agentwire-config

AgentWire Config (~/.agentwire/config.yaml)

Layout of ~/.agentwire/

File	Purpose
config.yaml	Main config (see structure below)
machines.json	Remote machines registry
scripts/	Machine-specific helper scripts (TTS management, startup, etc.)
voices/	Custom TTS voice samples
uploads/	Uploaded images for cross-machine sharing
artifacts/	Agent-generated HTML for artifact windows
wiki/	LLM-maintained knowledge base (Karpathy LLM Wiki pattern)
logs/	Audit logs for damage-control

Per-session config (posture, roles, voice) lives in .agentwire.yml in each project directory (see agentwire-project-config skill).

Machine Scripts (~/.agentwire/scripts/)

Each machine has a ~/.agentwire/scripts/ directory for machine-specific helper scripts (TTS management, startup hooks, service wrappers, etc.). This is the standard location — agents should look here first and put new scripts here.

Scripts in ~/bin/ should symlink to ~/.agentwire/scripts/ so they're callable from PATH but the source of truth is in one place.

These scripts are not managed by agentwire — they're local to each machine and not version controlled. They exist because different machines have different roles (GPU server runs TTS, Mac runs the portal, etc.) and need different glue scripts.

config.yaml Structure

server:
  host: "127.0.0.1"  # default; "0.0.0.0" allows LAN/phone access and requires the auth token (see SECURITY.md)
  port: 8765
  activity_threshold_seconds: 3  # Seconds before session considered idle
  ssl:
    cert: "~/.agentwire/cert.pem"
    key: "~/.agentwire/key.pem"
  # Auth token: unset = use ~/.agentwire/portal.token (auto-generated on first
  # non-loopback start; print/rotate with `agentwire portal token [--rotate]`).
  # Set a string to override the file; "" disables auth (loopback binds only —
  # the portal refuses to start on 0.0.0.0 with auth disabled).
  # auth_token: ""
  # Extra browser origins allowed on state-changing requests (exact
  # scheme://host[:port]). The portal's own origin and localhost always pass.
  # Needed when fronting with Cloudflare Tunnel:
  allowed_origins: []  # e.g. ["https://portal.example.com"]

projects:
  dir: "~/projects"
  worktrees:
    enabled: true
    suffix: "-worktrees"
    auto_create_branch: true
    copy_files: [".env", ".agentwire.yml", ".agentwire.tasks.yml"]   # gitignored files seeded into each new worktree
                           # (git worktree add only checks out tracked files,
                           #  so .env/secrets/local config don't carry over —
                           #  add ".env.local", ".envrc", etc. as needed).
                           # Keep .agentwire.yml/.agentwire.tasks.yml gitignored: a TRACKED
                           # copy means worktree runs use the committed version (HEAD) and
                           # silently ignore live edits — see agentwire-project-config skill.

tts:
  backend: "default"  # tier: default (in-process Kokoro, zero setup — ~200MB model
                      # auto-downloads on first portal start; speechSynthesis covers
                      # the wait) | custom (self-hosted shim at url)
  url: "http://localhost:8100"  # custom tier only — shim endpoint
  default_voice: "dotdev"
  voices_dir: "~/.agentwire/voices"  # Custom voice samples for cloning
  instructions: ""  # free-text prompt passed through to the shim
  options:  # opaque JSON passed to the shim; the bundled shim reads:
    backend: kokoro  # engine: kokoro | chatterbox | chatterbox-streaming | zonos-transformer | zonos-hybrid
  exaggeration: 0.5  # Voice expressiveness (0-1, Chatterbox)
  cfg_weight: 0.5  # CFG weight (0-1, Chatterbox)
  timeout: 60

stt:
  backend: "default"  # TIER (where transcription happens): default (portal-owned
                      # in-process Moonshine — bundled, auto-downloads on first boot,
                      # no setup; falls back to browser SpeechRecognition while it
                      # warms up or on py3.14+) | cloud (portal → hosted OpenAI-
                      # compatible transcription API, no shim daemon) | custom
                      # (self-hosted shim at url)
  engine: "auto"      # ENGINE (which model the self-hosted shim loads): auto | moonshine |
                      # whisper. Orthogonal to backend — used only by `agentwire stt start/serve`.
                      # `{backend: custom, engine: whisper}` = boot shim AND run faster-whisper.
  moonshine_model: "moonshine/base"  # moonshine engine only — ONNX model id (moonshine/tiny | moonshine/base)
  model: "base"       # whisper engine only — faster-whisper/openai-whisper model (tiny → large-v3)
  url: "http://localhost:8101"  # custom tier only — shim endpoint (also the `agentwire stt` port)
  cloud:  # cloud tier only — all fields optional, defaults shown
    base_url: "https://api.openai.com/v1"  # any OpenAI-compatible endpoint (Groq, Mistral, speaches, ...)
    model: "gpt-4o-mini-transcribe"
    api_key_env: "OPENAI_API_KEY"  # NAME of the env var holding the key — the key itself
                                   # never lives in config and never reaches the browser;
                                   # portal refuses to start if the var is unset
    language: ""  # optional ISO-639-1 hint
  timeout: 30
  silence_prepend_ms: 0  # prepend silence if your backend clips the first syllable
  instructions: ""  # free-text hint passed through to the shim
  options: {}  # opaque JSON passed to the shim (language hints, vocab biasing, ...)
  corrections: {}  # post-transcription find/replace, e.g. {"agent wire": "agentwire"}

agent:
  command: "claude --dangerously-skip-permissions"

dev:
  source_dir: "~/projects/agentwire-dev"  # agentwire source for TTS/STT venv

services:  # Where services run (for multi-machine setups)
  portal:
    machine: null  # null = local
    port: 8765
    session_name: "agentwire-portal"  # tmux session name
  tts:
    machine: "gpu-server"  # or null for local
    port: 8100
    session_name: "agentwire-tts"
  stt:
    session_name: "agentwire-stt"
  custom:  # User-defined service sessions — autostart on portal launch AND
           #   `agentwire up`, health-checked by the portal watchdog, shown in
           #   the portal's Services column. Manage with `agentwire services ...`.
           #   The notifications bridge is a built-in registry entry (override
           #   by defining a service with its name).
    - name: "agent-brain"          # tmux session name (required)
      project: "~/projects/brain"  # project dir; defaults to dev source dir
      autostart: true              # boot on portal launch / `agentwire up` (default true)
      roles: "brain"               # optional; overrides project .agentwire.yml
      posture: "bypass"            # optional; posture override
      restart: on-failure          # never | on-failure | always (watchdog respawn
                                   #   with 30s..10m exponential backoff; default on-failure;
                                   #   `agentwire services down` always sticks)
      healthcheck:                 # optional; defaults to tmux_session/60s
        kind: tmux_session         # tmux_session | http | command
        url: "http://..."          # for http (2xx = healthy)
        command: "curl -sf ..."    # for command (exit 0 = healthy)
        interval: 60               # seconds between watchdog checks
    - "simple-service"             # string shorthand = name only, all defaults

executables:  # Override executable paths (optional, auto-detected by default)
  ffmpeg: "/opt/homebrew/bin/ffmpeg"
  whisperkit-cli: "/opt/homebrew/bin/whisperkit-cli"
  hs: "/opt/homebrew/bin/hs"
  agentwire: "~/.local/bin/agentwire"

uploads:
  dir: "~/.agentwire/uploads"
  max_size_mb: 10
  cleanup_days: 7

artifacts:
  dir: "~/.agentwire/artifacts"
  max_size_mb: 10

wiki:
  dir: "~/.agentwire/wiki"           # Wiki vault location

portal:
  url: "https://localhost:8765"

channels:  # Outbound-only notifications. Only email + quo ship.
  # Keys are env-only: RESEND_API_KEY / QUO_API_KEY in ~/.agentwire/.env
  # (docs/wiki/security/secrets.md) — never in config.yaml.
  email:
    from_address: "Echo <[email protected]>"
    default_to: "[email protected]"
    banner_image_url: "https://yourdomain.com/images/banner.png"
    echo_image_url: "https://yourdomain.com/images/echo.png"
    echo_small_url: "https://yourdomain.com/images/echo-small.png"
    logo_image_url: "https://yourdomain.com/images/logo.png"
  quo:
    from_number: "+1234567890"  # E.164 or phone number ID (PNxxx)
    default_to: "+0987654321"

scheduler:
  autostart: true        # Start the scheduler daemon when the portal boots (default: true)
  dispatch_cooldown: 60  # Seconds between task dispatches (default: 60)
  dispatch_max_runtime: 14400  # Watchdog ceiling per dispatch in seconds; a hung ensure is killed and the task marked timeout (default: 4h, 0 disables)

usage_limit:             # Usage-limit recovery watchdog (docs/wiki/usage-limit-recovery.md)
  enabled: true          # Master switch for dialog detection/parking (default: true)
  exclude_sessions: []   # Session names never auto-parked (gates NEW parks only)

session_context:         # Context-bloat observability (Phase 0, observe-only — issue #442)
  warn_remaining_pct: 20 # Flag a session when its REMAINING context drops to/below this %.
                         # The Claude Code bar shows headroom, not usage, so LOW = bloated.
                         # Surfaced via `agentwire list --context` and MCP `sessions_context`.

worktree:                         # `agentwire worktree <name>` orchestration (WorktreeConfig).
                                  # Distinct from projects.worktrees above (the legacy
                                  # project/branch layout).
                                  # PRECEDENCE (#705): a project's .agentwire.yml `worktree:`
                                  # block (dir/base) overrides these for that repo; a
                                  # per-invocation --base flag beats both. Chain: flag →
                                  # project .agentwire.yml → this global block → built-ins.
  worktree_dir: ~/worktrees       # Root for worktrees, nested per project:
                                  # <worktree_dir>/<project>/<name>/ (mirrors ~/projects/)
  default_base: develop           # Base branch new worktrees fork from. OMIT to derive from
                                  # the repo's actual default branch (origin/HEAD, fallback to
                                  # current branch) — no hardcoded 'main'. --base always wins.
  default_project: ~/projects/my-repo  # Repo used when --project is omitted AND cwd isn't in a
                                  # git repo. Otherwise --project / the git root of cwd is used.
  naming: "{user}/{slug}"         # Optional branch-name template for NEW branches. Placeholders:
                                  # {name} (verbatim), {slug} (slugified), {user} (OS login).
                                  # Omit → branch == name verbatim. Only the git branch is
                                  # templated; the tmux session name stays {project}-{name}.

session:
  # No global default-role: a session's ROLE is derived from its spawn verb
  # (new → orchestrator or worker depending on branch, worktree → worker by
  # default, spawn → worker), then any --roles / .agentwire.yml roles: stack
  # on (worker) or replace (orchestrator) it. TOPOLOGY (worktree vs pane/main)
  # separately picks WHICH worker etiquette file (worker-worktree vs worker).
  # See resolve_roles.
  inject_soul: true          # Append the bundled 'soul' personality role to every human-facing
                             # session (appended last for recency weight). Headless roles
                             # (worker, task-runner, notifications) and soul/soul-* sessions
                             # are excluded automatically; per-session opt-out: --no-soul on new/dev

beta:
  # Opt-in gates for features that SHIP on main but stay off until asked for.
  # Every flag defaults to false, and "off" means ABSENT — not merely dormant:
  # a gated feature's role-prompt lines AND its MCP tool-description prose are
  # stripped before a model sees them, so a user who never enabled it pays no
  # tokens for it. Only a real YAML boolean turns one on — `voice_layer: "false"`
  # (quoted) stays OFF, deliberately. `agentwire doctor` reports each flag's
  # state either way. Config-only (no CLI verb); env override works via
  # AGENTWIRE_BETA__VOICE_LAYER=true.
  voice_layer: false         # The realtime voice buddy (`agentwire buddy`). Off: every buddy
                             # subcommand refuses, naming this key and OPENAI_API_KEY (which
                             # lives in ~/.agentwire/.env). See docs/wiki/voice-layer.md §0.