How Your Agent Works

Dedicated Infrastructure

Your AI agent runs on a dedicated virtual private server (VPS) — it's not shared with any other user. This ensures:

Privacy — Your data never mixes with other users
Performance — Consistent resources for your agent
Isolation — Issues with one agent can't affect others

The 6 Loops

Your agent runs 6 concurrent loops that handle different responsibilities:

Loop 1: Inbox Monitor (every 2 minutes)

Checks for inbound email replies from leads. Classifies responses by sentiment and updates lead status automatically.

Loop 2: IPC Watcher (every 5 seconds)

Handles real-time communication with the LeadClaw web app — processing commands, reporting status, and syncing data.

Loop 3: Task Scheduler (every 60 seconds)

Manages the outreach queue: picks leads to contact, drafts emails, schedules sends, and handles follow-ups.

Loop 4: Awareness Sweep (every 2 hours)

Performs lead generation — searches for new potential leads matching your active campaigns' criteria.

Loop 5: Owner Message Monitor

Your messages are delivered instantly via push notification from the web app. A backup check runs every 5 minutes to catch anything missed during agent restarts.

Loop 6: Heartbeat (every 2 hours)

Reports your agent's health status to the dashboard — uptime, memory usage, active campaigns, and error rate. The ECG-style visualization on your dashboard pulses when your agent is actively working.

AI Models

Your agent uses a multi-provider approach, routing each task to the most cost-effective model:

Gemini Flash — Primary model for email drafting, campaign execution, and lead research
Gemini Flash Lite — Classification tasks like sentiment analysis and lead scoring
Claude Haiku — Context summarization and simple decisions (fallback)
Claude Sonnet — Complex email drafting and campaign planning (fallback)
Claude Opus — Complex reasoning, self-improvement, novel edge cases

This multi-provider approach keeps costs extremely low while maintaining quality. Gemini handles the majority of tasks, with Claude models as intelligent fallbacks.