Browsie is a Chrome side-panel agent that reads the page like a human does, decomposes your goal into subgoals, then navigates, fills, clicks, and extracts until it's done. Runs against your local LLM by default — your data never leaves the box.
Most "browser agents" are one big prompt that loops until it hallucinates. Browsie isn't. Three small roles — the Director, the Actor, the Reflector — pass a checklist between them, and the agent itself stops to ask when it's unsure or auto-pauses when it's spinning. You can halt the run at any time.
Reads your goal once and decomposes it into 1–6 ordered subgoals, each with a clear "done when" predicate. Runs once at the start of the run, not every step.
Each turn: snapshot the active tab as a compressed accessibility tree, pick exactly one tool action — fill, click, extract, navigate — and hand it to the executor. No DOM hallucination.
After every action, the Reflector decides whether the current subgoal is satisfied. It advances the checklist, or it says "not yet" and the loop continues. When all subgoals are done, the run ends.
Browsie's actor only emits one of these tool calls per turn. That's the entire surface area. No code execution, no file write, no shell.
A short list, on purpose. Browsie isn't trying to be a platform. It's a focused cockpit for the 80% of browser tasks that are tedious, repetitive, and obviously a job for software.
The subgoal checklist is right there in the side panel. As the agent works, items tick green. When it's stuck, it tells you which one — not just "I failed."
Every action with the agent's one-line thought, color-coded by outcome. Scroll back through what it did, why, and what came back from the page.
Promote any goal to a reusable task. Schedule it to run on an interval or at a future time. Re-run by name. The morning routine writes itself.
Ask for one field, get a string. Ask for a list, get a CSV. The accessibility-tree snapshot keeps the agent honest about what's actually on the page.
Name, email, address, freeform preferences — saved locally, sent to the LLM only when the form needs them. The agent fills with your real data, not placeholders.
Refuses to submit, send, or buy by default. Click-loop, no-effect, and step-cap detectors auto-pause when the agent is spinning. You stay in the loop.
A browser agent reads everything on the page. That's a lot of trust. Browsie's answer is to never have the data in the first place — no cloud, no telemetry, no analytics, no contact with anyone but the LLM endpoint you point it at.
Configured for LMStudio / Ollama / any OpenAI-compatible endpoint. The default points at localhost. Page snapshots travel exactly as far as your inference server.
Your library, settings, saved tasks, and run history live in chrome.storage.local. Never synced. Never uploaded.
Zero outbound calls except (a) the LLM endpoint you configured and (b) the page the agent is currently working on. There are no third parties.
Browsie is unpacked-loadable in dev mode. Open the folder. Read the source. The whole agent fits in one short evening.
Browsie is distributed as an unpacked Chrome / Edge extension. Drop the folder in, point it at your local model, type a goal, and watch.
chrome://extensions. Enable Developer mode. Click Load unpacked and select the folder.[CFG], paste your endpoint and model name. LMStudio default works out of the box.