Examples and notebook¶

Runnable examples¶

Short scripts, one per use case, in examples/. Each runs on a laptop CPU with opendecider-nano (a 0.8 GB download on first use), and CI runs them against the released model whenever the package or the examples change (remote_backends.py, which needs a model server, is compile-checked).

script	what it shows
`quickstart.py`	three typed questions (`choice`, `score`, `noul`) about one ticket
`support_triage.py`	a queue of tickets in one call: route by team, flag churn risk, send unsure answers to a person
`agent_guardrail.py`	check an AI agent's JSON trace before its next step: continue, retry, ask the user or stop
`prompt_guard.py`	screen a question, a jailbreak, a document with injected instructions and a long report for prompt injection, and filter retrieved passages
`confident_automation.py`	on 2,000 labelled business decisions: how many you can automate at a given accuracy
`serve_client.py`	call `opendecider serve` over HTTP (Jev's `/v1/systemone` protocol), with retries
`remote_backends.py`	the same decision through LM Studio, Ollama or vLLM
`production_router.py`	a router for production: the model behind `opendecider serve`, a JSON audit line per decision, fallback on error, OpenTelemetry spans (0.4.0 or later)

git clone https://github.com/manjunathshiva/opendecider && cd opendecider
pip install opendecider pandas pyarrow
python examples/support_triage.py

Agent frameworks¶

One script per framework, in examples/agent_frameworks/. Each runs without an LLM API key; Agent frameworks explains each integration. They need opendecider 0.5.0 or later: pip install "opendecider[agno]>=0.5.0", with the extra each script needs.

script	framework	what it shows
`langgraph_router.py`	LangGraph	`DecisionRouter` as a graph's conditional edge: each ticket to a specialist, unsure ones to a person
`llamaindex_selector.py`	LlamaIndex	`DecisionSelector` picks a RouterQueryEngine's source without an LLM call
`agno_workflow.py`	Agno	`DecisionRouter.selector()` as a workflow Router's selector
`crewai_crew.py`	CrewAI	`TaskAssigner` picks the crew member for each task
`crewai_flow.py`	CrewAI	`DecisionRouter` in a Flow's `@router`
`agent_framework_workflow.py`	Microsoft Agent Framework	`DecisionRouter.cases()` as a switch-case edge group
`google_adk_router.py`	Google ADK	`DecisionRouterAgent` hands each request to a sub-agent
`pydantic_ai_agent.py`	PydanticAI	an agent calling `decision_toolset()`, including a retry on an invalid call
`strands_agent.py`	Strands Agents	`DecisionRouter` plus `decision_tools()`
`mcp_client.py`	MCP	what an AI assistant sees: `opendecider mcp`'s tools called over stdio
`mastra/`	Mastra (TypeScript)	the MCP tools through Mastra's `MCPClient`

TypeScript¶

In typescript/examples/, for @opendecider/client. Each takes the model as its argument (an opendecider serve URL, or ollama:...) and runs without an LLM API key: a scripted stand-in plays the agent's LLM, and the tool calls and the guard are real. CI runs them against opendecider serve with opendecider-nano.

script	what it shows
`quickstart.mjs`	typed decisions, a `Router` with a fallback, and the `Guard`
`ai-sdk-agent.mjs`	Vercel AI SDK: the tools in a `generateText` loop, and `guardMiddleware` blocking an attack
`mastra-agent.mjs`	Mastra: the tools in an agent, and `GuardProcessor` stopping an attack through the tripwire

Colab notebook¶

Open in Colab

On a free NVIDIA T4, about 15 minutes: nano on a support ticket and an agent trace, many states in one call, confident automation on the typed-decisions test split (with a chart), speed at 1–50 questions per call, opendecider serve, small and small-td, and rebuilding the benchmark tables. The notebook in the repository is saved with the outputs of a real T4 run, so you can read it without running it.

Live demo¶

Try it in your browser

opendecider-nano on a basic CPU Space: edit the state and the questions and see every option's probability.