Tell your browser what to do. Watch it do it.
Pickle Browser searches, opens real pages, reads them, and reports back with its sources. It runs on your machine, in a window you can watch, so you see every step and can take the wheel at any point. Use a model that runs locally for free, or bring the Claude or ChatGPT you already pay for.
Free · runs locally with no account · your pages never leave your device
Give it a goal. Watch it go.
Unedited screen capture: it searches, opens real pages, reads them, and writes up what it found, while you watch every step and can take over at any point.
Real run, at real speed. Nothing staged, edited, or re-enacted.
You choose where the thinking happens.
Pickle Browser is not tied to one AI company. Three ways to power it, and you can switch at any time in Settings.
The Ollama engine ships inside the app. On first run it downloads one open model sized to your hardware (about 1 GB), then works offline.
- Good at summarizing the page you're on, pulling out data, answering questions about what's in front of you
- Models Qwen2.5, Llama 3.2, Phi-3, Mistral and other Ollama models
- Page content never leaves your device
Straight answer: small local models are far less reliable on long multi-step research. They can lose the thread and go off topic. The app labels them and warns you, and you can switch mid-task.
Want more capability without paying? Two routes are free today, and both are one-time setup.
- Ollama Cloud gpt-oss 20B and 120B, qwen3-coder 480B, after a free sign-in
- NVIDIA Build 100+ hosted models with a free API key
- Fast enough for real multi-step research on any laptop
These run on someone else's servers, so the page text for that task is sent to them. Your history, notes and logins still stay local.
Already subscribed to Claude or ChatGPT? Use it. Nothing new to buy.
- Your API key for Claude, GPT or Gemini
- Or connect over MCP from Claude Desktop, Claude Code, Cursor or VS Code and drive this browser with 72 tools
- Your credits go further: it sends a clean page summary instead of raw HTML
Connected agents get the same approval gates. Buying, sending and anything irreversible still asks you first.
Most agent browsers run somewhere you can't see.
Cloud agent browsers are infrastructure for developers shipping automation: headless, remote, and starting from a blank session every time. Pickle Browser takes the opposite bet. It's a browser you use, that an agent drives while you watch. It does cost less to run, but that's a side effect of being efficient, not the reason to switch. The reason is that the work happens in front of you, on your hardware, and you can stop it mid-step.
$0 per run on the built-in model
Instead of piping raw page source into the model, it sends a compact structured summary. Measured across ten real sites: a median 156× fewer tokens than raw HTML (range 4×–663×), or about 3× against already-extracted page text. Tools that read an accessibility tree land in a similar range. The saving that always holds is that the bundled local model costs nothing per run.
Runs on your computer, not a cloud
It uses your real browser and logins, and page content never leaves your device. No hosted browser, no account to create. And because a model ships in the box, it even works offline.
Your model, your window, your call
Connect any MCP agent. You choose the model, Pickle Browser is just the browser. You see every tab it opens, can grab the wheel any time, and it always asks before it buys.
The honest comparison.
There are four common ways to give an agent a browser today. Here's where each one leaves you, and where Pickle Browser is different.
| Pickle Browser | Chrome+ extension | Cloud AI browsere.g. hosted assistants | Dev automationPlaywright MCP / scripts | |
|---|---|---|---|---|
| The browser you already browse in, with the logins you already have | ✓ | ✓ | ✗ their cloud | ~ separate profile |
| A model is included, so it works with no cloud account | ✓ bundled | ✗ | ✗ their model | ✗ bring your own |
| Page content never leaves the device | ✓ on the local model | ~ per extension | ✗ | ~ depends on your model |
| Watch it work and take the wheel mid-task | ✓ native tabs | ✓ | ~ | ~ headed, separate window |
| Asks before buying or anything destructive | ✓ | ✗ | ~ | ✗ |
| Token-efficient page reads | ✓ measured | ✗ raw HTML | ~ opaque | ✓ accessibility tree |
| Use any model you want | ✓ any MCP | ~ | ✗ locked in | ✓ any MCP client |
| Works with your agent, no code | ✓ | ~ | ✓ | ✓ config |
| No per-use or subscription cloud fees | ✓ local | ✓ | ✗ subscription | ~ needs a model |
✓ built in · ~ partial or depends on setup · ✗ not without giving something up
Reading a page shouldn't cost 40,000 tokens.
A single modern web page is often 30–60k tokens of raw HTML. Pickle Browser gives the agent a structured read (headings, links, and interactive elements with stable ids) that's a small fraction of that, and logs exactly what it saved. The savings aren't a slogan; they're a number in the toolbar.
Bring the agent you already use.
Pickle Browser speaks the Model Context Protocol, so any MCP client, like Claude Desktop, can drive your real browser. Add one server entry and the agent gets a full set of browse, read, and shop tools.
- Install Pickle Browser. The local runtime and MCP server come bundled.
- Add the Pickle Browser server. One entry in your agent client's MCP config.
- Ask your agent to browse. It drives the tab you're watching.
{
"mcpServers": {
"pickle": {
"command": "pickle-mcp"
}
}
}
The app shows a copy-paste config on its Connect screen, matched to your client.
One cart across every store.
Pickle Browser Search reads real product catalogs over the Universal Commerce Protocol when a store supports it, and clean web results when it doesn't. Your agent fills one cart across merchants; checkout splits it into a real cart at each store.
Everything the agent needs, in one place.
Not a chat box bolted onto a browser, but a full toolkit for you and your agent, all running on your machine.
Runs on your machine
The engine is built in and pulls an open model on first run. After that: no keys, no account, works offline, $0 per run.
Give your agent a personality
One editable Markdown file sets its tone, rules, and boundaries.
Reusable playbooks
Deep Research, Compare Prices, Fill Form… built in, and you can add your own.
Research → a cited report
Decompose a goal, watch each step, get a structured deliverable with sources.
Replay any run
Step through what the agent did, screenshot by screenshot.
See what you save
Every step shows tokens saved vs sending raw HTML. Measured, not promised.
One cart across stores
Shop over UCP + the web, checkout splits into a real cart per merchant.
Your tools, via MCP
Plug in Gmail, Slack, Notion, or your own MCP server for the agent to use.
Download and hand your agent the keys.
Free, no account. Set up local AI in one click on first run, or connect the model you already use.
One codebase, all three platforms · ~120 MB · early build
The bundled local model is best for quick single-page work: summarizing, extracting, and answering about the page you are on. Small models are far more error-prone on long multi-step research, so connect a larger or cloud model when accuracy matters.