Context is precious. Execution is sacred. Trust is earned.Python 3.10–3.13

Architecture

Agent-Gantry separates discovery, retrieval, schema conversion, and execution so teams can adopt one layer at a time.

Registry

Stores tool definitions inferred from decorators, explicit schemas, modules, MCP servers, and framework adapters.

Sync manager

Fingerprints tool definitions and syncs changes to the vector store only when needed.

Semantic router

Embeds queries, searches candidate tools, applies score thresholds and optional rerankers, and can use LLM intent classification.

Execution engine

Runs sync or async callables behind security policies, retries, rate limits, circuit breakers, telemetry, and callbacks.

Runtime flow

User prompt → semantic query → vector candidates → rerank / threshold → provider schema → model call → ToolCall → secure execution → telemetry/result

Deployment modes

  • Library-only: import AgentGantry into an app and decorate existing tools.
  • Framework middleware: bridge into Microsoft Agent Framework, LangChain, LangGraph, LlamaIndex, CrewAI, and others.
  • Protocol service: expose tools through MCP or A2A for cross-agent interoperability.