Continued door hal25

Continued puts an AI agent in your browser sidebar and lets you decide which model it talks to: a local Ollama server on your own machine (no key, no account), or OpenAI, Google Gemini, Anthropic or the Helmholtz Blablador gateway with your own API key. There is no Continued backend and no telemetry; page content goes only to the provider you choose.

What it can do
- Answer questions about the page you're on, search it, follow its links, and fetch other URLs without opening tabs.
- Work across tabs: compare several open pages, list them, open or navigate them.
- Fill fields and click buttons, always after asking you first. Password fields are never touched.

Four modes, fixed permissions
- Chat: plain conversation about the current page.
- Agent: reads automatically; anything that changes a page asks for approval.
- Auto-Edit: opens and navigates on its own; clicks and form fills still ask.
- Plan: writes a step-by-step plan first, then executes it with one click.

Harness builder
Design your own workflows on a visual canvas: steps run in sequence, in parallel lanes, or once per item, each with retry, timeout and "run only if" rules. The model is only consulted where you put an LLM step and is always told which earlier steps failed. Four built-ins to start from: summarise this page, compare open tabs, research a list of URLs, answer a question from this page.

Open source (MIT): https://github.com/ataulhaleem/continued. Also available as a VS Code extension.