agentwire-dev 1.55.0

Multi-session voice web interface for AI coding agents

A self-hosted, keyboard-driven cockpit for running a whole fleet of Claude Code agents at once.

From a wall of tmux panes to one cockpit. Worktrees, command palette, scheduler, voice — every layer stacks, so you ship far more than you ever could hand-juggling terminals. Your machine, your keys, no telemetry.

---

Why AgentWire

Running several Claude Code agents by hand means a wall of tmux panes: constant context-switching, copy-pasting between sessions, tracking which branch each agent sits on, babysitting every one. It doesn't scale past one or two before your attention — not the model — becomes the bottleneck. AgentWire turns that wall into one cockpit, and its capabilities compound:

• It compounds. Worktrees + command palette + quick-action keys + tab-switching
  • scheduler + voice each stack on the others to multiply how much one person can run at once. That's the whole point — not any single feature.
• Orchestrate many agents, not watch one. Every session is a window; F3 fans them into a Mission-Control collage; Tab cycles between them; Cmd/Ctrl+K runs any action. Council deliberates, briefing mode fans out, worktree sessions isolate parallel work — tmux- and Claude-Code-native from the ground up.
• Autonomous and unattended. A scheduler runs recurring jobs (research, test runs, cleanup, doc-drift); gates verify real work happened; repo tasks open draft PRs, not commits to main; reliability plumbing (verified delivery, dead-letter, email-on-fail, usage-limit recovery) means nothing silently fails.
• Your machine, your keys, no telemetry. Runs entirely on your hardware, loopback by default. No cloud account, no third party in the loop — your code and your API keys never leave the machine. 300+ damage-control blocks make running unattended sane.

Use the official app to watch one session; use AgentWire to command a whole fleet — on your own hardware, by keyboard, with voice layered on top.

---

What It Does

A self-hosted, desktop-style portal where every agent session is a window — reachable from any device on your LAN, running on your own hardware with your own keys.

You → AgentWire Portal → tmux sessions → a fleet of Claude Code agents
 ⌨️🎤      (WebSocket)         📺                    🤖🤖🤖

From your laptop, phone, or tablet on your network:

• Every session is a draggable window; F3 collage, Tab to cycle
• Command palette (Cmd/Ctrl+K) runs any action; quick-action keys throughout
• Branch a session into an isolated git worktree for parallel work
• Put repeat work on a scheduler; hand hard calls to a council
• Push-to-talk in, spoken summaries back — drive sessions by ear while off-tab

The outcome: the whole fleet feels like one machine. You go from babysitting one session to commanding many.

---

Quick Start

# Install (isolated tool env — works on any Python, including Homebrew/Debian PEP 668 setups)
uv tool install agentwire-dev

# Setup (interactive)
agentwire init

# Run
agentwire portal start
# Open http://127.0.0.1:8765 in Chrome — your cockpit is live (voice works immediately too)

# Start your first agent session
agentwire new -s hello -p ~/projects/hello

Optional: clone this repo (git clone https://github.com/dotdevdotdev/agentwire-dev ~/projects/agentwire-dev) and agentwire dev opens a guided helper session inside it — setup walkthrough, project wiring, issue filing, forking. It requires the source checkout; everything above works without one.

Requirements: Python 3.10+, tmux, Claude Code. Optional: ffmpeg (only for host-mic push-to-talk capture and voice cloning — browser voice input works without it).

Honest setup time: under a minute to a working voice portal with a genuinely good voice — Kokoro-82M runs on CPU out of the box (one-time ~200 MB model download in the background; the browser voice covers the wait). ~15 minutes for the full experience: cloned voices via a self-hosted TTS shim, Whisper-grade transcription, phone-from-anywhere (certs + token).

Phone / LAN Access

The portal binds to loopback (127.0.0.1) by default. To access the portal from your phone, tablet, or other devices on your local network:

1. Generate SSL certificates (required for microphone access over non-loopback connections):

   agentwire generate-certs
   

2. Enable LAN access: set server.host: 0.0.0.0 in ~/.agentwire/config.yaml.
3. Get your auth token: non-loopback connections require a bearer token. Print it with:

   agentwire portal token
   

4. Connect: Open https://<your-machine-ip>:8765 on your phone and enter the token when prompted.

Origin checks reject cross-site browser requests on every bind. Keep the portal on a trusted LAN — never port-forward it or run it on a public-facing VPS. For internet access, use Cloudflare Tunnel + Zero Trust. See SECURITY.md for details.

Platform-specific instructions

macOS:

brew install tmux uv        # ffmpeg optional: only for host-mic PTT + voice cloning
uv tool install agentwire-dev

Bare pip install agentwire-dev fails on Homebrew Python 3.12+ with error: externally-managed-environment (PEP 668). uv tool install (or pipx install agentwire-dev) sidesteps that by installing into its own isolated environment.

Ubuntu/Debian:

sudo apt install tmux       # ffmpeg optional: only for host-mic PTT + voice cloning
curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install agentwire-dev

pipx install agentwire-dev works everywhere too. Prefer plain pip? Use a venv: python3 -m venv ~/.agentwire-venv && source ~/.agentwire-venv/bin/activate && pip install agentwire-dev.

WSL2: Same as Ubuntu. Audio is limited; use as remote worker with portal on Windows host.

tmux config matters. Default tmux has no mouse scroll, a tiny scrollback, and broken copy UX — see Recommended tmux config, or let agentwire init install it for you.

---

Features

The cockpit, then the layers that stack on it:

Feature	Description
Desktop UI	Every session is a draggable window; F3 collage grid, Tab to cycle, Cmd/Ctrl+K command palette
Multi-Session	Run multiple agents on different projects simultaneously
Git Worktrees	Same project, multiple branches, parallel agents — each in an isolated worktree session
Worker Orchestration	Spawn worker panes, coordinate tasks, auto-reap when idle
Council	Fan a prompt to multiple "soul" sessions (brain, conscience, critic…), synthesize with attribution
Scheduler	Recurring autonomous tasks with gates, priorities, and usage-limit recovery — draft PRs, not commits to main
Briefing Mode	A terse human-facing anchor fans out verbose worktree correspondents, then briefs you on cue
Voice	Push-to-talk in, neural TTS out on CPU (Kokoro); voice cloning via a self-hosted GPU shim; global macOS hotkeys dictate to a background session from any app
Remote Machines	SSH into GPU servers and talk to agents there
Safety Hooks	300+ dangerous commands blocked (rm -rf, force push, etc.), all logged
Outbound Channels	Email (Resend) + SMS (Quo / OpenPhone) for cross-device notifications
Session Roles	Leader/worker patterns for multi-agent workflows

---

How It Works

1. Create a session:

agentwire new -s myproject -p ~/projects/myproject

2. Open the portal: Visit http://127.0.0.1:8765 in Chrome (or your phone/tablet with LAN access configured). Each session is a window — open as many as you want and Tab between them, or F3 for the collage grid.

3. Drive it by keyboard: Cmd/Ctrl+K runs any action; quick-action keys send to the focused session; branch a session into a worktree for parallel work; put repeat jobs on the scheduler.

4. Add voice on top: Hold the mic to speak (in instant mode the transcript appears for a quick glance — Enter sends it), and agent responses are spoken back — Kokoro neural voice out of the box, so you can drive sessions and hear summaries by ear while your eyes are elsewhere.

---

Multi-Agent Orchestration

AgentWire supports orchestrator/worker patterns for complex tasks:

# .agentwire.yml in your project (keep it gitignored — it's personal config,
# and tracked copies break worktree dispatch; agentwire adds it to .gitignore for you)
type: claude-bypass
roles:
  - agentwire
  - voice

Sessions can spawn workers:

agentwire spawn --roles worker  # Creates a worker pane
agentwire send --pane 1 "Implement the auth module"

Workers execute tasks autonomously while the orchestrator coordinates.

---

Safety

AgentWire blocks dangerous operations before they execute:

• rm -rf /, git push --force, git reset --hard
• Cloud CLI destructive ops (AWS, GCP, Firebase, Vercel)
• Database drops, Redis flushes, container nukes
• Sensitive file access (.env, SSH keys, credentials)

agentwire safety check "rm -rf /"
# → ✗ BLOCKED: rm with recursive or force flags

agentwire safety status
# → 312 patterns loaded, 47 blocks today

All decisions logged for audit trails.

---

Voice Configuration

Two tiers, both sides:

default (zero setup, what a fresh install gets): Chrome speech recognition in, Kokoro-82M out — a genuinely good neural voice (top of the TTS Arena at 82M params), 32 preset voices across 8 languages, pure CPU, identical on every OS. The model (~200 MB) downloads in the background on first portal start; browser speechSynthesis covers speech until it's ready (and remains the last-resort fallback). No GPU, no certs, no commands.

custom (bring your own model): any HTTP shim implementing the voice shim contract — ~30 lines wraps anything (Deepgram, whisper.cpp, an expressive emotion-tag model). Voice cloning, GPU engines, emotion control live here. The bundled servers are reference shims:

# ~/.agentwire/config.yaml
tts:
  backend: "custom"
  url: "http://localhost:8100"     # agentwire tts start (kokoro CPU / chatterbox GPU / zonos)
  options:
    backend: kokoro
stt:
  backend: "custom"
  url: "http://localhost:8101"     # agentwire stt start (moonshine ONNX, CPU)

Shims can declare capabilities (emotion tags, style instructions) via GET /capabilities — agentwire injects the shim's tool_prompt into the agent's say tooldef so agents actually use them.

Prefer text-only?

Instant mode already needs nothing — just don't press the mic. Agent speech plays through the browser; mute the tab (or close it — with no browser connected, speech plays on local speakers, which you can silence at the system level).

---

CLI Reference

Session Management

agentwire list                    # List sessions
agentwire new -s <name> -p <path> # Create session
agentwire kill -s <name>          # Kill session
agentwire send -s <name> "prompt" # Send to session
agentwire output -s <name>        # Read output

Worker Panes

agentwire spawn --roles worker    # Spawn worker in current session
agentwire send --pane 1 "task"    # Send to worker
agentwire output --pane 1         # Read worker output
agentwire kill --pane 1           # Kill worker

Voice Commands

agentwire say "Hello"             # TTS (auto-routes to browser)
agentwire send -s NAME "Done"     # Inject text into a session
agentwire listen start/stop       # Voice recording
agentwire voiceclone list         # Custom voices

Remote Machines

agentwire machine add gpu --host 10.0.0.5 --user dev
agentwire new -s ml@gpu           # Create session on remote
agentwire tunnels up              # SSH tunnels for services

Safety & Diagnostics

agentwire doctor                  # Auto-diagnose issues
agentwire safety status           # Check protection status
agentwire hooks install           # Install Claude Code hooks
agentwire network status          # Service health check

---

Documentation

Full reference: docs/wiki/INDEX.md

Quick links:

• Troubleshooting
• Portal API
• Remote Machines
• Voice Shim Contract · Self-Hosted TTS
• Safety Hooks

---

Community

• Issues - Bug reports
• Website - Docs and demos

---

License

AgentWire is free and open source under the Apache License 2.0 — your machine, your keys, no telemetry.

Free-forever pledge: the tool is free forever. No paid tier, no open-core, no feature paywall. Commercial training & enablement for teams is offered separately by dotdev — the tool itself is never gated.

Contributions are accepted under the Developer Certificate of Origin (a Signed-off-by line), not a CLA.

---

AgentWire: For people who have better things to do.

---

Maintained by dotdev. Reach out via email at [e-mail] or on GitHub @dotdevdotdev.