I tried to find five new devtools this week that had nothing to do with coding agents. I found one. So here’s the one, plus four agent tools I’d actually use.
1. Run a 35B model on a Mac in 3 GB of RAM
Edge0 streams the experts of a mixture-of-experts model from your SSD only when it needs them, so their 35B model peaks at about 3 GB of memory. They report 15 to 18 tokens per second on a Mac Mini M4 Pro, which is fast enough to actually use. The weights still take about 23 GB of disk.
2. tmux for people running five agents at once
Corgi is a terminal multiplexer that shows which agent is working, which one finished and which one needs you, and F1 jumps to the next one waiting. It works with Claude Code, Codex, Pi and OpenCode, across all your accounts.
3. Let your coding agent hear what it made
My favorit of the week is earshot. Agents can render a whole product video in code and still can't tell you the whoosh lands 80 ms after the cut. earshot gives them a CLI and MCP server that describes sounds in words, finds the best moment in a long recording and scores a film's audio, all locally on CPU.
4. An open source alternative to Grok Bot
OpenMausBot turns your agents into contacts in a chat app. Each bot is a Claude, Codex or Grok CLI running on your machine with your existing logins, and gets its own personality, model and computer (a cloud Linux desktop or a local VM).
5. See how risky a PR is before you hit merge
Farol is a free CLI and GitHub Action that runs after your AI review and tells whoever approves how risky the merge is and which files carry the risk. The score comes from static analysis (dependency graph, security and lint checks, breaking API changes), which I trust more than an LLM grading its own work.
Building a devtool, or found one I should know about? Hit reply. I'd love to hear about it.
Jonathan (@jonathimer)