Saved tasks, Memory Usage Monitoring, Retry Limits, Voice mode

Users can now save tasks/chats and rerun them with different inputs.
Voice mode is now available and users can also monitor their memory usage.

new stuff

• Save as a task: Save a useful conversation as a reusable task. Instructions can be updated and be given different inputs.
• Voice input: Click the microphone and speak. Users can choose between Browser dictation, Groq, OpenRouter, Deepgram, or ElevenLabs.
• Background notifications: Agents can notify users while working on the background tasks when they need an answer or approval.
• Memory benchmark:pnpm benchmark:memory to measure Chrome and DSH memory usage while agents work across multiple tabs parallelly.
• Multiple Chrome profiles: agents can run across different Chrome profiles at the same time using one DSH instance.

small improvements

• Added retry limits so the agent doesn't keep repeating the same failed action.
• Improved handling of changed pages and stale element references.
• Improved discovery of controls inside shadow DOM.
• Fixed saved chats returning to their tabs and reopening closed tabs.
• Made saved-task generation more reliable when the model returns an incomplete response.
• Made provider errors clearer, including OpenRouter spending-limit errors.
• Fixed Markdown tables and kept the chat header visible while scrolling.
• Made screenshots easier to drag into the message box.
• Renamed /actions to /tools and improved the slash-command and theme pickers.

updating

Stop DSH, run the installer again, and reload the extension at chrome://extensions.
If you use multiple Chrome profiles, reload it in each one, then start DSH again.

Your settings are kept.

v0.1.1 - Meet dsh Browser Agent

The first release of dsh Browser Agent is here!

I wanted an agent that could sit beside the page I’m on, understand what I’m looking at, and help me get things done in the browser.
Something I could talk to, watch as it works, and take over from whenever I need to.

This is the first version of that idea, built around Chrome’s side panel and powered by DeepSeek Harness.

What you can do

• Ask it to explain a page, follow links, fill out forms, or find something on a website.
• Work with the tabs you already have open, including sites where you’re signed in.
• Keep separate conversations for different tabs and come back to saved chats.
• Attach images or documents to give it more context.
• Choose which browser tools it can use, with approval prompts on by default for clicks, typing, and navigation.
• Make yourself at home with light and dark themes.

Give it a try

You’ll need Chrome, Node.js 22.19+ or 24+, and an API key for your chosen model provider.

Follow the setup instructions, open the side panel, and start with something simple: “Summarize this page.”

DeepSeek Harness runs on your computer.
Messages and the page content the agent uses go to your chosen model provider.

A starting point

There’s a lot more I want to do with this, but I’m excited to put this first version into people’s hands.

If you try it, I’d love to hear what you used it for, where it helped, and where it got in your way.
Open an issue with a bug, an idea, or something you wish it could do.

Thanks for giving it a place in your browser.