Agentic Chat

When no domain agent matches a message, KULVEX falls back to agentic chat — a multi-round tool_use loop powered by the LLM.

How It Works

User sends a message that doesn’t match any agent pattern
KULVEX injects 12 curated tools into the system prompt (~984 tokens)
The LLM generates a response, optionally calling tools
If tools are called, results are fed back to the LLM
Repeat up to 5 rounds until the LLM produces a final text response

User: "What's my energy production this week compared to last week?"
  │
  ├── Round 1: LLM calls get_solar_history(days=14)
  ├── Round 2: LLM analyzes data, generates comparison
  └── Final: Natural language response with numbers

Tool Set

The ALWAYS-available tool set includes:

Tool	Description
`get_weather`	Current weather + forecast
`get_solar`	Solar energy production
`get_home_status`	Smart home device states
`control_device`	Turn on/off home devices
`get_presence`	Who’s home
`get_security`	Security system status
`search_web`	Web search
`send_message`	Send via messaging channel
`get_system_info`	Server stats
`run_backup`	Trigger backup
`analyze_url`	Fetch and analyze a URL
`get_calendar`	Calendar events

Security: ToolGuard

Every tool call goes through ToolGuard before execution:

Validates parameters against expected schemas
Blocks dangerous operations (rm -rf, DROP TABLE, etc.)
Rate limits tool calls per session
Logs all tool executions for audit

Cloud vs Local

Local (Mnemo): Uses the same agentic loop but with Ollama-compatible tool_use format
Cloud (Claude): Uses Anthropic’s native tool_use API, which is more reliable for complex multi-step tasks

Limits

Max 5 rounds per message to prevent infinite loops
12 tools in the ALWAYS set to fit within context window
If the LLM fails to produce a valid tool call, the round is skipped

Agent List Creating Agents