1Install & first run

WebAgent Labs runs on every Chromium-based browser — Chrome, Microsoft Edge, Brave, Opera, Vivaldi, and Arc.

Install from the Chrome Web Store

1. Open the WebAgent Labs listing on the Chrome Web Store.
2. Click Add to Chrome → Add extension.
3. Pin the WebAgent Labs icon to the toolbar (click the puzzle icon → pin).
4. Click the WebAgent Labs icon to open the side panel.

Install from source (Edge / Brave / Opera or for developers)

1. Clone the repo: git clone https://github.com/profoncode-debug/WebWright
2. Open your browser's extensions page (chrome://extensions/, edge://extensions/, etc.).
3. Toggle Developer mode.
4. Click Load unpacked and select the cloned folder.

2Choosing a provider & model

Open Settings (gear icon, top-right of the sidebar). WebAgent Labs supports 8 LLM providers, each on its own tab:

• Ollama Cloud — free tier available. Get a key at ollama.com.
• Ollama Local — runs models on your own machine. Free forever. Zero data leaves your device.
• OpenAI — GPT-4o and newer. Bring your own API key.
• Anthropic / Claude — Claude Sonnet 4 and newer.
• Google Gemini — Gemini 2.0 Flash, free tier available.
• DeepSeek — DeepSeek-Chat, free tier.
• OpenRouter — one key for 400+ models from every major lab.
• Custom endpoint — any OpenAI-compatible or Ollama-compatible URL.

Model slots

You can configure four separate models per provider so the right model handles each job:

• Agent Model — picks actions in Agent Mode. Use a strong reasoning model.
• Vision Model — interprets annotated screenshots during vision escalation. Must be vision-capable.
• Chat Model — handles page Q&A in Chat Mode.

3Using Agent Mode

Agent Mode is for tasks where the agent must do things — navigate, click, type, fill forms, complete multi-step flows.

How to start a task

1. Open the side panel and navigate to the page you want to work on.
2. Type your goal in the input field at the bottom in plain English.
3. Press Enter — WebAgent Labs auto-detects that it's an action and runs the agent. (Use Ctrl + Enter to force Agent mode if the wording is ambiguous.)
4. The agent works on your current tab and runs step by step. You see every action in the live log above the input.

Example goals

Pausing & stopping

• Click Stop at any time to halt the agent immediately.
• The agent automatically stops after Max Steps actions (default 20) or Wall Timeout seconds (default 300).

4Using Chat Mode

Chat Mode answers questions about the page you're currently viewing. Multi-turn conversation, full page context.

How to start a chat

1. Navigate to the page you want to ask about.
2. Type your question in the side panel input.
3. Press Enter — WebAgent Labs recognises it as a question and answers in Chat mode (same input as Agent; no separate button).

If you navigate the tab or switch to a different one mid-conversation, the page context refreshes automatically on your next message — your conversation is kept, only the page content updates.

Text or screenshot? It decides by itself

There's nothing to configure. Chat sends the page text by default (fastest, cheapest), and automatically attaches a live screenshot when it's needed — on PDFs and visual pages, or whenever your question is about something visual (charts, images, layout, design).

Example questions

5Using Voice Mode

Voice Mode lets you operate WebAgent Labs hands-free — speak your request, and it speaks the answer back. The same intent routing applies: a question is answered in chat, an instruction launches the agent.

How to start

1. Click the mic button in the side panel to turn Voice Mode on.
2. The first time, your browser asks for microphone permission — click Allow.
3. Just talk. The voice bar shows what's happening: Listening → Thinking → Working (if it's running a task) → Speaking → back to Listening.
4. Click the mic again (or the ✕ on the bar) to turn it off.

Browser support (not available in Brave)

Voice Mode uses your browser's built-in speech recognition. It works in Chrome and Edge. It is not available in Brave, which disables the speech-recognition service for privacy — clicking the mic there just shows a message asking you to switch to Chrome or Edge.

Good to know

• While the agent is working, the mic is paused (bar shows Working) so it never mishears the task narration — it resumes listening the moment the task finishes.
• It remembers your recent voice turns, including actions it performed, so follow-ups like "what did you just do?" work.
• Text-to-speech runs on your device. URLs, markdown, and emojis are stripped so replies sound natural.

6Vision Escalation explained

When DOM-based actions fail, WebAgent Labs automatically climbs a 4-tier ladder. You don't trigger this manually — the agent escalates on its own when it detects it's stuck.

1. Tier 1 — DOM Analysis. Default. The agent reads the page's interactive elements and picks one. Fast and token-efficient.
2. Tier 2 — Vision + 80 marks. Triggered when a DOM click fails, an element isn't found, or the agent gets stuck in a loop. A screenshot is taken with up to 80 numbered markers overlaid on interactive elements, then sent to a vision-capable LLM.
3. Tier 3 — Vision + 160 marks. If Tier 2 still can't resolve the action, marker density is doubled for harder pages with many small elements.
4. Tier 4 — Raw Coordinates. Last resort. The LLM picks an X,Y point on the screenshot and WebAgent Labs dispatches a synthetic click via the Chrome DevTools Protocol.

The agent also has anti-loop detection: if it repeats the same action three times, oscillates between two elements, or detects that a step claimed success but the page state didn't change, it changes strategy or escalates.

7Settings reference

Open the gear icon in the side panel header.

• Provider — which LLM service to use. Default: Ollama Cloud.
• Endpoint, API Key, Models — set per provider.
• Max Steps — auto-stop the agent after N actions. Default: 20. Raise this for complex tasks; lower it to keep the agent on a short leash.
• Step Delay — pause between actions, in milliseconds. Default: 2000. Lower it for speed; raise it on slow sites.
• LLM Timeout — max wait per LLM call, in ms. Default: 100000 (100 s). Raise for slow models.
• Wall Timeout — max total task duration, in ms. Default: 300000 (5 min).
• Voice Mode — uses the browser's built-in speech recognition (Chrome/Edge). Not available in Brave.

8Tips for best results

The accuracy of any AI agent is a function of two things you control: how specifically you prompt it, and which model you point it at.

Prompting

• Name the site when relevant. "On Flipkart…" beats "find me a phone."
• State a stop condition. "…and open the cheapest in-stock listing" beats "buy headphones" (which runs forever).
• Spell out filters and constraints up front — under $X, in stock, before this date.
• Avoid pronouns. Restate the actual goal each time instead of "do that again."

Model choice

• Frontier reasoning models (GPT-4o, Claude Sonnet 4, Gemini 2.0 Flash, gpt-oss:120b-cloud) handle long agent loops most reliably.
• Mid-tier models (DeepSeek-Chat, Gemini Flash, gemma4:31b) are good for everyday navigation but may stall on complex multi-step flows.
• Small local models (qwen2.5-coder:7b, llava:13b) work for simple tasks but expect retries on complex sites.
• Always pair text-only providers with a vision model. Without one, vision escalation can't recover from DOM failures.

9Troubleshooting

"Ollama Cloud rejected the request (403)"

Your API key is missing, invalid, or has been rate-limited. Fix:

1. Go to ollama.com and sign up (free).
2. Generate a personal API key.
3. Paste it into Settings → Ollama Cloud → API Key.

"Local Ollama rejected the request (403)"

Ollama is blocking the extension's origin. Set the environment variable OLLAMA_ORIGINS=* and restart Ollama.

The agent is stuck in a loop

The agent's anti-loop detection usually catches this. If it doesn't:

• Click Stop to halt manually.
• Rephrase the goal to be more specific (sites, stop condition, constraints).
• Switch to a stronger model.
• Make sure a Vision Model is configured so vision escalation can fire.

Side panel won't open

Click the puzzle icon in your browser toolbar → pin WebAgent Labs. Then click its icon directly.

10Keyboard shortcuts

• Enter in the input → smart send (auto-routes to Chat or Agent)
• Ctrl + Enter in the input → force an Agent task
• Shift + Enter in the input → new line (multi-line prompt)
• Click Stop while the agent is running → halt it immediately