Teach OpenClaw to do a repetitive browser task once. It runs automatically forever — no AI cost after that.
You show it the task. It records the steps. Every future run is pure Playwright — fast, free, deterministic. When something breaks, a cheap local model notices and a strong model fixes it. You pay for AI once, not every run.
ClawReflex is an OpenClaw skill that adds browser workflow graduation to your OpenClaw setup.
| Without ClawReflex | With ClawReflex |
|---|---|
| Strong model runs every step, every time | Strong model teaches once, then steps aside |
| Full token cost on every run | Token cost only when something breaks |
| Agent re-discovers the site on every run | Compiled Playwright macro — deterministic |
| Failure is silent or hard to interpret | Screenshot + plain-English error on every failure |
- OpenClaw installed and running
- Node.js ≥ 20
macOS:
open install/install-macos.command
# or: bash install/install-linux.shLinux:
bash install/install-linux.shWindows (PowerShell):
.\install\install-windows.ps1Restart OpenClaw after install. Then run npm run doctor to confirm everything is connected.
Open the OpenClaw dashboard and type:
teach a workflow
The skill asks what site, what task, what counts as success. You confirm. It records and compiles. Then:
# Copy the example input and fill in your values
cp ~/.openclaw/workspace/clawreflex/workflows/<slug>/macro/input.example.json \
~/.openclaw/workspace/clawreflex/workflows/<slug>/macro/input.json
# Edit input.json with your real credentials and settings
# Run it
npm run supervise <slug> ~/.openclaw/workspace/clawreflex/workflows/<slug>/macro/input.jsonYou'll see:
✓ my-workflow [step 7/7] download file
Status : success
Verdict : accept (confidence: 98%)
Rationale: Deterministic lane completed cleanly.
Run completed successfully.
Next run: npm run supervise my-workflow input.json
npm run doctor # check environment and workflow health
npm run teach <observed-run.json> # compile a workflow from a recorded session
npm run supervise <slug> <input.json> # run + validate + supervise
npm run run <slug> <input.json> # run macro only (no supervision)
npm run run-all # run all enabled workflows
npm run list # show all workflows with health status
npm run verify <slug> # re-check validation rules without running
npm run repair <slug> # generate a repair prompt for OpenClaw
npm run capture-auth <slug> # save a browser session (for MFA/SSO sites)
npm run schedule add <slug> --cron "…" # register a scheduled run
npm run schedule list # list all schedules
npm run schedule remove <slug> # remove a schedule
npm run export <slug> # export a sanitized shareable workflow
npm run import <file-or-url> # import a workflow
npm run report # token cost and ROI summary
npm run help # show this command listBy default, ClawReflex uses a local deterministic classifier — no model cost at all on clean runs. To add an AI supervisor for suspicious runs:
Ollama (local, free):
export SUPERVISOR_KIND=ollama
export SUPERVISOR_MODEL=phi3:mini
export SUPERVISOR_BASE_URL=http://localhost:11434HuggingFace or any OpenAI-compatible API:
export SUPERVISOR_KIND=openai-compat
export SUPERVISOR_MODEL=microsoft/Phi-3-mini-4k-instruct
export SUPERVISOR_BASE_URL=https://api-inference.huggingface.co/v1
export SUPERVISOR_API_KEY=hf_yourTokenHereCopy .env.example to .env and fill in your values. See config-templates/ for full setup examples.
When a Playwright locator fails, ClawReflex can fall back to a local vision model to find the element by screenshot:
ollama pull llava
export CLAWREFLEX_OCR_KIND=ollama-vision
export CLAWREFLEX_OCR_MODEL=llava
export CLAWREFLEX_ENGINE_PATH=/path/to/clawreflex/engine/src- How it works
- Beginner setup
- Runbook
- Troubleshooting
- Security and privacy
- Expert extension guide
- Limitations
npm test # 135 unit tests
SKIP_E2E=0 node --test tests/e2e.test.js # e2e (requires playwright install)MIT