Self-improving AI Employees for everyone
Android (android_tool)
| Field | Value |
|---|---|
| Category | tool |
| Backend handler | AndroidTool.execute_op via ToolNode |
| Tests | test_android_tool.py |
| Tool name | android |
Purpose
Give an existing AI agent a natural-language tool for the shared local Android phone. Setup and startup happen explicitly in Workspace → Mobile. This is separate from the legacy Android relay services.
The paired android-phone-skill is registered in nodes/visuals.json for the auto-add-skill policy. Internal phone steps publish iteration, max_iterations, phase, execution ID and run ID through update_node_status, scoped to the Android node and workflow. The square tool node and Workspace display those steps while the parent AI agent remains in its current tool-call turn. Manual takeover changes the phase to “Waiting for you”; resumed tasks retain their consumed step budget.
Inputs (handles)
| Handle | Required | Purpose |
|---|---|---|
input-model |
No | Optional single enabled OpenAI, Anthropic, or Gemini model connector overriding node/global selection for the embedded mobile-use engine |
Parameters
The AI-visible schema accepts only prompt: a required string, 1–20,000 characters. Whitespace-only requests are rejected by the shared execution path. Extra tool arguments are forbidden.
Saved configuration uses MobileParams: prompt defaults to empty, max_steps defaults to 40 (1–200), and timeout_s defaults to 900 (30–3600 seconds). In tool mode, the model supplies the request while the saved configuration controls the step/time limits. The global model is used by default; a parent agent’s custom model is not inherited automatically.
Model selection defaults to model_source="global", reading the toolbar’s current saved provider/model on each task. model_source="custom" uses the node’s provider and model; blank custom model uses the provider default. One connected model takes priority over both modes. Missing or unsupported global providers produce actionable errors. Supported providers are OpenAI, Anthropic and Gemini.
Outputs (handles)
output-tool connects to an AI agent's input-tools. The result contains response, outcome, run_id, and artifacts (currently empty). Standard tool dispatch returns a flat result or error; there are no main-flow handles.
Logic Flow
- Standard tool dispatch validates the model input separately from saved configuration.
- The operation combines the prompt with saved limits and calls the shared Mobile Agent execution path.
- That path verifies local owner access, saved workflow/execution identity, and the model connector.
- The shared runtime queues the task and grants a fenced device lease before starting the isolated mobile-use worker.
- The worker returns its result; cleanup revokes capabilities and drains pending device operations.
Decision Logic
Missing setup or a stopped phone produces an actionable error; execution never accepts SDK licenses or starts installation. Human control pauses agent use. Tasks use the same queue, cancellation and manual-resume rules as Mobile Agent. Automatic activity retries are disabled to avoid replaying phone actions.
Side Effects
Model API requests, device reads and mutations through the broker, a task worker subprocess, and bounded operational logs. The plugin reads saved model configuration and provider credentials server-side. Prompts and screenshots reach the selected model provider but are excluded from diagnostic logs.
External Dependencies
The installed optional mobile-use environment, managed Android SDK/AVD, local embedded Temporal worker, and a supported model credential. Node import only registers capabilities; it starts no emulator or downloads.
Edge cases & known limits
Windows x64 managed Android only; one shared phone and one video viewer. No iOS, remote device farm, arbitrary device attachment, or independent worker process. The task schema cannot override workflow scope, device serial, credentials, or saved execution limits.
Related
Live Android activity
Workspace shows the active phone task's model/provider, step budget, total elapsed time, current phase duration, and time since the last reported activity. The expandable activity panel retains the latest 40 events: model requests/responses and actual device operations (screen plus accessibility-tree reads, taps, scrolling, typing, navigation, and app actions). Completed operations include duration. The last task remains visible after completion, failure, or cancellation until another task runs; history is in memory and clears on restart. A 30-second quiet period shows a notice, not a claim that the engine is stuck. Use the last phase to distinguish a model wait from device I/O; take manual control using Use phone if needed. The panel is outside the phone canvas and hidden in full-screen mode. Activity contains operational summaries, never private model reasoning, typed text, URLs, or prompt content.