Hi, I am the developer of Murshid and would appreciate technical feedback. Murshid is a Chromium-based, open-source browser licensed under GPL-3.0-or-later. It is currently a 0.1.0 development preview.
Structure and workflow
WebBrain runs in a browser side panel. It supports questions about pages, selected text and PDFs, and uses reading, clicking and typing for browser tasks. The “Plan before Act” flow shows the steps first so the user can approve, change or stop the plan. A Companion component also provides a locally running endpoint.
The design question I want to discuss is the boundary between approving a plan and approving one consequential action. Summarizing a page, filling in a form and submitting that form should not necessarily share the same permission level. I want to make that distinction clear without overwhelming users with approval prompts. These are design questions for feedback, not a claim that the current preview handles every case perfectly.
What would you recommend?
- After a plan is approved, which actions should require another explicit confirmation? Form submission, file uploads and account settings are especially interesting cases.
- Where should users see which tab and content the assistant can currently access? Are permissions easier to understand per tab or per task?
- If a task stops halfway through, what should an activity history show to distinguish completed actions from actions that never happened?
As this is an early preview, a separate test profile is more appropriate than a daily work account. The Windows build is not code-signed and the macOS build is not notarized yet. I would love to hear concrete examples and criticism from people building browser automation or human-approved agent interfaces.
Cevaplar
1 yanıt