Idun Agent Engine

Turn any LangGraph-based agent into a production-grade API in minutes.

Idun Agent Engine is a lightweight runtime and SDK that wraps your agent with a FastAPI server, adds streaming, structured responses, config validation, and optional observability — with zero boilerplate. Use a YAML file or a fluent builder to configure and run.

Installation

pip install idun-agent-engine

• Requires Python 3.12+
• Ships with FastAPI, Uvicorn, LangGraph, SQLite checkpointing, and optional observability hooks

A note on PyPI's JSON metadata

Both TestPyPI's and public PyPI's JSON API at /pypi/idun-agent-engine/<version>/json report entry_points: null and omit idun-agent-standalone from requires_dist. This is a Warehouse rendering quirk: the wheel itself wires the idun console script and bundles the standalone package. To verify post-install:

python -c "import idun_agent_engine, idun_agent_standalone, idun_agent_schema"
idun --help

Both commands succeed when the install is healthy.

Quickstart

1) Minimal one-liner (from a YAML config)

from idun_agent_engine.core.server_runner import run_server_from_config

run_server_from_config("config.yaml")

Example config.yaml:

server:
  api:
    port: 8000

agent:
  type: "langgraph"
  config:
    name: "My Example LangGraph Agent"
    graph_definition: "./examples/01_basic_config_file/example_agent.py:app"
    # Optional: conversation persistence
    checkpointer:
      type: "sqlite"
      db_url: "sqlite:///example_checkpoint.db"
    # Optional: provider-agnostic observability
    observability:
      provider: langfuse   # or phoenix
      enabled: true
      options:
        host: ${LANGFUSE_HOST}
        public_key: ${LANGFUSE_PUBLIC_KEY}
        secret_key: ${LANGFUSE_SECRET_KEY}
        run_name: "idun-langgraph-run"

Run and open docs at http://localhost:8000/docs.

2) Programmatic setup with the fluent builder

from pathlib import Path
from idun_agent_engine import ConfigBuilder, create_app, run_server

config = (
    ConfigBuilder()
    .with_api_port(8000)
    .with_langgraph_agent(
        name="Programmatic Example Agent",
        graph_definition=str(Path("./examples/02_programmatic_config/smart_agent.py:app")),
        sqlite_checkpointer="programmatic_example.db",
    )
    .build()
)

app = create_app(engine_config=config)
run_server(app, reload=True)

Endpoints

All servers expose these by default:

• POST /agent/invoke: single request/response
• POST /agent/stream: server-sent events stream of ag-ui protocol events
• GET /health: service health with engine version
• GET /: root landing with links

Invoke example:

curl -X POST "http://localhost:8000/agent/invoke" \
  -H "Content-Type: application/json" \
  -d '{"query": "Hello!", "session_id": "user-123"}'

Stream example:

curl -N -X POST "http://localhost:8000/agent/stream" \
  -H "Content-Type: application/json" \
  -d '{"query": "Tell me a story", "session_id": "user-123"}'

LangGraph integration

Point the engine to a StateGraph variable in your file using graph_definition:

# examples/01_basic_config_file/example_agent.py
import operator
from typing import Annotated, TypedDict
from langgraph.graph import END, StateGraph

class AgentState(TypedDict):
    messages: Annotated[list, operator.add]

def greeting_node(state):
    user_message = state["messages"][-1] if state["messages"] else ""
    return {"messages": [("ai", f"Hello! You said: '{user_message}'")]}

graph = StateGraph(AgentState)
graph.add_node("greet", greeting_node)
graph.set_entry_point("greet")
graph.add_edge("greet", END)

# This variable name is referenced by graph_definition
app = graph

Then reference it in config:

agent:
  type: "langgraph"
  config:
    graph_definition: "./examples/01_basic_config_file/example_agent.py:app"

Behind the scenes, the engine:

• Validates config with Pydantic models
• Loads your StateGraph from disk
• Optionally wires a SQLite checkpointer via langgraph.checkpoint.sqlite
• Exposes invoke and stream endpoints
• Bridges LangGraph events to ag-ui stream events

Observability (optional)

Enable provider-agnostic observability via the observability block in your agent config. Today supports Langfuse and Arize Phoenix (OpenInference) patterns; more coming soon.

agent:
  type: "langgraph"
  config:
    observability:
      provider: langfuse   # or phoenix
      enabled: true
      options:
        host: ${LANGFUSE_HOST}
        public_key: ${LANGFUSE_PUBLIC_KEY}
        secret_key: ${LANGFUSE_SECRET_KEY}
        run_name: "idun-langgraph-run"

Configuration reference

• server.api.port (int): HTTP port (default 8000)
• agent.type (enum): currently langgraph (CrewAI placeholder exists but not implemented)
• agent.config.name (str): human-readable name
• agent.config.graph_definition (str): absolute or relative path/to/file.py:variable
• agent.config.checkpointer (sqlite): { type: "sqlite", db_url: "sqlite:///file.db" }
• agent.config.observability (optional): provider options as shown above
• mcp_servers (list, optional): collection of MCP servers that should be available to your agent runtime. Each entry matches the fields supported by langchain-mcp-adapters (name, transport, url/command, headers, etc.).

Config can be sourced by:

• engine_config (preferred): pass a validated EngineConfig to create_app
• config_dict: dict validated at runtime
• config_path: path to YAML; defaults to config.yaml

MCP Servers

You can mount MCP servers directly in your engine config. The engine will automatically create a MultiServerMCPClient and expose it on app.state.mcp_registry.

mcp_servers:
  - name: "math"
    transport: "stdio"
    command: "python"
    args:
      - "/path/to/math_server.py"
  - name: "weather"
    transport: "streamable_http"
    url: "http://localhost:8000/mcp"

Inside your FastAPI dependencies or handlers:

from idun_agent_engine.server.dependencies import get_mcp_registry

@router.get("/mcp/{server}/tools")
async def list_tools(server: str, registry = Depends(get_mcp_registry)):
    return await registry.get_tools(server)

Or outside of FastAPI:

from langchain_mcp_adapters.tools import load_mcp_tools

registry = app.state.mcp_registry
async with registry.get_session("math") as session:
    tools = await load_mcp_tools(session)

Examples

The examples/ folder contains complete projects:

• 01_basic_config_file: YAML config + simple agent
• 02_programmatic_config: ConfigBuilder usage and advanced flows
• 03_minimal_setup: one-line server from config

Run any example with Python 3.13 installed.

CLI and runtime helpers

Top-level imports for convenience:

from idun_agent_engine import (
  create_app,
  run_server,
  run_server_from_config,
  run_server_from_builder,
  ConfigBuilder,
)

• create_app(...) builds the FastAPI app and registers routes
• run_server(app, ...) runs with Uvicorn
• run_server_from_config(path, ...) loads config, builds app, and runs
• run_server_from_builder(builder, ...) builds from a builder and runs

Production notes

• Use a process manager (e.g., multiple Uvicorn workers behind a gateway). Note: reload=True is for development and incompatible with multi-worker mode.
• Mount behind a reverse proxy and enable TLS where appropriate.
• Persist conversations using the SQLite checkpointer in production or replace with a custom checkpointer when available.

Roadmap

• CrewAI adapter (placeholder exists, not yet implemented)
• Additional stores and checkpointers
• First-class CLI for idun commands

Contributing

Issues and PRs are welcome. See the repository:

• Repo: https://github.com/Idun-Group/idun-agent-platform
• Package path: libs/idun_agent_engine
• Open an issue: https://github.com/Idun-Group/idun-agent-platform/issues

Run locally:

cd libs/idun_agent_engine
poetry install
poetry run pytest -q

License

MIT — see LICENSE in the repo root.