Insights
Tactical playbooks and market analysis for teams using AI browser automation.
Newsletter
Get updates on browser automation
Filter by tag
Clear filterWe forked Google's chrome-devtools-mcp to share one real Chrome across several AI agents. It crashed on memory, raced on connections, and shipped broken builds. Here is why we dropped the MCP server and rebuilt the part we needed as a small browser skill.
Claude Code is excellent. I still chose OpenCode for the coding agent at Vibe Technologies. Here are the five reasons — open-source forkability, subagent observability, remote control via opencode serve, model flexibility, and where Claude Code still wins.
OpenCode's web UI isn't mobile-friendly. Here's how Claude Code remote control over Tailscale became the interim solution for managing AI coding sessions from a phone.
Running OpenCode in server mode on a VM and exposing it via Tailscale turns it from a one-person CLI tool into a persistent coding service. The software-engineer agent on OpenClaw can then open sessions, supervise progress, and follow up — acting more like a staff engineer than a solo coder.
Practical token-optimization stack at Vibe Technologies for OpenCode coding agents — LST for log compression, RTK (Rust Token Killer) as a CLI proxy, Caveman mode for terse output, and orchestrator + subagent role split that keeps expensive models doing only expensive work.
Running opencode in serve mode across 10+ repos daily. We kept hitting the same UX gaps and reliability bugs. At some point we stopped filing issues and started shipping patches.
A deep dive into fixing opencode for production use as a long-running coding agent: MCP memory leaks, tool deadlocks, session recovery, and why Vibe Browser's extension architecture avoids the chrome-devtools-mcp memory trap entirely.
How Vibe Technologies structures its AI coding stack: starting with Claude Code, switching to OpenCode for multi-model flexibility, setting up orchestrator and subagent roles, and running parallel agents 24/7.