Config Schema

Triggerfish is configured through triggerfish.yaml, located at ~/.triggerfish/triggerfish.yaml after running triggerfish dive. This page documents every configuration section.

Secret References Any string value in this file can use the secret:

prefix to reference a credential stored in the OS keychain. For example, apiKey: "secret:provider:anthropic:apiKey" resolves the value from the keychain at startup. See Secrets Management for details. :::

Full Annotated Example

# =============================================================================
# triggerfish.yaml -- Complete configuration reference
# =============================================================================

# ---------------------------------------------------------------------------
# Models: LLM provider configuration and failover
# ---------------------------------------------------------------------------
models:
  # The primary model used for agent completions
  primary:
    provider: anthropic
    model: claude-sonnet-4-5

  # Optional: separate vision model for image description
  # When the primary model doesn't support vision, images are automatically
  # described by this model before reaching the primary.
  # vision: glm-4.5v

  # Streaming responses (default: true)
  # streaming: true

  # Provider-specific configuration
  # API keys are referenced via secret: syntax and resolved from the OS keychain.
  # Run `triggerfish dive` or `triggerfish config migrate-secrets` to set up.
  providers:
    anthropic:
      model: claude-sonnet-4-5
      # apiKey: "secret:provider:anthropic:apiKey"

    openai:
      model: gpt-4o

    google:
      model: gemini-pro

    ollama:
      model: llama3
      endpoint: "http://localhost:11434"

    lmstudio:
      model: lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF
      endpoint: "http://localhost:1234"

    openrouter:
      model: anthropic/claude-sonnet-4-5

    zenmux:
      model: openai/gpt-5

    zai:
      model: glm-4.7

  # Ordered failover chain -- tried in sequence when primary fails
  failover:
    - claude-haiku-4-5 # First fallback
    - gpt-4o # Second fallback
    - ollama/llama3 # Local fallback (no internet required)

  # Failover behaviour
  failover_config:
    max_retries: 3 # Retries per provider before moving to next
    retry_delay_ms: 1000 # Delay between retries
    conditions: # What triggers failover
      - rate_limited # Provider returned 429
      - server_error # Provider returned 5xx
      - timeout # Request exceeded timeout

# ---------------------------------------------------------------------------
# Logging: Structured log output
# ---------------------------------------------------------------------------
logging:
  level: normal # quiet | normal | verbose | debug

# ---------------------------------------------------------------------------
# Channels: Messaging platform connections
# ---------------------------------------------------------------------------
# Secrets (bot tokens, API keys, passwords) are stored in the OS keychain.
# Run `triggerfish config add-channel <name>` to enter them securely.
# Only non-secret configuration appears here.
channels:
  telegram:
    ownerId: 123456789 # Your Telegram numeric user ID
    classification: INTERNAL # Default: INTERNAL

  signal:
    endpoint: "tcp://127.0.0.1:7583" # signal-cli daemon endpoint
    account: "+14155552671" # Your Signal phone number (E.164)
    classification: PUBLIC # Default: PUBLIC
    defaultGroupMode: mentioned-only # always | mentioned-only | owner-only
    groups:
      "group-id-here":
        mode: always
        classification: INTERNAL

  slack:
    classification: PUBLIC # Default: PUBLIC

  discord:
    ownerId: "your-discord-user-id" # Your Discord user ID
    classification: PUBLIC # Default: PUBLIC

  whatsapp:
    phoneNumberId: "your-phone-number-id" # From Meta Business Dashboard
    classification: PUBLIC # Default: PUBLIC

  webchat:
    port: 8765 # WebSocket port for web client
    classification: PUBLIC # Default: PUBLIC (visitors)

  email:
    smtpApiUrl: "https://api.sendgrid.com/v3/mail/send"
    imapHost: "imap.gmail.com"
    imapPort: 993
    imapUser: "you@gmail.com"
    fromAddress: "bot@example.com"
    ownerEmail: "you@gmail.com"
    classification: CONFIDENTIAL # Default: CONFIDENTIAL

# ---------------------------------------------------------------------------
# Classification: Data sensitivity model
# ---------------------------------------------------------------------------
classification:
  mode: personal # "personal" or "enterprise" (coming soon)
# Levels: RESTRICTED > CONFIDENTIAL > INTERNAL > PUBLIC

# ---------------------------------------------------------------------------
# Policy: Custom enforcement rules (enterprise escape hatch)
# ---------------------------------------------------------------------------
policy:
  rules:
    - id: block-external-pii
      hook: PRE_OUTPUT
      priority: 100
      conditions:
        - type: recipient_is
          value: EXTERNAL
        - type: content_matches
          pattern: "\\b\\d{3}-\\d{2}-\\d{4}\\b" # SSN pattern
      action: REDACT
      message: "PII redacted for external recipient"

    - id: rate-limit-browser
      hook: PRE_TOOL_CALL
      priority: 50
      conditions:
        - type: tool_name
          value: browser
        - type: rate_exceeds
          value: 10/minute
      action: BLOCK
      message: "Browser tool rate limit exceeded"

# ---------------------------------------------------------------------------
# MCP Servers: External tool servers
# ---------------------------------------------------------------------------
mcp_servers:
  filesystem:
    command: "deno"
    args: ["run", "--allow-read", "--allow-write", "mcp-filesystem-server.ts"]
    classification: INTERNAL

  github:
    command: "npx"
    args: ["-y", "@modelcontextprotocol/server-github"]
    classification: CONFIDENTIAL

# ---------------------------------------------------------------------------
# Scheduler: Cron jobs and triggers
# ---------------------------------------------------------------------------
scheduler:
  cron:
    jobs:
      - id: morning-briefing
        schedule: "0 7 * * *" # 7 AM daily
        task: "Prepare morning briefing with calendar, unread emails, and weather"
        channel: telegram
        classification: INTERNAL

      - id: pipeline-check
        schedule: "0 */4 * * *" # Every 4 hours
        task: "Check Salesforce pipeline for changes and notify if significant"
        channel: slack
        classification: CONFIDENTIAL

      - id: pr-review-check
        schedule: "*/15 * * * *" # Every 15 minutes
        task: "Check open PR tracking files and query GitHub for new reviews"
        classification: INTERNAL

  trigger:
    interval: 30m # Check every 30 minutes
    classification: INTERNAL # Max taint ceiling for triggers
    quiet_hours: "22:00-07:00" # Suppress during these hours

# ---------------------------------------------------------------------------
# Notifications: Delivery preferences
# ---------------------------------------------------------------------------
notifications:
  preferred_channel: telegram # Default delivery channel
  quiet_hours: "22:00-07:00" # Suppress normal/low priority
  batch_interval: 15m # Batch low-priority notifications

# ---------------------------------------------------------------------------
# Agents: Multi-agent routing (optional)
# ---------------------------------------------------------------------------
agents:
  default: personal # Fallback agent
  list:
    - id: personal
      name: "Personal Assistant"
      channels: [whatsapp, telegram]
      tools:
        profile: "full"
      model: claude-opus-4-5
      classification_ceiling: INTERNAL

    - id: work
      name: "Work Assistant"
      channels: [slack, email]
      tools:
        profile: "coding"
        allow: [browser, github]
      model: claude-sonnet-4-5
      classification_ceiling: CONFIDENTIAL

# ---------------------------------------------------------------------------
# Voice: Speech configuration (optional)
# ---------------------------------------------------------------------------
voice:
  stt:
    provider: whisper # whisper | deepgram | openai
    model: base # Whisper model size
  tts:
    provider: elevenlabs # elevenlabs | openai | system
    voice_id: "your-voice-id"
  wake_word: "triggerfish"
  push_to_talk:
    shortcut: "Ctrl+Space"

# ---------------------------------------------------------------------------
# Webhooks: Inbound event endpoints (optional)
# ---------------------------------------------------------------------------
webhooks:
  endpoints:
    - id: github
      path: /webhook/github
      # Webhook secret is stored in the OS keychain
      classification: INTERNAL
      actions:
        - event: "pull_request.opened"
          task: "Review PR and post summary"
        - event: "pull_request_review"
          task: "A PR review was submitted. Read tracking file, address feedback, commit, push."
        - event: "pull_request_review_comment"
          task: "An inline review comment was posted. Read tracking file, address comment."
        - event: "issue_comment"
          task: "A comment was posted on a PR. If tracked, address feedback."
        - event: "pull_request.closed"
          task: "PR closed or merged. Clean up branches and archive tracking file."
        - event: "issues.opened"
          task: "Triage new issue"

# ---------------------------------------------------------------------------
# GitHub: GitHub integration settings (optional)
# ---------------------------------------------------------------------------
github:
  auto_merge: false # Default: false. Set true to auto-merge approved PRs.

# ---------------------------------------------------------------------------
# Groups: Group chat behaviour (optional)
# ---------------------------------------------------------------------------
groups:
  default_behavior: "mentioned-only"
  overrides:
    - channel: slack
      channel_name: "#ai-assistant"
      behavior: "always"

# ---------------------------------------------------------------------------
# Remote: Remote access (optional)
# ---------------------------------------------------------------------------
# ---------------------------------------------------------------------------
# Web: Search and fetch configuration
# ---------------------------------------------------------------------------
web:
  search:
    provider: brave # Search backend (brave is the default)
# API key is stored in the OS keychain

# ---------------------------------------------------------------------------
# Remote: Remote access (optional)
# ---------------------------------------------------------------------------
remote:
  tailscale:
    serve: true
    funnel:
      enabled: true
      paths: ["/webhook/*"]
  auth:
# Auth token is stored in the OS keychain

Section Reference

models

Key	Type	Description
primary	object	Primary model reference with provider and model fields
primary.provider	string	Provider name (anthropic, openai, google, ollama, lmstudio, openrouter, zenmux, zai)
primary.model	string	Model identifier used for agent completions
vision	string	Optional vision model for automatic image description (see Image and Vision)
streaming	boolean	Enable streaming responses (default: true)
providers	object	Provider-specific configuration (see below)
failover	string[]	Ordered list of fallback models
failover_config.max_retries	number	Retries per provider before failover
failover_config.retry_delay_ms	number	Delay between retries in milliseconds
failover_config.conditions	string[]	Conditions that trigger failover

channels

Each channel key is the channel type. All channel types support a classification field to override the default classification level.

All secrets (tokens, API keys, passwords) are stored in the OS

keychain, not in this file. Run triggerfish config add-channel <name> to enter credentials securely. :::

classification

Key	Type	Description
mode	"personal" or "enterprise"	Deployment mode (coming soon — currently both use the same classification levels)

policy

Custom rules evaluated during hook execution. Each rule specifies a hook type, priority, conditions, and action. Higher priority numbers are evaluated first.

mcp_servers

External MCP tool servers. Each server specifies a command to launch it, optional environment variables, a classification level, and per-tool permissions.

scheduler

Cron job definitions and trigger timing. See Cron and Triggers for details.

notifications

Notification delivery preferences. See Notifications for details.

web

Key	Type	Description
web.search.provider	string	Search backend for web_search tool (currently: brave)

See Web Search and Fetch for details.

logging

Key	Type	Default	Description
level	string	"normal"	Log verbosity: quiet (errors only), normal (info), verbose (debug), debug (trace)

See Structured Logging for details on log output and file rotation.

github

Key	Type	Default	Description
auto_merge	boolean	false	When true, the agent auto-merges PRs after receiving an approving review. When false (default), the agent notifies the owner and waits for an explicit merge instruction.

See the GitHub Integration guide for full setup instructions.