01
A native agent core
The orchestration loop, provider clients, tool registry, and streaming pipeline compile with your application. There is no separate Python or Node process to deploy and supervise.
// Capabilities
Built for serious agents
PasClaw combines a compact native runtime with the primitives needed to build dependable tool-using agents in Object Pascal.
Native PascalTool callingStreamingCross-platform
02
Tools are first-class
Register Pascal procedures with typed schemas, let the model invoke them, validate their arguments, and return observations to the next reasoning turn.
03
Built for applications
Use PasClaw from a CLI, desktop UI, service, chat channel, or embedded workflow. The same core supports interactive and unattended operation.
04
Designed to extend
Add providers, tools, skills, MCP servers, hooks, and memory strategies without replacing the central agent loop.
From the documentation
Technical details
Implementation notes drawn from the PasClaw repository documentation.
Built-in tool surface
The core catalog covers filesystem work, shell and code execution, web retrieval, memory, sessions, knowledge bases, vaults, messaging, and installed skills.
- Read tools can fan out across worker threads while mutating tools stay serial.
- execute_code handles multi-line scripts without repeated inference round trips.
- MCP servers and skill manifests register additional tools into the same loop.
Delegation and communication
Agents can delegate focused work to subagents and deliver results through configured channels.
- Foreground and background subagent modes keep research out of the parent context.
- send_message is only available when a channel is configured.
- Vault and knowledge-base tools keep sensitive or curated material behind explicit retrieval calls.
Common questions
Frequently asked questions
Keep exploring
See how the rest of PasClaw fits together.