Yad (يد), self-hostable, privacy-first AI browser automation agent

Built by Abdellah Ouadoudi

Website / homepage / live URL

• The YAD website (live landing page): https://yadagent.com, own domain, purchased 2026-08-25, migrated live 2026-08-26. Also live in Dutch as a second language, every page has an -nl counterpart (for example /yad-nl, /yad-business-nl). Served from the same server as before (nginx vhost proxies / to the /yad route on the x402 server). The old https://wazir-x402.duckdns.org/yad URL still works and serves the same content, but yadagent.com is now the canonical one.
• Chrome Web Store listing (install page): https://chromewebstore.google.com/detail/dacfhekkemkiikecbjffmbdcohddodea, live and installable today.
• Microsoft Edge Add-ons: submitted 2026-08-26, in review (Microsoft's own estimate is about 7 business days).
• Brave: supported today, no separate submission needed. Brave installs the same Chrome Web Store extension directly, verified working via a real native-messaging handshake test, not assumed.
• Landing page source lives in site/index.html; the deployed copy lives at /app/euler-liquidator/data/yad-landing.html on the server. nginx config for the domain: /etc/nginx/sites-available/yadagent on the Hetzner box.

Yad ("the hand" in Arabic) is a thin Chrome extension (the Hand) controlled by a locally hosted brain (the Companion). It works inside the user's real, logged-in browser session, learns a task once, and then repeats it deterministically at near-zero cost.

Yad is two separate installs, on purpose. Part A, the extension, from the Chrome Web Store. Part B, the companion, as an npm package: npx yadagent pair (native messaging, pairs with Part A, your real logged-in browser) or npx yadagent serve (standalone, no extension, headless Chromium, your own Ollama instance). See packages/companion/npm-package or the package on npm.

Benchmark results (gpt-4o-mini, July 2025): 92% fully correct on 25-task open benchmark · 100% on 5 authenticated tasks · 100% pass+partial · avg 2.7 steps · avg 8.7s per task → See Benchmark section below.

Benchmark

YAD includes a fully open, reproducible benchmark suite. Every task definition, scoring script, and result file is checked into this repository, anyone can reproduce the numbers by running pnpm benchmark.

Results (July 2025, gpt-4o-mini via OpenRouter)

Results by difficulty (main suite):

Task categories

• product-research (8 tasks): navigate catalogue, find cheapest/rated item, open product page, read description
• content-extraction (7 tasks): read quotes, list authors, filter by tag, navigate to page 2, visit author profile
• web-research (5 tasks): Wikipedia infobox facts, publication years, geography
• news-monitoring (5 tasks): HackerNews titles, points, most-commented story
• authenticated (5 tasks): full login flow on saucedemo.com, sort, filter, navigate product pages
• academic-research (5 tasks): ArXiv paper titles, authors, abstracts
• reference-lookup (3 tasks): Cambridge Dictionary definitions
• code-research (2 tasks): GitHub public repository info

How to run

# Install deps
pnpm install

# Run full 25-task benchmark (requires YAD companion on localhost:3747 + Chrome extension)
pnpm benchmark

# Run authenticated tasks
pnpm benchmark --tasks data/benchmark-tasks-auth.jsonl

# Run extended competitor-coverage tasks
pnpm benchmark --tasks data/benchmark-tasks-extended.jsonl

# Run a single task
pnpm benchmark --id bk-001

# Run with visible browser
pnpm benchmark --headed

# Use a custom model
YAD_MODEL=gpt-4o pnpm benchmark

Scoring: pass (1.0) = agent finished + all expected keywords matched in summary. partial (0.5) = agent finished but task has no verifiable keywords. fail (0.0) = agent did not complete the task.

Methodology notes

YAD uses accessibility tree snapshots (roles, refs, text) rather than screenshots. This makes it faster and cheaper than vision-based agents, but blind to purely visual structure (canvas, custom dropdowns, icon-only controls). A vision fallback for those cases is a known gap, not yet built.

The agent loop runs a multi-step plan-execute cycle with:

• 5-call same-URL goal-drift guard (with LLM judge check)
• 6-call no-progress guard
• Self-healing recovery loop: 8 stuck signals → LLM generates recovery plan → stored in recovery-store for future runs
• Extract-loop guard: forces finish after 2 consecutive extracts on the same URL (model already has the data)
• 7-day action-cache for zero-LLM-cost replays on repeated tasks

The benchmark measures whether the agent completes the goal and puts the correct data in its finish summary.

Comparison with other web agents

Key differentiators:

1. Persistent logged-in session. All major competitors (browser-use, Skyvern, WebVoyager) start a fresh browser per task. YAD works inside the user's real Chrome with all cookies and sessions intact. This is the difference between "demo on an empty account" and "real daily automation."

2. Accessibility tree, not screenshots. Reading roles and text directly is faster and cheaper than sending images to a vision model. YAD's avg task time is 8.7s vs WebVoyager's minutes per task.

3. Recovery loop, not yet claimed as full learning. YAD detects 8 types of stuck signals (goal-drift, consecutive failures, url-regression, etc.), automatically generates an LLM recovery plan in under 3 seconds, and stores successful patterns in a recovery-store for instant reuse on the same site and signal. That is proven recovery, not general-purpose learning across sites, and this README will not claim more than that until it is built.

4. Near-zero cost on repeated tasks. The action-cache stores proven action sequences (7-day TTL). Repeated tasks run with 0 LLM calls, returning in about 2 seconds. Cache hit rate grows with usage.

5. NanoBrowser comparison. The closest Chrome-extension competitor. As of July 2025, NanoBrowser publishes no benchmark scores. YAD is the only Chrome-extension web agent with an open, reproducible benchmark.

Benchmark scope note, read this before quoting a number from this README

YAD's 25-task suite is intentionally smaller than WebVoyager's 643-task benchmark, and the two were not run under identical conditions. The tasks cover similar categories (product research, content extraction, web research, news monitoring) on deterministic test sites (books.toscrape.com, quotes.toscrape.com) and live sites (Wikipedia, HackerNews). The static sites ensure reproducibility across runs; the live sites may cause minor variance.