v0.3.0 — First run, fixed

Everything here is about the first five minutes. 0.2.0 was the first release under the name Flowkite; installing it, it turned out, was the easy part.

The panel would not take a task until you found two hidden settings

The side panel unlocks when an agent has a model, but its setup screen said "configure your API keys in the settings page". Follow that exactly and you came back to the same screen, with nothing saying that two more picks were needed in dropdowns further down the options page.

• A saved provider is now enough to start. Saving one points the Planner and Navigator at its flagship model and the Fast agent at its cheap tier, when the provider names one. Only ever on a setup with nothing configured, so it can never move a choice you made yourself.
• Setup happens in the panel. Pick a provider, paste a key, and you are in the chat. No second tab, no hunting.
• Ollama sits in that same list, and swaps the key field for a server address. It is the free path and it was buried.

A wrong key is caught immediately

Before, a mistyped key surfaced at the worst moment: the first task, mid-run, as a model error, after you had already approved a plan. Saving now asks the provider whether the key works, using the cheapest request that proves it, which costs no tokens.

Only an outright rejection stops the save. A timeout, a rate limit, a provider that offers no way to check: none of those block you, because none of them mean your key is bad.

Alt+Shift+F opens the panel

There was no keyboard shortcut at all. Rebindable at chrome://extensions/shortcuts.

Also

• The end-to-end harness runs again. Chrome 137 removed --load-extension for automated sessions and 151 closed the escape hatch, so pnpm e2e had been starting a browser with no extension in it. It now installs the extension the supported way, on an ordinary Chrome.

Install

Not on the Chrome Web Store yet.

1. Download flowkite-0.3.0.zip below and unzip it.
2. Open chrome://extensions/ and turn on Developer mode.
3. Click Load unpacked and select the unzipped folder.

Chrome or Edge. Bring an API key for OpenAI, Anthropic, Gemini and the rest, or point it at a local Ollama and pay nothing.

Upgrading from 0.2.0: your providers, models, history and settings are untouched. Nothing here re-configures an install that already works.

v0.2.0 — Flowkite

First release under the name Flowkite. The previous one (v0.1.13) predates the rename and is missing everything below — if you installed from it, replace it with this build.

Flowkite is still 100% free and local: no account, no backend, no telemetry you can't turn off. The only cost is your own LLM provider bill.

Control and safety

• Approval modes. Approval is now one choice in the composer instead of two settings buried in Options. manual gates every page-touching action including navigation, planner gates the plan (previous behaviour), auto gates nothing — including before money and credentials, so switching it on takes an acknowledgement that names exactly what it turns off. Changing the mode pushes into a task that is already running rather than only saving to storage.
• Human handoff. The agent can park itself with ask_user and hand the tab back to you for a login or a captcha. Your credentials go straight into the page and never through the model.
• Privacy dashboard. A new Options pane answers "what left this machine?" from local records only: token totals per model, every host the agent itself visited, and every webhook delivery. Capped, never uploaded, one-click clearable. Your own browsing is never logged — only the agent's navigation.
• No more numbered boxes on your page. The highlight overlay is gone by default; it is still drawn while vision is on, because the model grounds its indices on it, and now only for as long as the capture instead of the whole model round-trip.

Cost and time

• Budget brake. Enter your own per-model USD prices (nothing is shipped or assumed) and a per-task cap. The task pauses mid-run with a continue/stop card. Unpriced models count as unknown, never as zero.
• Scheduled tasks. Daily unattended runs in their own background tab. Scheduling pre-approves the plan, sensitive actions auto-decline, and results land in history with a notification.
• A cheaper, more correct step. The page is parsed once per step instead of three to seven times — which also fixes a real bug, since every re-parse renumbered the elements and the index the model picked could resolve to a different one. History trimming, which was written but never actually called, now runs. Retries are three attempts with backoff in one place, instead of two stacked layers that could have fired 21 requests for one rate-limited step.

New abilities

• Vision escalation. When a step fails, the next state read attaches a numbered screenshot even with vision off, so the model can re-ground instead of retrying against the same blind description. Bounded by the failure budget.
• Two-way webhook. Task outcomes POST to one URL you choose (https anywhere, plain http only on localhost). The response may carry {"followUp": "next task"} and Flowkite runs it unattended — capped at three chained follow-ups, opt-in, off by default.
• Task templates. A pinned prompt with {slots}; Tab walks the slots and send is held until none are left. Fresh installs get two examples.
• "Ask Flowkite" context menu on any page or selection, opening the side panel with the composer pre-filled.
• extract_content returns to you, capped and injection-wrapped, with pipe tables rendered as real tables and a Copy CSV button.
• One task, one answer. A live status line that overwrites in place, one persisted result, and the step trail one click beneath it — instead of four separate bubbles. The line is a live region, so screen readers still get the progress.
• Tab grouping and a completion chime. Tabs the agent opens are collected into one labelled Chrome tab group, so an automated tab is obvious among your own. The chime rises on success and falls on anything that did not finish, well under system alert volume.

Fixes

• API keys are no longer written to logs.
• Ollama's CORS help link and 403 message point at docs.ollama.com/faq — the old URL 404s.
• The quick start link and the X link point at the real destinations.
• A leak in the DOM parser that added two window listeners per highlighted element, per parse, unbounded.

Install

Not on the Chrome Web Store yet.

1. Download flowkite-0.2.0.zip below and unzip it.
2. Open chrome://extensions/ and turn on Developer mode.
3. Click Load unpacked and select the unzipped folder.

Chrome or Edge. Bring an API key for OpenAI, Anthropic, Gemini, or point it at a local Ollama.

v0.1.13 — first Flowkate release

First release of Flowkate, a fork of Nanobrowser focused on keeping the user in control of what the agent does.

Install

1. Download flowkate.zip below and unzip it.
2. Open chrome://extensions/ and turn on Developer mode (top right).
3. Click Load unpacked and select the unzipped folder.

Then open the side panel, click the settings icon and add an LLM API key.

What this fork adds

You stay in control

• Plan preview — the agent shows its plan and waits for approval before touching a page.
• Confirmation for sensitive actions — buying, deleting, submitting a form, downloading a file or typing into a password field all stop and ask first. The check reads the DOM element the agent picked rather than the model's own account of what it is doing, so a page that steers the model still has to get past you.
• Undo — roll back the last step; the page navigates back and the agent forgets the step, then re-plans.

On-device memory — the agent can remember your preferences across sessions, stored in chrome.storage.local only. Options → Memory lists every stored fact with per-entry delete and a master switch.

More reliable grounding — empty DOM parses are retried before the agent concludes a page is empty, and a screenshot is attached automatically when the DOM yields nothing at all. Stale element clicks are retried once against fresh state.

Hybrid model routing — an optional cheap Fast model handles routine steps, while failures, dense pages, screenshot reasoning and the first step of a new plan escalate to your main model.

Parallel research — independent lookups run concurrently in their own tabs and merge. Those tabs are read-only by construction and cannot click, type or submit.

Notes

• Not on the Chrome Web Store yet; unpacked install only.
• Chrome and Edge are supported. Firefox and Safari are not.
• See PRIVACY.md for what is stored and what is sent to your LLM provider, and PERMISSIONS.md for why each permission is requested.