Agent-Gantry

Universal Tool Orchestration Platform for LLM-Based Agent Systems

Context is precious. Execution is sacred. Trust is earned.

Agent-Gantry is a Python library (v0.16.0) for building agents that can discover, select, and execute the right tools without flooding every prompt with every schema your organization owns. It combines semantic retrieval, provider schema conversion, secure execution, framework bridges, MCP/A2A interoperability, persistence adapters, and observability into one tool orchestration layer.

Documentation

The project documentation is now an Astro + React + TypeScript site with an implementation journey, interactive tool lifecycle walkthrough, integration matrix, and production operations guidance.

npm install
npm run dev      # local docs server
npm run build    # type-check and static build
npm run preview  # verify the generated site styling

Start with the rich docs in src/pages/index.astro. The published site lives at codehalwell.github.io/Agent-Gantry.

Install

uv add agent-gantry
# or
pip install agent-gantry

Useful extras:

uv add "agent-gantry[openai]"
uv add "agent-gantry[anthropic]"
uv add "agent-gantry[google-genai]"
uv add "agent-gantry[lancedb,nomic]"
uv add "agent-gantry[mcp,a2a]"
uv add "agent-gantry[agent-frameworks]"
uv add "agent-gantry[all]"

Quick start

from openai import AsyncOpenAI
from agent_gantry import AgentGantry, set_default_gantry, with_semantic_tools

client = AsyncOpenAI()
gantry = AgentGantry()
set_default_gantry(gantry)

@gantry.register(tags=["weather"])
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"The weather in {city} is 72°F and sunny."

@with_semantic_tools(limit=3, dialect="openai")
async def ask_llm(prompt: str, *, tools=None):
    return await client.chat.completions.create(
        model="gpt-5.5",
        messages=[{"role": "user", "content": prompt}],
        tools=tools,
    )

await ask_llm("What's the weather in San Francisco?")

Agent-Gantry automatically fingerprints registered tools, syncs definitions to the configured vector store, retrieves semantically relevant tools, and converts schemas to the requested provider dialect.

Core capabilities

• Semantic tool routing: reduce prompt context by retrieving top-k relevant tools instead of injecting every tool.
• Register once, run anywhere: emit schemas for OpenAI-compatible APIs, Anthropic, Gemini, framework adapters, MCP, and A2A paths.
• Secure execution: run tools through policies, capabilities, timeouts, retries, rate limits, circuit breakers, callbacks, and telemetry.
• Persistence and retrieval: use in-memory defaults, LanceDB, Qdrant, Chroma, pgvector, OpenAI/Nomic/sentence-transformers embeddings, and rerankers.
• Framework coverage: Microsoft Agent Framework plus LangChain, LangGraph, LlamaIndex, CrewAI, Google ADK, Pydantic AI, OpenAI Agents SDK, Haystack, Agno, Strands Agents, and DSPy.
• MCP both ways: consume local (stdio) and remote (Streamable HTTP / SSE) MCP servers, and serve your registry to Claude Desktop, Claude Code or any remote client with gantry.serve_mcp() / agent-gantry serve-mcp --module my_app.tools — two meta-tools instead of the whole tool list.
• Skills, retrieved by meaning: load any Agent Skills (SKILL.md) directory with gantry.add_skills_from_directory(...) and inject only the skills relevant to each prompt.
• Bundled Claude Skill: install with agent-gantry install-skill --claude or target a project-local skills directory.
• Selection without embeddings: point a decision model at the catalogue instead of a vector store. JevSelector replaces embed-then-search for tools, skills and MCP servers; JevReranker refines the shortlist when the catalogue is too large to send. Both fail open to semantic routing.

Selecting tools without a vector store

Semantic routing embeds the query and searches. A selector asks a decision model directly, which needs no embedder, no vector store and no sync — and can return nothing when nothing fits, which top-k cannot express.

from agent_gantry import AgentGantry, JevSelector

gantry = AgentGantry(selector=JevSelector(threshold=0.3))  # reads TYPESAFE_API_KEY

@gantry.register(tags=["email"], examples=["show my messages", "any new mail"])
def list_inbox(limit: int = 10) -> list[str]:
    """List the most recent messages sitting in the inbox."""
    return []

@gantry.register(tags=["email"], examples=["email Bob about the meeting"])
def send_email(to: str, body: str) -> str:
    """Send an email message to a named recipient."""
    return "sent"

# With a key, this honours the negation and returns list_inbox alone.
# Without one, selection fails open and semantic routing answers instead.
tools = await gantry.retrieve_tools("show my messages, but do not send anything")

The same selector covers retrieve_skills() and retrieve_mcp_servers(). Above a few hundred entries, keep semantic search and refine it instead:

from agent_gantry import AgentGantry, JevReranker

gantry = AgentGantry(reranker=JevReranker())   # reorders the vector-search shortlist

Write examples=[...] on your tools before reaching for either: on our own benchmark that moved the default embedder from 1/5 to 5/5, more than any model change did. See agent_gantry/adapters/selectors/README.md for the measured trade-offs.

Manual retrieval and execution

from agent_gantry import AgentGantry
from agent_gantry.schema.execution import ToolCall

gantry = AgentGantry()

@gantry.register(tags=["finance"])
def calculate_tax(amount: float) -> float:
    """Calculate US sales tax for an amount."""
    return amount * 0.08

tools = await gantry.retrieve_tools("What is the tax on $100?", limit=5)
result = await gantry.execute(ToolCall(
    tool_name="calculate_tax",
    arguments={"amount": 100.0},
))

Development

uv sync --all-extras
uv run pytest
npm install
npm run build

License

MIT