The built-in agent

PhysiClaw isn’t only a set of tools an outside model calls — it ships its own agent, a brain that can drive the phone on its own.

You can run PhysiClaw two ways. As a plain MCP server, it hands the twelve tap/swipe/peek tools to whatever agent you already use — Claude Desktop, an IDE, your own client — and that external model does the deciding. As a built-in agent, PhysiClaw is the model: it runs its own look → decide → act loop, in its own process, with no external client attached. Same robot, same tools — the difference is whose mind is in charge.

Coming from OpenClaw?

OpenClaw is an agent that drives software through the screen it lives on. PhysiClaw’s built-in engine follows the same principles — a native tool-call loop, normalized at the provider boundary — but points them at a physical phone. It’s OpenClaw’s playbook with a body: the agent that used to click pixels now reaches out and taps glass.

What the agent adds over a plain MCP server

A plain MCP server is reactive: it sits still until some external client sends a tool call. The built-in agent closes that gap — it can be the initiator.

It owns the loop

The agent runs the full look → decide → act cycle itself: peek to see, choose a bbox and a gesture, tap, then peek again to check. No external model in the loop.

It runs unattended

It wakes on its own — on a schedule or when the phone screen changes — operates the phone, and goes back to sleep. Nobody has to be sitting at a client.

It remembers

A persistent memory carries facts across wakes, so the agent isn’t starting cold every time. (See Memory & skills.)

It learns routines

Skills are reusable, app-specific playbooks the agent discovers and follows — “how to send a WeChat message,” “how to place a grocery order” — instead of re-figuring-out each app every time.

How the agent thinks

Every wake runs the same loop you already met in How it works — the agent just drives it instead of an external client:

 wake ──► LOOK ──► DECIDE ──► ACT ──► LOOK ──► … ──► close
 trigger  peek     pick a      tap /   peek      (DONE / WAIT /
 fires    (camera) bbox +      swipe  again,      FAIL / IDLE)
                   gesture            re-decide

A few rules keep the loop honest. Each turn is shaped as exactly [note, one-other] — one running-summary note plus one real action — so the agent takes one step, records why, and never fires a burst of taps blind. Every turn ends by looking at the result, so a popup or a slow load is just the next state to react to, not a script to fall out of. And each session ends with a one-word verdict — DONE, WAIT, FAIL, or IDLE — that says what happened and whether to follow up.

One brain, many models

The built-in agent isn’t tied to one vendor. It runs on Anthropic, OpenAI, Google, Moonshot, or Qwen through its own engine, or shells out to Anthropic’s claude CLI as a subprocess. You pick the model with one config line. Models covers the full list and how to switch.

When to use which

You don’t have to choose once and for all — the same install does both.

	Plain MCP server	Built-in agent
Who decides	your external client (Claude Desktop, an IDE)	PhysiClaw itself
Starts a task	you, by prompting the client	a trigger: a schedule or a screen change
Runs unattended	no — needs a client connected	yes — wakes, acts, sleeps
Memory & skills	up to your client	built in

Reach for the plain server when you want to keep an existing agent in the loop and just give it hands. Reach for the built-in agent when you want PhysiClaw to run on its own — a recurring chore, a watch-and-react task, a phone that does things while you’re away.