A2A Servers¶

Browser, desktop, and mobile agents can be exposed as A2A servers.

Browser Server¶

uisurf_agent run browser_agent \
  --mode a2a \
  --host 0.0.0.0 \
  --port 8001

If --port is omitted, the browser server defaults to 8001.

Desktop Server¶

uisurf_agent run desktop_agent \
  --mode a2a \
  --host 0.0.0.0 \
  --port 8002

If --port is omitted, the desktop server defaults to 8002.

Mobile Server¶

uisurf_agent run mobile_agent \
  --mode a2a \
  --host 0.0.0.0 \
  --port 8003 \
  --mobile-serial 5C060DLCR002MM

If --port is omitted, the mobile server defaults to 8003. If --mobile-serial is omitted, the server uses MOBILE_DEVICE_SERIAL or the first authorized ADB device.

Runtime Controls¶

A2A servers read .env through python-dotenv. Shared controls:

Variable	Description
`AGENT_HOST`	Default bind host when `--host` is omitted.
`MAX_OBSERVATION_IMAGES`	Number of recent image-bearing observations that retain image payloads.
`INCLUDE_THOUGHTS`	Global default for thought streaming when supported.
`AUTO_MODE`	Global default for automatic approval of safety-gated A2A actions.

Per-agent controls:

Agent	Port	Public URL	Max steps	Auto mode
Browser	`BROWSER_AGENT_PORT`	`BROWSER_AGENT_PUBLIC_URL`	`BROWSER_AGENT_MAX_STEPS`	`BROWSER_AGENT_AUTO_MODE`
Desktop	`DESKTOP_AGENT_PORT`	`DESKTOP_AGENT_PUBLIC_URL`	`DESKTOP_AGENT_MAX_STEPS`	`DESKTOP_AGENT_AUTO_MODE`
Mobile	`MOBILE_AGENT_PORT`	`MOBILE_AGENT_PUBLIC_URL`	`MOBILE_AGENT_MAX_STEPS`	`MOBILE_AGENT_AUTO_MODE`

Observation settings also have per-agent overrides:

BROWSER_FAST_MODE
DESKTOP_OBSERVATION_DELAY_MS
MOBILE_OBSERVATION_DELAY_MS
BROWSER_OBSERVATION_SCALE
DESKTOP_OBSERVATION_SCALE
MOBILE_OBSERVATION_SCALE

Provider Selection¶

The same provider flags are available in A2A mode:

uisurf_agent run browser_agent \
  --mode a2a \
  --provider gemini \
  --model-id gemini-3-flash-preview

Human Confirmation¶

When a model emits a tool call requiring confirmation, the A2A executor returns an approval_required payload and moves the task into input_required.

The pending confirmation is resolved when the client sends an approval or denial response for that task.

MCP Mode¶

The CLI exposes --mode mcp as a reserved mode, but MCP server mode is not implemented yet. It exits with a clear error instead of starting a server.