Open source2026Author

Continuum Extension

Browser capture surface for agent context

Pull context out of a PR, an AI chat or an error page and route it to an agent, with credentials stripped before it ever leaves the tab.

Problem

Most of the context an agent needs is in a browser tab, the PR discussion, the error page, the chat where the approach was worked out. Copying it by hand is the reason it never gets recorded.

But a capture tool that scrapes pages is also a credential exfiltration tool if it is careless, so the redaction has to happen before anything leaves the page rather than server-side afterwards.

What I built

A capture surface with the security boundary at the page, not the backend.

Captures from GitHub PRs and issues, AI chats, docs and error pages
Secret redaction for AWS, GitHub and OpenAI credentials before anything leaves the page
Per-site permissions and private-tab blocking
react-flow task graph with pause, resume and reject on individual steps
MV3 content scripts over a Python native-messaging host, bridging to the local CLI on 127.0.0.1:7357

Capture, redacted first

The order matters. Redaction happens in the page, before the native-messaging hop, so a credential never reaches the host process at all.

Something worth keeping is on screen
Is this site allowed, and is the tab private?
Private tab, or site not permitted, capture refused
Allowed, capture proceeds
Redact in-page
AWS, GitHub and OpenAI credential patterns stripped before the message is constructed.
Native-messaging host
Local only, 127.0.0.1:7357, no remote endpoint.
Recorded into .continuum/

Limitations and failure modes

Redaction is pattern-based. It covers the credential formats it knows and cannot promise to catch a secret that looks like ordinary text.

Attribution

Built by me: The extension and the native-messaging host.

Continuum
Open-source memory for AI coding agents