Idun Engine turns a Python file into a production agent in minutes. Tools, memory, guardrails, MCP, observability, all wired by default. GPL 3.0, model-agnostic, MCP-native. Run it on your laptop, your VPC, or your air-gapped cluster.
1# first agent · 30 seconds2$ pip install idun-agent-engine3✓ installed · # langgraph · adk · mcp · guardrails4$ idun agent serve --config config.yaml5✓ running on :8000 · streaming SSE6$ curl :8000/agent/invoke7✓ live · OpenTelemetry on · PII guard ✓
Idun Engine is a Python package. You write a LangGraph or Google ADK agent, point a YAML config at it, and let the engine wrap it as a production service. Streaming SSE, traces, guardrails, MCP tools, the whole stack.
For solo demos, idun-agent-standalone bundles a chat UI, an admin panel and a local trace viewer in one process. Same engine underneath.
# graph.pyfrom langgraph.prebuilt import create_react_agentfrom langchain.chat_models import init_chat_modelgraph = create_react_agent(model=init_chat_model("gpt-4o"),tools=[search_tool, hubspot_tool],prompt="You are a customer support agent.",)
Bring your graph. Engine wraps it as a streaming FastAPI service. CrewAI and Haystack on the roadmap.
Stdio + streamable_http MCP servers, declared in YAML, auto-loaded at startup. LangChain + ADK tool adapters.
LangGraph checkpointers. SQLite for dev, Postgres for prod. Per-agent, per-thread, per-user scoping.
Built on Guardrails AI hub: PII (email · phone · CC · SSN), NSFW, toxic, jailbreak, competitor, bias, gibberish, topic.
Langfuse, Arize Phoenix, LangSmith, GCP Cloud Trace + Logging. Every plan, tool call, token cost is a span.
AG-UI + CopilotKit for the browser. Slack, Teams, Discord, Google Chat, WhatsApp adapters bundled.
| Provider | Modes | Region |
|---|---|---|
| OpenAIlangchain-openai | ●chat · tools · stream | US · EU |
| Anthropiclangchain-anthropic | ●chat · tools · stream | US · EU |
| Google Geminilangchain-google-vertexai | ●chat · tools · stream | europe-west9 |
| Mistrallangchain-mistralai | ●chat · tools · stream | FR · Le Chat |
| Llama via vLLMself-host | ●chat · tools | your VPC |
| AWS Bedrocklangchain-aws | ●chat · stream | eu-west-3 |
| Azure OpenAIlangchain-openai | ●chat · tools · stream | france-central |
| Coherelangchain-cohere | ●chat · embed | EU |
Engine doesn't ship a model layer of its own. We delegate to LangChain's init_chat_model() so swapping providers is a one-line change. Your tools, guardrails and traces don't notice.
Self-host Llama on vLLM. Call Mistral on Le Chat. Route to Claude for long context. Route to GPT-4o-mini for cheap classification. Same agent definition.
model: "gpt-4o"model: "claude-sonnet-4-5"# no other change required
idun-agent-standalone is the engine wrapped for solo devs. Chat UI, admin panel, local SQLite trace viewer, all bundled in one Python process. No Docker compose, no infra, no platform.
Run idun-standalone init, open localhost, ship a working agent in five minutes. When you outgrow it, the same agent code runs on full Engine in production. Nothing to rewrite.
Idun Engine is licensed GPL-3.0-only. The full source (runtime, integrations, scaffolds, examples) lives in one repo. No CLA-traps, no separate "enterprise" runtime hidden behind a contract. Idun Platform is a commercial product on top of Engine; Engine never depends on it.
The roadmap is public, RFCs are public, every PR is reviewed in the open. Fork it. Audit it. Run it for ten years.
Three layers, each optional except the engine. Run the engine alone in your VPC. Add Standalone for solo dev ergonomics. Add Platform when you reach fleet-scale governance.
Multi-tenant control plane on top of one-or-many engines. SSO, RBAC, AI Act registry, FinOps. Commercial product.
Wraps the engine with a chat UI, an admin panel, a local trace viewer. One-command init. Same engine underneath.
FastAPI process. Wraps your LangGraph or ADK graph and wires MCP, memory, guardrails, channels, OTel. The only required layer.
| Capability | Idun Engine runtime · GPL 3.0 | LangChain ecosystem LangChain · LangGraph · py · TS | CrewAI multi-agent framework · MIT | Google ADK vendor SDK · GCP | Inhouse your team · from scratch |
|---|---|---|---|---|---|
| Production runtime | ●native | DIY | DIY | partial | DIY |
| Self-host (your VPC, on-prem) | ●yes | ●yes | ●yes | ●yes | ●yes |
| Type-safe MCP servers, config-driven | ●native | partial | no | partial | no |
| OTel traces by default | ●on by default | no | no | no | no |
| Langfuse · Phoenix · LangSmith · GCP | ●native | partial | partial | no | DIY |
| Guardrails (PII · NSFW · jailbreak ...) | ●9+ built-in | no | no | no | DIY |
| AG-UI + CopilotKit streaming | ●native | no | no | no | no |
| Slack · Teams · WhatsApp · Discord | ●bundled | partial | no | no | DIY |
| JWT / SSO + session scoping | ●yes | no | no | no | DIY |
| Lock-in | ●none · OSS | none | none | high | none |
Engine is a runtime on top of LangGraph + ADK, not a replacement for the libraries themselves.
$ pip install idun-agent-engine$ idun agent serve→ live on :8000$ curl :8000/agent/invoke→ OTel on · audit on