qwen3-vl-8b · Qwen-Agent · OpenTrust

ΨClaw — control your machine like magic.

A fine-tuned desktop companion model for OpenClaw. PsiClaw operates across your browser, native apps, and file system — with persistent memory, API-first execution, and a confirmation-first safety policy.

Why this exists

Mission profile

ΨClaw is built on qwen3-vl-8b — a model specifically designed for PC/mobile GUI operation — and ships as the default desktop companion inside OpenClaw.

Observe browser, native apps, and filesystem state in full.

Prefer direct API calls. Fall back to visual automation only when needed.

Require confirmation before any irreversible action.

Persist user identity via OpenTrust — memory as reasoning, not retrieval.

Active desktop sessions
12
+3 today
Approval queue
04
2 high priority
Task success rate
93.4%
+4.2% this week
Mean intervention time
11s
-6s vs baseline

Desktop-aware observation

PsiClaw reads the full desktop surface: browser tabs, native apps, active windows, focused elements, and system state — all as structured context for the model.

API-first, browser as fallback

When a web service has a discoverable API, PsiClaw calls it directly. DOM automation is the fallback for visual-only tasks, keeping execution fast and reliable.

Persistent identity via OpenTrust

User preferences, working patterns, and long-term context are reasoned over — not just retrieved. PsiClaw learns you across sessions, not just within them.

Safe action loop

Operator workflow

1

Observe current desktop state: active windows, browser state, running processes, clipboard, and recent interaction history.

2

Determine whether an API skill is available for the target action. Prefer direct API over DOM automation when possible.

3

Propose the next action with confidence, rationale, and risk level. Request approval for irreversible or high-impact steps.

4

Execute in a scoped, observable way. Re-read state after each action before deciding the next step.

5

Capture the full trace — observation, reasoning, action, outcome — for replay, evals, and future fine-tuning.

What the demo proves

Product story

Operator console: live desktop state, proposed actions, and approval queue.
Desktop Gym: structured training scenarios across browser, native apps, and terminal.
Trace explorer: full causal chain from observation → API/DOM → outcome.
Eval dashboard: success rate, intervention burden, API routing, memory quality.