Architecture Overview

Agentic Forge is built from modular components that work together to create efficient, interoperable AI agents with smart tool management.

Components

Component	Description
Orchestrator	Standalone component managing the LLM conversation loop with model routing, tool routing, and hooks for observability.
Armory	Protocol gateway that aggregates multiple tool sources and exposes them through a unified MCP interface.
Tool RAG	Dynamic tool selection using semantic search to reduce context usage and improve accuracy.
Interfaces	CLI, Web UI, WebSocket API, and Python SDK for interacting with the system.
Anvil	CLI and web-based inspector for testing and debugging MCP servers during development.

Key Innovations

Protocol Interoperability

Different LLMs output tool calls in different formats (OpenAI, Anthropic, Gemini, etc.). Agentic Forge normalizes these through Pydantic AI and translates to whatever backend format is needed.

Token Optimization with TOON

Tool results are converted from JSON to TOON (Token-Oriented Object Notation) at the gateway level, achieving 15-40% token savings (scaling with data size) without requiring changes to existing MCP servers or tools.

Smart Tool Selection

Instead of loading all tools into context, Tool RAG uses semantic search to dynamically select only relevant tools. Our benchmarks show 60% context reduction when combined with TOON format.

Interactive Tool UIs with MCP Apps

Tools can include interactive UI components that render in the chat. Our location picker demonstrates this—users click on a map to select coordinates instead of typing them. Built on the MCP Apps specification with sandboxed iframes and JSON-RPC messaging.

Data Flow

Technology Stack

Layer	Technology	Purpose
Framework	Pydantic AI	Model abstraction, type-safe tools
Protocol	MCP (Streamable HTTP)	Tool communication standard
Format	TOON	Token-efficient data encoding
Search	Vector embeddings	Semantic tool matching
API	WebSocket + REST	Real-time streaming + management