I run multiple Claude Code and OpenCode sessions at the same time — sometimes on my laptop, sometimes on remote devboxes over SSH. The friction that bugged me most was context switching: I'd be thinking through a problem, pacing around, and then have to sit down, find the right terminal window, and type out what I wanted the agent to do. I kept wishing I could just say it out loud.
So I built AgentWire. It's a self-hosted CLI + browser portal that sits on top of tmux. You open the portal on your phone or tablet, push-to-talk, and your voice gets transcribed and routed to whichever agent session you pick. The agent talks back via TTS. It works across machines — SSH into a remote box, and those sessions show up in the same portal.
The core loop: speak a prompt, agent works on it, agent speaks the result back. You can be on the couch, on a train, wherever — as long as your portal is accessible (I use a Cloudflare tunnel).
It also does multi-agent orchestration. One session can be the "leader" that spawns worker panes, delegates tasks, and collects summaries. This is useful when you want parallel work across a codebase without the agents stepping on each other.
You can run everything on one machine, or distribute it across several like I do — I have TTS running on a Windows WSL box with a GPU, several dev boxes (local and remote) running various projects, and my main local machine running the portal and STT.
Install: pip install agentwire-dev (or uv tool install agentwire-dev). No cloud, no accounts, no sign-up. Your audio stays on your machines. TTS runs on your own GPU or serverless (RunPod), STT via Whisper.
I use this daily. The docs are thin, and I'm adding tons of QoL and better error paths daily too. But the core voice workflow and session management are solid.
Curious about:
1. Is push-to-talk the right model? Or would continuous listening / wake-word work better for coding?
2. If you run multiple AI coding agents, what would you want from a coordination layer?
3. What's the first thing that would need to improve for you to actually try this?
Python, tmux, MCP. Works with Claude Code, OpenCode, or anything that runs in a terminal.
prradox•1h ago
So I built AgentWire. It's a self-hosted CLI + browser portal that sits on top of tmux. You open the portal on your phone or tablet, push-to-talk, and your voice gets transcribed and routed to whichever agent session you pick. The agent talks back via TTS. It works across machines — SSH into a remote box, and those sessions show up in the same portal.
The core loop: speak a prompt, agent works on it, agent speaks the result back. You can be on the couch, on a train, wherever — as long as your portal is accessible (I use a Cloudflare tunnel).
It also does multi-agent orchestration. One session can be the "leader" that spawns worker panes, delegates tasks, and collects summaries. This is useful when you want parallel work across a codebase without the agents stepping on each other.
You can run everything on one machine, or distribute it across several like I do — I have TTS running on a Windows WSL box with a GPU, several dev boxes (local and remote) running various projects, and my main local machine running the portal and STT.
Install: pip install agentwire-dev (or uv tool install agentwire-dev). No cloud, no accounts, no sign-up. Your audio stays on your machines. TTS runs on your own GPU or serverless (RunPod), STT via Whisper.
I use this daily. The docs are thin, and I'm adding tons of QoL and better error paths daily too. But the core voice workflow and session management are solid.
Curious about:
1. Is push-to-talk the right model? Or would continuous listening / wake-word work better for coding?
2. If you run multiple AI coding agents, what would you want from a coordination layer?
3. What's the first thing that would need to improve for you to actually try this?
Python, tmux, MCP. Works with Claude Code, OpenCode, or anything that runs in a terminal.