Create Context Graph supports 8 agent frameworks. The framework choice only affects backend/app/agent.py in the generated project — everything else (FastAPI backend, Neo4j client, frontend, data, and schema) stays the same.
Choosing a framework
Interactive wizard
Run the CLI and select your framework at the relevant step:
CLI flag
The --framework flag accepts one of the eight keys listed below.
Supported frameworks
PydanticAI
Claude Agent SDK
OpenAI Agents SDK
LangGraph
CrewAI
Strands
Google ADK
Anthropic Tools
Key: pydanticaiType-safe agents with @agent.tool decorators and dependency injection via RunContext. PydanticAI validates tool inputs and outputs against Pydantic models. Full token-by-token text streaming via agent.run_stream().Required API key: ANTHROPIC_API_KEYStreaming: Full (token-by-token text + real-time tool calls) Key: claude-agent-sdkAnthropic’s own SDK with dict-based tool definitions and a bounded agentic loop (max 15 iterations). Full token-by-token text streaming via client.messages.stream().Required API key: ANTHROPIC_API_KEYStreaming: Full (token-by-token text + real-time tool calls) Key: openai-agentsOpenAI’s agent framework with @function_tool decorators and Runner.run(). Full token-by-token text streaming via Runner.run_streamed().Required API key: OPENAI_API_KEYopenai-agents requires OPENAI_API_KEY. The CLI will warn you if this variable is not set when you select this framework.
Key: langgraphLangChain’s graph-based agent runtime with @tool decorators and create_react_agent(). Full token-by-token text streaming via graph.astream_events().Required API key: ANTHROPIC_API_KEYStreaming: Full (token-by-token text + real-time tool calls) Key: crewaiMulti-agent framework with Agent, Task, and Crew abstractions using @tool decorators. Uses Anthropic as the LLM provider. CrewAI is synchronous and runs in a worker thread via asyncio.to_thread() with thread-safe async bridging.Required API key: ANTHROPIC_API_KEYStreaming: Tools only (real-time tool call events; text response arrives all at once) Key: strandsAgent framework using Anthropic models with @tool decorators. Strands is synchronous and runs in a worker thread via asyncio.to_thread() with thread-safe async bridging.Required API key: ANTHROPIC_API_KEYStreaming: Tools only (real-time tool call events; text response arrives all at once) Key: google-adkGoogle’s Agent Development Kit with FunctionTool and Gemini models. Full token-by-token text streaming via runner.run_async().Required API key: GOOGLE_API_KEYgoogle-adk requires GOOGLE_API_KEY. The CLI will warn you if this variable is not set when you select this framework.
Key: anthropic-toolsModular agent framework with a @register_tool registry and a bounded Anthropic API agentic loop (max 15 iterations). Full token-by-token text streaming via client.messages.stream().Required API key: ANTHROPIC_API_KEYStreaming: Full (token-by-token text + real-time tool calls) Framework comparison
Full streaming means token-by-token text delivery plus real-time tool call visualization. Tools only means tool call events stream in real-time, but the text response arrives all at once after the agent finishes. All frameworks use the same SSE (Server-Sent Events) protocol.
What changes between frameworks
Only one file varies between frameworks: backend/app/agent.py. This file contains:
- Agent initialization and model configuration
- Tool definitions generated from your domain’s
agent_tools ontology
- The
handle_message() async function that the FastAPI routes call
- For full-streaming frameworks: the
handle_message_stream() async function for SSE text streaming
Each framework template uses idiomatic patterns for that framework (decorators, classes, registries) but exposes the same interface to the rest of the application.
What stays the same
Regardless of which framework you choose, the generated project always includes:
- FastAPI backend (
main.py, config.py, routes.py, models.py) — identical across all frameworks
- Neo4j clients (
context_graph_client.py, gds_client.py, vector_client.py) — shared graph access layer
- Frontend (Next.js + Chakra UI v3 + NVL graph visualization) — framework-agnostic
- Data and schema (
cypher/schema.cypher, fixture data, ontology YAML)
- Infrastructure (
docker-compose.yml, Makefile, .env)
Generating a second project with a different framework
To compare frameworks side by side, re-run the CLI with a new output directory:
The two projects share the same schema, data, and frontend. Only agent.py and the agent-specific dependencies in pyproject.toml differ.
Framework-specific dependencies
The generated backend/pyproject.toml includes only the dependencies needed for the chosen framework. Examples:
pydanticai → pydantic-ailanggraph → langgraph, langchain-anthropicstrands → strands-agents[anthropic]google-adk → google-adk, nest-asyncio
All frameworks share common dependencies: fastapi, uvicorn, neo4j, pydantic, and python-dotenv.
Conversation memory uses local sentence-transformers embeddings by default — no OPENAI_API_KEY required. If you set OPENAI_API_KEY in your .env, the generated project automatically upgrades to OpenAI text-embedding-3-small embeddings.
