Now is the Best Time to be a Duct Tape Engineer | by Adam Derewecki | Jun, 2026 | MediumSitemapOpen in appSign up<br>Sign in
Medium Logo
Get app<br>Write
Search
Sign up<br>Sign in
Now is the Best Time to be a Duct Tape Engineer
Adam Derewecki
3 min read·<br>Just now
Listen
Share
Press enter or click to view image in full size
I have two small kids, which means my free time is largely governed by school schedules. The gaps I actually get to think about code, projects, anything, are often in the car. Driving to pickup, driving to dropoff, the fifteen minutes between.<br>I already had OpenAI and Claude’s advanced voice modes on my phone. I’d use them on drives. But I could never actually do anything. I could talk through a problem, think out loud, get some ideas, and then park the car with nothing shipped, nothing changed, nothing that would still be there when I sat back down at a keyboard hours later.<br>What I wanted was to say “hey Siri, call Claw Phone” and have the audio system in my Toyota become an IDE.<br>So I built it. You call a number, your AI agent picks up with live access to your Gmail, Google Calendar, and any other data you’ve wired in. Ask it whether someone replied to your email. Ask it to draft a message. Ask it to look something up in your notes. The stack is three things glued together: Twilio carries the audio, OpenAI’s Realtime API handles the voice conversation, and the Realtime model has exactly one tool: ask_claude(query). That routes your question to a Claude Code subprocess running on my homelab NUC, with MCP servers already warm and connected to my actual accounts.
Claude Code has write access to the codebase it’s running in. That was always the design. The agent that answers your call can edit its own source files. Pair that with a supervisor loop that restarts the server on exit, and the full deployment workflow becomes: edit some code, push a commit, call the number, say “deploy yourself.” The server exits cleanly and comes back on the new version. I’ve been using it on school runs.<br>— — —<br>I didn’t build Twilio. I didn’t build the Realtime API. I didn’t build Claude Code or MCP. What I did was see that if you connected them in a specific way, you’d get something that solved my exact problem. That’s always been the duct tape engineer’s skill: pattern-matching across components, seeing the shape of a solution before you’ve written a line. What’s different now is that the components are extraordinary. Each one of those pieces was, five years ago, a company.<br>The gap between having the idea and having it working has basically collapsed. The Twilio integration, the WebSocket plumbing, the subprocess management, the MCP wiring, none of it felt hard. I kept a Claude window open the whole time and it kept handing me exactly the right next piece. I had it running in an afternoon.<br>ClawPhone is open source at github.com/derwiki/clawphone. You need a Twilio number, an OpenAI key, an Anthropic key, and a machine to leave running. The README has the full setup.
AI
Software Engineering
Programming
Productivity
Side Project
Written by Adam Derewecki<br>200 followers<br>·22 following
Hi! I’m Adam. I live in San Francisco, write code, take pictures, and practice yoga.
Help
Status
About
Careers
Press
Blog
Privacy
Rules
Terms
Text to speech