Tobias Strebitzer (@tstrebitzer) header
Tobias Strebitzer (@tstrebitzer) profile picture

Entrepre-nerd and software engineer trying to make sense of all this AI stuff.

indonesia
Joined February 2011
I forked Chromium to fix my browser automation headaches with Claude Code. It's called Shuttle: every agent gets its own identity, its own logins and its own tab group. The MCP server is built in. Free and open source: shuttlebrowser.com
9
3
10
8,043
Tobias Strebitzer retweeted
Introducing Vitals CLI  The Activity Monitor alternative that thinks in apps, not processes. Now as a command. > brew install --cask foredotstudio/tap/vitals $15 lifetime, only 13 left → 3 days free 👉 vitalsmac.com
48
33
16
86
6,185
I forked Chromium with Claude Code so my agents can use a real browser. It's called Shuttle: every agent gets its own identity, its own logins and its own tab group in your window. The MCP server is built in. Free and open source (BSD-3).
5
5
291
Small MCP surface: 12 browser tools + 24 page tools, powered by @vercel agent-browser (bundled). Per-session CDP endpoints for Playwright, for scripted e2e tests. Removed from Chromium: Google sign-in, sync, telemetry, Google AI, translate, password manager and more.
1
1
101
The MCP server is built in (HTTP, loopback-only by default). No middleman. No debugging dialog, no automation bar, navigator.webdriver is false, and no window pops up until an agent needs one.
1
54
So I forked Chromium and moved the agents in. Every agent is a named identity (name, color, avatar) with its own storage. Mine log in once and stay logged in. Each one works in its own color-coded tab group. Biggest UX win for me: I can see what every agent is doing.
1
39
- Agents taking over each other's tabs on a shared profile - Every login gone on close - 270 MB per chrome-devtools-mcp instance before Chrome even starts. Ten agents: ~2.7 GB - Cloud browsers: unreliable, expensive, blocked on datacenter IPs
1
42
- Throwaway browser instances and profiles piling up - "Allow remote debugging?" on every connect - "Chrome is being controlled by automated test software" on every page - A dozen windows after an hour, and no idea which tab belonged to which agent
1
63
I use Claude Code for a lot of browser work: tools that were never built for agents, smoke-testing my own work, searching the web for things I care about. Things kept going wrong 👇
2
100
Tobias Strebitzer retweeted
No matter how good AI gets at copywriting, design, or video… Handing it to someone with no directing skill is like giving a junior a team of world-class talent and zero brief. They’ll just produce tons of mediocre results faster.
4
1
1
68
6/ Turns out what separates a lattice from radial art is position, not contrast. Measuring transition concentration per axis: synthetic checkerboards 1.000, real radial art 0.211 to 0.313. The roses now pass first try, cutting three API calls to one.
1
11
5/ It got me a prototype quickly. It did not survive a real asset set: a brass compass rose failed my guard three times per variant, convicted of being a painted checkerboard by its own alternating facets, and shipped a drop shadow fused into the alpha.
1
8
4/ Where this came from: @jidefr wrote the approach up for nano banana in December 2025, and his article is what got me started. The math under it is much older, Smith and Blinn's difference matting from SIGGRAPH 1996.
1
1
21
3/ So go around the model. Render on white, edit only the background to black, then solve: a = 1 - (white - black) F = black / a Exact matte, no threshold, no despill, no halo.
1
10
2/ The reflex is a background remover. Threshold, chroma key, matting network. They are all guessing, and they guess on the same pixels every time: glass, smoke, hair, glow, antialiased edges. Everything that is genuinely semi-transparent.
1
26
1/ Image models cannot emit an alpha channel. Ask one for a transparent background and it paints the Photoshop checkerboard straight into the RGB pixels. Here is that answer sliding across a coloured surface, fake checkerboard and all.
1
1
4
138
8/ Warning: Claude Code and the Agent SDK run real processes on your infrastructure, and - unless hardened diligently - have full access to your filesystem, credentials, OS environment and internal services. How far would you trust a setup like this.
48
7/ Claude Worker also allows for setting up multiple profiles, so each user can have their own settings, memory and tool access. This was a strong requirement for my team operating agents across varying business functions.
1
21