Skip to main content

Documentation Index

Fetch the complete documentation index at: https://raveculture-mintlify-x402-fallbacks-1774312113.mintlify.app/llms.txt

Use this file to discover all available pages before exploring further.

​
Architecture

How Agentbot is built. A reference for contributors, self-hosters, and anyone who wants to understand the system.

​
System Overview

┌──────────────────────────────────────────────────────────────┐
│                        USERS                                 │
│         Telegram · Discord · WhatsApp · Web Dashboard        │
└──────────────────────────┬───────────────────────────────────┘
                           │
                    ┌──────▼──────┐
                    │   VERCEL    │
                    │  Next.js 16 │
                    │  (Frontend) │
                    └──────┬──────┘
                           │
              ┌────────────┼────────────┐
              │            │            │
       ┌──────▼──────┐ ┌──▼───┐ ┌──────▼──────┐
       │   RENDER    │ │Redis │ │  Neon PG    │
       │  Backend    │ │Cache │ │  Database   │
       │  (Express)  │ │      │ │             │
       └──────┬──────┘ └──────┘ └─────────────┘
              │
    ┌─────────┼─────────┐
    │         │         │
┌───▼───┐ ┌──▼──┐ ┌────▼────┐
│Worker │ │Ollama│ │  A2A    │
│Service│ │ AI  │ │  Bus    │
└───────┘ └─────┘ └─────────┘

​
Components

​
Frontend (Vercel)

Stack: Next.js 16 (App Router) + React 19 + Tailwind CSS v4 The web dashboard, onboarding flow, billing, and all user-facing pages. Deployed on Vercel with automatic Git-based deploys. Key directories:
  • web/app/ — Pages and API routes
  • web/app/api/ — Server-side API proxies (provision, billing, auth)
  • web/app/components/ — Shared UI components
  • web/app/lib/ — Utilities (auth, Stripe, security)

​
Backend (Render)

Stack: Express.js + TypeScript The core API server. Handles agent provisioning, deployments, A2A communication, and skill management. Key services:
  • POST /api/provision — Create new agents (public, no auth)
  • POST /api/deployments — Deploy agent containers (auth required)
  • GET /api/agents — List and manage agents
  • POST /api/ai/* — AI provider proxy (OpenRouter, Anthropic, etc.)
  • GET /api/browse/* — File explorer (tree, read, write, git sync)
  • GET /api/logs/:agentId/stream — Live log tail (SSE)
  • GET /api/usage/* — Usage tracking (tokens, costs, tool metrics)
Protected endpoints use JWT auth middleware that verifies the token, attaches user context, and sets the RLS context on the database connection. See the auth API for details.

​
Worker service (Render)

A standalone background job processor built on BullMQ. It connects to the same Redis instance used by the backend and processes jobs from two queues:
QueueJob typesConcurrency
tasksscheduled-task — cron jobs per agent5
skill-execution — run a skill on behalf of an agent
provisionnew-agent — provision a new agent asynchronously2
The worker also handles:
  • A2A message routing between agents
  • Webhook processing (Stripe, Telegram, Discord)
The worker listens for SIGTERM and drains in-flight jobs before shutting down.

​
Database (Neon PostgreSQL)

Stores:
  • User accounts and authentication
  • Agent configurations and metadata
  • Billing, subscriptions, and usage tracking
  • Skill marketplace data
  • A2A message history
All user-scoped tables are protected by PostgreSQL row-level security (RLS) policies. Each authenticated request sets a database-level user context so queries automatically return only the calling user’s data. Admin users bypass RLS and can access all rows. See Security for the full list of protected tables.

​
Cache (Redis)

Used for:
  • Rate limiting
  • Session management
  • Real-time agent status
  • A2A message queues
  • Webhook deduplication

​
AI Layer (Ollama + BYOK)

Two modes:
  • Ollama (self-hosted): Local inference for self-hosted deployments (Llama 3, Mistral)
  • BYOK: Users bring their own API keys for OpenRouter, Anthropic, OpenAI, Google, or Groq

​
Watchdog

Monitors agent gateway processes for health and auto-recovers from failures:
CapabilityDetails
Health checksEvery 2 minutes (configurable via WATCHDOG_CHECK_INTERVAL)
Crash detectionListens for gateway exit events
Crash-loop detection3 crashes in 5-minute window
Auto-repairKill → wait 5s → restart → verify health (max 2 attempts)
NotificationsTelegram + Discord alerts for crashes, repairs, and recovery
Degraded state5-second retry interval when health check fails
See Watchdog for the full API.

​
File Explorer + Git Sync

Browser-based workspace management — no SSH needed:
  • File tree: Recursive directory listing with depth control
  • Read/Write: View and edit files inline (1MB limit)
  • Git status: See uncommitted changes
  • Git diff: Per-file or full workspace diffs
  • Git sync: One-click commit + push to GitHub
  • Git log: Browse recent commit history
See File Explorer API for the full API.

​
Usage Tracker

SQLite-backed token and cost tracking per agent, model, and time period:
  • Per-event recording (session, agent, provider, model, tokens, cost)
  • Daily aggregation with upsert
  • Tool event monitoring (success rate, duration)
  • Query APIs: summary, by-agent, by-model, daily totals
See Usage Tracking API for the full API.

​
A2A Bus (Agent-to-Agent)

Enables agents to communicate directly:
  • Request/response patterns
  • Fire-and-forget messages
  • Skill delegation between agents
  • Cross-tenant isolation enforced

​
Data Flow: Agent Provisioning

User fills form (frontend)
       │
       ▼
POST /api/provision (Vercel)
       │
       ▼
POST /api/provision (Render backend)
       │
       ├── Validate input (plan, provider, tokens)
       ├── Generate userId + agentId
       ├── Store config in PostgreSQL
       ├── Allocate port + subdomain
       ├── Deploy Docker container
       ├── Configure Caddy reverse proxy
       │
       ▼
Return agent URL + stream key to user

​
Data Flow: Agent Runtime

User sends message (Telegram/Discord/WhatsApp)
       │
       ▼
Webhook → Worker Service
       │
       ├── Route to correct agent container
       ├── Load agent config + memory
       ├── Execute skills if needed
       ├── Call AI provider (BYOK or Ollama)
       ├── Check A2A bus for delegated tasks
       │
       ▼
Response sent back to user

​
Self-Hosting

Agentbot is fully open source. To self-host:
  1. Frontend: Deploy web/ to Vercel or any Next.js-compatible host
  2. Backend: Run agentbot-backend/ on any Docker-capable server
  3. Worker: Run the worker entry point from agentbot-backend/ using Dockerfile.worker
  4. Database: Neon Postgres (free tier available) or any PostgreSQL 15+
  5. Redis: Any Redis 7+ instance
  6. AI: Ollama for local inference, or configure BYOK providers
See Installation for full setup instructions.

​
Tech Decisions

ChoiceWhy
Next.js 16App Router, server components, Vercel-native DX
Express (not serverless)Long-running agent processes need persistent connections
Neon PostgresServerless, scales to zero, generous free tier
RedisFast rate limiting and session management
DockerIsolated agent environments per user
BYOK over resellingUsers pay providers directly — no markup, no lock-in
Base (not Ethereum)Low fees, fast finality, Coinbase ecosystem