mirror AgentMessage / AgentEvent from the Zenoh gateway. See Concepts → Messaging.

When to use Chat vs. bubbaloop agent chat

  • Chat tab — best for visual workflows, multi-camera vision, remote (LAN) operators, anyone who already has the dashboard open.
  • bubbaloop agent chat — best for scripting (pipe-friendly stdout), SSH sessions, and quick one-liners. Same agent, same memory, same tools.