Documentation

ClaudePaw is a multi-tenant platform for deploying autonomous AI agents. Agents run ReAct loops with real tools, persistent memory, and composable skills — all on Cloudflare's edge.

What is an agent?

An agent is not a chatbot. It receives a message, reasons about what to do, executes tools (web search, shell commands, browser automation), observes the results, and repeats until the task is complete. This is the ReAct (Reason + Act) pattern.

Platform architecture

Every agent session runs in a Durable Object — an isolated, stateful compute unit at the edge. The agent loop:

  • Builds context from system prompt, skills, memory, and conversation history
  • Sends to the LLM via AI Gateway
  • If the LLM returns tool calls, executes them and loops
  • If the LLM returns text, streams it to the client
  • Persists memory and usage data after each conversation

Two compute tiers

Durable Object — handles the agent loop, context, memory, and lightweight tools (search, URL reading, memory ops). Always running.

Sandbox Container — handles heavy tools: shell exec, browser automation, file I/O. One per org, spun up on demand, auto-sleeps after 10 minutes.

Next steps