The framework everything is built on. Cloudflare's Agents SDK (v0.3.10)
provides the stateful agent architecture, WebSocket handling, tool orchestration, and conversation management
that make ZeroBot possible.
- Each agent extends Agent<Env, KitState> — the base class for all agents (unlimited)
- Built-in WebSocket support for real-time streaming chat
- @callable() decorators expose RPC methods for agent-to-agent communication
- Automatic state persistence via Durable Objects — agents remember everything
- Tool orchestration loop: message → LLM → tool_use → result → loop until done
- Only 2 production dependencies total:
agents and @cloudflare/playwright
▶ Click for details
The entire application runs as a Cloudflare Worker. No servers, no containers. Deployed
globally across 330+ edge hops with sub-millisecond cold starts.
- Web UI (15 pages), API endpoints, and webhook handlers all run in a single Worker
- Slack, Discord, and WhatsApp webhooks processed at the edge
- All business logic, LLM routing, and tool orchestration in one deployment
- Automatic scaling: handles 1 request or 10,000 concurrently
- Zero traditional servers to maintain, patch, or scale
▶ Click for details
Each AI agent is a Durable Object instance that maintains persistent state across
conversations. Create as many agents as you need.
- Unlimited agents — e.g. Kit (chief-of-staff), Scout (marketing), Closer (sales), Sentinel (security)
- WebSocket connections for real-time chat streaming
- Each user gets their own isolated DO instances (multi-tenant by design)
- Agents coordinate via RPC calls (e.g., one agent asks another for analysis)
- State survives between requests — agents "remember" the conversation
▶ Click for details
All structured data lives in D1 — conversations, messages, memories, skills, audit logs,
experiments, and more. Tenant-isolated on every query.
- Conversations, messages, and memory entities with full-text search
- Skills engine, notification system, and achievement tracking
- Audit log for every security-relevant action (compliance-ready)
- A/B testing framework with statistical analysis
- Every query includes user_id for strict tenant isolation
▶ Click for details
Identity files, knowledge documents, uploaded images, and daily memory summaries — all
stored in R2 with zero egress fees.
- Agent identity files (SOUL.md, CONTEXT.md) that define each agent's personality
- User profile (USER.md) with system-prompt-level trust
- Knowledge base documents and uploaded files (images, PDFs)
- Daily memory summaries generated by nightly cron jobs
- Paths scoped per-user: {userId}/{botId}/ for complete isolation
▶ Click for details
The speed layer. Rate limiting, personality caching, embedding caches, and config
storage — all with global edge reads in under 1ms.
- Rate limiting: per-token and per-user with dual-layer enforcement
- Personality file caching with 7-day TTL and stale-while-revalidate
- Embedding caches to avoid re-generating vectors
- MCP tool definition caching (1-hour TTL per connection)
- Config storage for changelog, roadmap, and feature flags
▶ Click for details
Semantic memory recall. When an agent searches its memory, Vectorize finds the most
relevant entries using cosine similarity on embeddings.
- Memory embeddings stored as high-dimensional vectors
- Cosine similarity search finds semantically related memories
- Per-user namespace isolation prevents cross-user data leakage
- Multi-signal scoring: semantic similarity + importance + recency + reinforcement
- Enables "fuzzy" recall — agents find relevant context even with different wording
▶ Click for details
Cloudflare's GPU fleet handles lightweight AI tasks: intent classification, memory
operations, embedding generation, context compaction, and content-level DLP.
- Intent classification: routes messages to the right LLM (simple vs complex)
- Memory consolidation: deduplicates entries before storage
- Contradiction detection: identifies conflicting memories and auto-resolves
- Context compaction: summarizes old messages to keep context lean
- Embedding generation (bge-base-en-v1.5) for semantic search
- Content-level DLP: Cloudflare AI API inspects content for PII and sensitive data at
the inference layer
- Audio transcription: Deepgram Nova-3 + Whisper models for voice and audio processing
▶ Click for details
All external LLM calls (Claude, Gemini, Perplexity) route through AI Gateway for
caching, logging, cost tracking, granular DLP, and guardrails.
- 8-way routing: Workers AI, Claude Sonnet/Haiku/Opus, Perplexity, Gemini Flash/Pro/Image
- Response caching reduces redundant API calls and costs
- Real-time cost tracking and usage analytics per model
- Granular DLP: Content inspection on every LLM request/response — PII, credentials,
and sensitive data intercepted before leaving the edge
- AI Guardrails: Configurable safety rules enforce content policies, block prompt
injection, and prevent harmful outputs
- Rate limiting prevents runaway spending
- Automatic fallback if gateway is unavailable
▶ Click for details
kitbot.0arc.ai sits behind Cloudflare Access for SSO and zero-trust authentication. No
VPN needed — identity-based access at the edge.
- Zero-trust security: every request authenticated before reaching the Worker
- SSO integration for seamless sign-in
- No VPN infrastructure to maintain
- Access policies enforced at the edge, not in application code
- Combined with bearer token auth (SHA-256) for API-level security
▶ Click for details
Your agents can browse the web using Cloudflare's headless browser API. Take
screenshots, extract content, and interact with pages on demand.
- Headless Chromium powered by @cloudflare/playwright
- Screenshot capture and content extraction from any URL
- Used for research tasks, competitor monitoring, and link previews
- Runs in Cloudflare's infrastructure — no browser instances to manage
- Integrated as an agent tool: agents decide when to browse
▶ Click for details
kitbot.0arc.ai routes through Cloudflare DNS with full SSL/TLS, DDoS protection, and WAF
— all managed from one dashboard.
- Custom domain with automatic SSL certificate management
- Cloudflare DNS for fast, reliable resolution worldwide
- Built-in DDoS protection at no extra cost
- Web Application Firewall (WAF) blocks malicious requests
- Full TLS encryption from user to Worker
▶ Click for details
Automated background jobs: memory consolidation, morning briefings, social monitoring
every 30 minutes, and pending task processing.
- Nightly memory consolidation: deduplicates and summarizes the day's conversations
- Morning briefings: prepares a daily summary of priorities and events
- Social media monitoring: scans Reddit and YouTube every 30 minutes
- Pending memory processing: handles deferred memory writes
- No external scheduler needed — runs natively on Cloudflare
▶ Click for details