No ceiling
Every graph compiles to standalone Python (LangGraph) you own. Hit a limit? Drop into a Custom Code node — it round-trips verbatim.
Composable nodes
Input, Agent, Tools, Router, Evaluator, Memory, Responder/Revisor. Wire them into anything from a simple reflex to full Reflexion.
Run it live
Stream your agent in the playground. The built-in fake model needs no API key — bring OpenAI or Anthropic when you're ready.
From idea to ownable agent, in three moves.
Draw the graph
Start from a template, or drop nodes and connect them. The canvas is just state — nodes and edges.
Run & inspect
Chat with it in the playground. Watch tokens stream and tool calls fire, turn by turn.
Own the code
Open the Code view and copy idiomatic LangGraph. Zero Calypr dependency, zero lock-in.
The agent ladder, ready to run.
Eight archetypes from a single reflex up to tool-using ReAct and self-revising Reflexion. Load one, run it, read its code.
Simple reflex
Reacts to the latest input.
Model-based
Remembers the conversation.
Goal-based
Plans toward a goal.
Utility-based
Generates N, keeps the best.
Reflection
Critiques and revises itself.
Learning
Adapts from feedback.
ReAct
Reasons and calls tools in a loop.
Reflexion
Researches, then self-revises.
Export Python you’d actually merge.
Every node carries its own code generator, so the graph you draw projects to idiomatic LangGraph — grouped imports, a typed State, one function per node, the canonical tool loop. A round-trip test proves the generated module runs identically to the canvas.
"""ReAct Agent — generated by Calypr. Owns no Calypr dependency."""
def node_agent(state: State) -> dict:
"""Model-based agent: respond using the full conversation state."""
model = init_chat_model("gpt-4o-mini", temperature=0.7).bind_tools([web_search])
messages = state.get("messages") or []
reply = model.invoke([SystemMessage(content=system), *messages])
return {"messages": [reply]}
def build_graph():
graph = StateGraph(State)
graph.add_node("agent", node_agent)
graph.add_node("tools", ToolNode([web_search]))
graph.add_conditional_edges("agent", tools_condition, {"tools": "tools", END: "out"})
graph.add_edge("tools", "agent")
return graph.compile()From a prompt to an agent you own.
No black box. No lock-in. Draw it, run it, take the code with you.