How to use LLM Browser from first launch to real workflows

1. Learn the two parts of the app

The first part is the browser you already understand: tabs, an address bar, bookmarks, history, downloads, settings, and web pages. The second part is the bridge panel, the assistant chat docked in the same window. You browse normally until you want help, then you ask the bridge panel to read or act on the page in front of you.

A helpful mental model is simple: the assistant can only be useful when it has a clear page, a clear task, and a clear stopping point. Instead of saying "sort this out", say what page to use, what result you want, and when it should stop for your review.

2. Install it and browse normally first

Install LLM Browser like any desktop app. Open it, search the web, open two or three tabs, and sign in to any site you plan to use. Do this before giving the assistant a task. It helps you confirm that the browser itself is working, that the page is loaded, and that you are looking at the right account.

Open Settings → Browser to set your homepage, search engine, theme, bookmark bar, and privacy preferences. None of this requires AI. It is just the browser foundation the assistant will later work inside.

3. Connect an AI model

Open Settings → LLM Connection. This panel decides which model powers the assistant. Choose the connection type that matches how you want to work:

• API Key: use a hosted provider by pasting an API key and choosing a model.
• OAuth: sign in through a supported work provider when that is how your organization manages access.
• Local Server: use Ollama, LM Studio, llama.cpp, or another local endpoint running on your machine.
• Local AI: load a GGUF model file directly for an on-device setup with no separate server.
• CLI Pipe: send prompts to a command-line tool and read the output back into the browser.
• Terminal: run a saved terminal command as the model backend for custom local workflows.

Fill in the fields for the tab you choose, then press Test. Wait for the green Connected banner. If the test fails, fix the model name, endpoint, token, or local server before asking the assistant to browse.

4. Pick cloud AI or local AI on purpose

A hosted model is usually stronger and easier for long tasks. It can be the better first choice for comparing pages, reading messy sites, or working through a multi-step web chore. A local model keeps the work on your machine and avoids per-token billing, which is better for private browsing, internal dashboards, client notes, inbox triage, or anything you do not want sent to a cloud model.

You do not have to choose forever. Keep a local model for everyday private chores and switch to a stronger hosted model only when a task needs it.

5. Start with tiny visible tasks

Your first prompts should be small. Good starter prompts are:

• "Summarize the visible page in five bullet points."
• "Find the price, cancellation rule, and main limitation on this page."
• "Compare this tab with the next tab and make a two-column table."
• "List the buttons or links I would need to click to finish this task, but do not click them yet."
• "Fill the visible form fields from these details, leave anything uncertain blank, and stop before submit."

These prompts teach you how the assistant reads labels, switches tabs, and reports progress before you ask it to do longer work.

6. Use a safe prompt pattern

For real websites, use this pattern:

1. Tell it the goal: "Find three suitable jobs", "compare prices", or "fill this form".
2. Give the source of truth: "use only the visible page", "use my CV text below", or "use the details in this message".
3. Set boundaries: "do not submit", "do not send messages", "do not buy anything", or "ask before changing settings".
4. Request a result format: "make a table", "give me a checklist", or "summarize what you changed".

A strong example is: "On this page, fill the application fields from the profile below. Leave anything uncertain blank. Stop before submitting. At the end, list every field you filled and every field you skipped."

7. Fill forms without losing control

LLM Browser is useful for repetitive forms because it can read field labels, type into boxes, choose dropdowns, tick checkboxes, and stop for your review. The safest habit is to make the assistant prepare the form, not finish the transaction. Ask it to stop before submit, then review the visible page yourself.

For important forms, add this sentence: "If a field is ambiguous, leave it blank and explain what you need from me." That prevents the assistant from guessing where a human choice is required.

8. Turn long chores into checkpoints

Long tasks work better when split into checkpoints. Instead of "find and apply to jobs", try:

1. "Find ten relevant jobs and put them in a table with links."
2. "Remove anything that does not match these constraints."
3. "Open the top three and summarize the application requirements."
4. "Fill one application draft and stop before submit."

This is slower than one giant prompt, but it is much easier to review, correct, and trust.

9. Tune reasoning, system prompt, and step limit

Every connection has settings that change how the assistant behaves. Use a lower reasoning level and a short step limit for fast chores such as extracting a price or filling a small form. Raise the reasoning level and step limit for multi-page research, messy forms, or tasks where it must compare details.

The system prompt is where you set standing rules. For example: "Always stop before submitting forms. Prefer exact page evidence. Leave uncertain fields blank. Summarize actions at the end." A simple standing rule like that makes the assistant more predictable.

10. Save repeatable workflows

When a prompt works, save it. A good saved workflow is a checklist, not a vague wish. For example:

"Research a product: open the top three official product pages, collect price, platform support, privacy policy notes, trial length, and cancellation terms. Make a table with source links. Do not sign up or buy anything."

Reusing checklist prompts gives the assistant a stable shape to follow and gives you an easy way to improve the workflow over time.

11. Use local AI for private browsing

If privacy matters, run the model locally. Start Ollama or LM Studio, choose Local Server, enter the local endpoint and model name, then press Test. Or choose Local AI, select a GGUF model file, set the served model name and context size, then test it.

Local models are often smaller, so give them smaller jobs: summarize the visible page, extract a table, clean up form text, compare two tabs, or draft a response. For complicated reasoning across many pages, switch to a stronger model if you are comfortable doing so.

12. Use the Bridge Server when another app should drive the browser

Developers and automation builders can open Settings → Bridge Server, enable the local control server, choose the port, and copy the bearer token. External clients can then control LLM Browser over JSON-RPC through HTTP or WebSocket. That is useful when another agent, script, or desktop app should navigate the same real browser you can see.

Keep the token private, regenerate it if needed, and turn the server off when you do not need outside control.

13. Troubleshooting checklist

• The assistant does nothing: check that the model connection shows Connected and the page is fully loaded.
• The answer is vague: ask it to use only visible page evidence and quote exact labels or links.
• It misses fields: zoom out, scroll to the form, and ask it to list fields before filling them.
• A local model is slow: lower the step limit, choose a smaller task, or use a smaller/faster model.
• You are nervous about a site: add "stop before submit" and watch the task step by step.

Troubleshooting is easier when you check the connection type first. If API Key fails, confirm the key, provider base URL, model name, and billing access. If OAuth fails, sign out and sign in again, then check whether your work account allows the model you selected. If Local Server fails, open Ollama or LM Studio, make sure the model is loaded, and test the local endpoint in the browser before returning to LLM Browser. If Local AI is slow or crashes, reduce context size, choose a smaller GGUF, or shorten the task. If CLI Pipe or Terminal fails, run the command by hand first and make sure it returns plain text.

For page automation problems, ask the assistant to inspect before it acts. A useful reset prompt is: "List the visible buttons, links, forms, and tables on this page. Do not click anything yet." Once you see what it thinks is available, give it one precise action at a time.

14. A complete beginner workflow

Here is a full safe workflow you can copy:

1. Open the site you want to work on and sign in yourself.
2. Open the page where the task starts.
3. Check that the AI connection is green.
4. Ask: "Read this page and tell me what actions are available. Do not click anything yet."
5. Choose one action and ask the assistant to do only that step.
6. Ask for a summary of what changed.
7. Review the page yourself before any final submit, send, purchase, or delete action.

Once that feels natural, combine steps into longer prompts. The best LLM Browser users do not hand over everything at once; they guide the assistant through visible, reviewable steps.

15. Workflow playbooks you can copy

These are practical starting points. Paste the prompt, replace the bracketed parts, and keep the stopping rule until you trust the workflow.

Research a buying decision

"Open the official pages for [product A], [product B], and [product C]. For each one, collect price, platform support, privacy notes, refund or cancellation terms, and the strongest limitation you can find. Use only the pages you open. Make a table with source links. Do not sign up or buy anything."

Extract a table from a messy page

"Read the visible page and find the table/list about [topic]. Extract the rows into a clean table with columns for name, link, date, status, and notes. If a value is missing, write 'not shown' instead of guessing."

Prepare a form safely

"Fill the visible form from the details below. Leave anything uncertain blank. Do not submit. When finished, list every field you filled, every field you skipped, and anything you need me to decide."

Compare job listings

"Find five roles matching [title], [location or remote rule], and [salary or seniority]. Put them in a table with company, role, link, location, salary if shown, must-have skills, and why it fits. Do not apply yet."

Clean an inbox or dashboard without risk

"Scan the visible inbox/dashboard and group items into urgent, reply later, archive candidate, and ignore. Do not archive, delete, send, or mark anything. Give me the list first so I can approve actions."

16. Review before final actions

LLM Browser can help with real websites, but final actions deserve a human pause. Review before a submit, send, purchase, delete, unsubscribe, password change, permission change, or account setting change. The best pattern is to let the assistant prepare the page, explain what it changed, and stop. Then you decide whether to press the final button yourself.

Where to go next

For privacy-first use, read the local LLM browser guide. For machine-to-machine automation, read how to control LLM Browser over the local API. For everyday web chores, see the guide to using an AI browser to fill forms.