01 / 06
Tools
How the Agent Calls Code
Implement the four-method ToolInterface and the model can call your code. Built-in tools include file read, code search, database query and HTTP.
Ecosystem
Six subsystems. One agent.
Every phpClaw agent runs on the same six building blocks, on all 8 adapters. Open one to see how it works, verified from source.
02 / 06
Guard
What Stops a Bad Tool Call
Eight guards run by default and scan every message before the model sees it. A blocked message throws GuardException.
03 / 06
Skills
Domain Knowledge Without Fine-Tuning
Load knowledge from Markdown files, arrays or a remote URL. phpClaw lists each skill for the model, which loads the one a question needs; a message that names a skill gets it added straight away.
04 / 06
Hooks
The Agent Lifecycle You Can Watch
Lifecycle events such as agent.before, tool.after and guard.blocked. Listen to log, trace or alert without touching the agent.
05 / 06
Providers
Choose the Model Backend
Anthropic, OpenAI, Gemini, Groq, DeepSeek, Mistral, Ollama or any OpenAI-compatible endpoint. Add fallbacks and phpClaw tries them in order.
06 / 06
Memory
What the Agent Remembers
File, Redis or in-process storage for conversations and run state. Load a conversation by its id and the agent picks up where it left off.