LocalPilot
A browser agent in Chrome's side panel, powered by an AI model on your own PC. LocalPilot opens pages, reads them, clicks, types, fills forms, saves data to your PC and takes screenshots, while you watch and approve what it does. The model runs on your computer (Ollama, LM Studio, llama.cpp, vLLM, LiteLLM, FreeLLMAPI, or any OpenAI/Anthropic-compatible server), so pages and chats never go to a cloud account.
▶ Watch the 60-second tour, with sound · Install in 3 steps
- Your real Chrome: your logins, cookies and tabs; no account, no subscription, no telemetry
- You stay in control: it asks before acting on a new site, and you can stop it at any time
- Opt-in CAPTCHA solver that runs in your own browser and hands the tab to you when a check needs a person (tested sites)
- Saved tasks on a schedule that replay their steps without the AI, and notify you when the result changes
/test writes a bug report for any site; /record saves a captioned GIF of a task- Works with Claude Code through an MCP server, with its own tab for each session
- Status: 0.10.0 · unpacked extension for Chrome 116+ (Chromium browsers work too) · the PC-bridge installer is Windows-only
Contents
See it work
Every clip was recorded by LocalPilot itself, on real websites. More runs, reports and spreadsheets are in examples/.
Why LocalPilot
| |
|---|
| Private by default. | The model is yours. Point it at a local server and no page, screenshot or prompt leaves your machine. No account, no cloud dashboard. |
| Your real browser. | It works in your normal Chrome profile — your logins, cookies, extensions and tabs — not in a separate automated browser. |
| You stay in control. | Approvals per site (allow once / always / deny), “ask before every action”, a plan-first mode, a visible working frame with Stop, and a pause card whenever a check needs you. |
| CAPTCHA solving, locally. | Off by default. When enabled, it clicks simple checks like a person would — and when it cannot pass one, it stops, tells you, and continues after you solve it. See the tested sites → |
| More than a browser driver. | PC files in allowlisted folders, job-board recipes, a website test report, GIF recording, saved tasks on a schedule, and an MCP server for Claude Code. |
The panel shows every step, asks before acting on a new site, and names the model in use.
Quick start
1. Load the extension
- Download or clone this folder.
- Open
chrome://extensions, turn on Developer mode, click Load unpacked and pick the folder.
- Pin LocalPilot and press Alt+Shift+L to open the side panel.
2. Connect a model
The easiest free setup is Ollama:
ollama pull qwen2.5vl:7b
setx OLLAMA_ORIGINS "chrome-extension://*" # Windows; once, then restart Ollama
On macOS/Linux, start Ollama with OLLAMA_ORIGINS="chrome-extension://*" ollama serve instead.
In the panel: gear icon → Preset “Ollama (localhost:11434)” → Test & list models → pick a model → Save.
Any other OpenAI/Anthropic-compatible server works the same way. The full list and troubleshooting are in docs/REFERENCE.md.
3. Give it a task
Open a site, open the panel, and type:
Summarize this page in three bullet points.
Then try: “Find the contact link and open it.” · “Fill this form with my name Jane Doe, but do not submit.” · “Save every result on this page as a CSV.”
The optional PC bridge (local files) and the MCP server are two short steps: PC bridge · Claude Code / MCP.
What it can do
| Feature | What you get |
|---|
| Browse and act | Navigate, read the page as a compact accessibility tree with reusable refs, click by ref or numbered screenshot mark, type, scroll, handle tabs and dialogs. |
| Fill forms | React/Vue inputs, selects, checkboxes and web components; saved values (never shown to the model), password fields refused; the job-application helper stops before submitting. |
| Collect and export | Extract repeating lists, read a site's own JSON via captured network responses, write CSV/JSON/Markdown into allowlisted PC folders. |
| Job search | Recipes for BOSS直聘, 猎聘, 51job, 智联 and LinkedIn — collected, de-duplicated, saved as CSV with an import draft — plus a guided application filler. |
| Test a website | /test crawls the site read-only, checks for JS errors, broken links, unlabeled fields, tiny phone layouts and slow pages, then writes a report with screenshots and a GIF. |
| Record | /record … saves a captioned GIF of the task, with a red circle on every click. |
| Schedule | Save a task and run it daily, weekly, every N minutes, or as a page watch that notifies you when something changes. |
| Drive it from Claude Code | The MCP server exposes the same browser tools to Claude Code. Each session gets its own tab in one unfocused window, so sessions never take over each other's tabs, or yours. |
| CAPTCHA checks | Optional local solver for click-based checks; hands over to you when it cannot pass. Details → |
The complete tool list, every setting and the internals are in docs/REFERENCE.md.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
