{% extends "base.html" %} {% block head %} {% endblock %} {% block content %}
🤖 AI Model
Primary model i
Primary model
The AI model used for all responses. DuckClaw uses LiteLLM, so any supported provider works. Prefix with groq/, gemini/, or ollama/ for non-Claude models. Example: gemini/gemini-2.0-flash
Reasoning model i
Reasoning model
Used for reflection and response synthesis — complex multi-step thinking. Leave blank to use the primary model. Examples: claude-sonnet-4-5-20251001, deepseek/deepseek-r1, groq/deepseek-r1-distill-llama-70b
Vision model i
Vision model
Used for screenshot analysis, image understanding, and camera skills. Leave blank for auto-detection (falls back to gemini/gemini-2.0-flash if primary can't do vision). Requires a vision-capable API key — set it in ~/.duckclaw/.env or via the CLI setup. Examples: gemini/gemini-2.0-flash, claude-haiku-4-5-20251001, gpt-4o
Text to speech i
Text to speech model
Used for text-to-speech synthesis and voice output. Default: gemini/gemini-2.0-flash (supports TTS natively, free tier available). Requires GEMINI_API_KEY in ~/.duckclaw/.env. Alternative: groq/whisper-large-v3 (fast, free tier).
Max tokens i
Max tokens
Maximum tokens the model generates per response. Range: 256–32768. Recommended: 4096.
Temperature i
Temperature
Controls randomness. 0 = deterministic. 1–2 = more creative. Recommended: 0.7.
Cost tracking i
Cost tracking
Estimates token cost for each LLM call and logs it.
Restart DuckClaw to apply changes.
🛡️ Global Permission Settings
Default tier i
Default permission tier
Fallback tier for unclassified actions.
🟢 safe — silent auto-approve
🔵 notify — auto-approve + notification
🟡 ask — require your confirmation
🔴 block — never allow
Audit log i
Audit log
Records every action to a persistent log. Required for the Audit Log page to work.
Notify on safe i
Notify on safe
Show a notification even for SAFE-tier actions (normally silent).
Restart required to apply.
⚙️ Action Permission Rules
Change the tier for individual actions. = custom override. HARDCODED = cannot be changed.
Action Type Current Tier Default
Loading rules…
🔒 Security
Prompt injection defense i
Prompt injection defense
Scans all messages and responses for injection patterns. Strongly recommended.
Context isolation i
Context isolation
Wraps external data in a security fence so the LLM treats it as data only — prevents indirect injection attacks.
Restart required to apply.
🖥️ Dashboard
Port i
Dashboard port
The local port where the DuckClaw dashboard is served. Default: 8741. Restart required after changing.
Change via ~/.duckclaw/duckclaw.yaml
Restart required to apply.
🧠 Memory read-only
Database i
SQLite database path
Stores all user facts, conversation history, and ingested file records.
{{ config.memory.db_path }}
Vector index i
ChromaDB vector index path
Stores semantic embeddings of conversations and documents for memory search.
{{ config.memory.chroma_path }}
Max facts i
Max facts
Maximum number of user facts held in memory. Older facts are pruned when limit is reached.
{{ config.memory.max_facts }}
Search results {{ config.memory.semantic_search_results }}
{% endblock %} {% block scripts %} {% endblock %}