TLDR;

For those in a hurry:

The entire goal of this project is to create a framework to allow anyone to create, publish and consume AI Agents (as we do with any other codebase, plugin, package, etc), safely;
Triks come in two modes: conversational (full LLM agents that handle multi-turn conversations) and tool (native tools exposed directly to your main agent);
A handoff model routes conversations between your main agent and trik agents. The main agent calls talk_to_<trik> to hand off, and the trik calls transfer_back when done;
Security is enforced through Type-Directed Privilege Separation (TDPS) — constrained output types ensure the main agent never sees unconstrained free-form text from triks;
Everything is kept under control through a Gateway that validates schemas, manages handoffs, and enforces limits;
Everything runs locally;

This project is in its infancy and many more updates, features, fixes and improvements are planned. Any help in the right direction is greatly appreciated. Check the Trikhub repository for more info.

Core Concepts

Deep dive into TrikHub’s architecture and design principles. TrikHub is built on a simple principle: constrain what the main agent sees from trik outputs.

This section covers:

Manifest - The contract that defines a Trik
Agent Modes - Conversational vs Tool mode
Security Model - How type safety prevents injection
Sessions - Multi-turn conversations and references
Storage - Persistent key-value data for Triks
Configuration - Managing API keys and secrets
Cross-Environment - Running triks across runtimes