Create Context Graph
Scaffold domain-specific full-stack agents with graph memory and decision traces.
Per-dimension scores and reasoning
Evidence shows: CLI provides --dry-run option to preview generation without creating files, indicating least-privilege awareness; but no explicit permission tiers or sandboxing. User confirmation: interactive wizard exists, but non-interactive mode may execute directly without forced confirmation. Data flow transparency: README details data flow (e.g., NAMS write path hybrid shape), but no runtime data flow diagram or logs. Sensitive data handling: Claude Code connector claims to auto-redact secrets, but no implementation details. Dependency security: pyproject.toml lists dependencies, but no vulnerability scanning or lock files. External effects: CLI generates projects and may connect external services, but no explicit risk warnings. Rollback: --reset-database option exists, but no rollback mechanism for generated projects. Source attribution: README clearly labels Neo4j Labs project, but publisher unverified.
Self-consistency: README and pyproject.toml versions match (0.13.1), test files show fixes for v0.12.0, no obvious contradictions. Dependency availability: dependency list complete, but no lock files or version range verification. Failure messages: tests show failure handling, but no user-visible error message examples.
Audience and scenarios: README clearly targets developers, provides multiple usage scenarios (NAMS, self-hosted, custom domains). Capability boundaries: CLI options clearly defined, but no explicit limits on generated project scale or complexity. Trigger precision: CLI parameters clear, but no automated trigger mechanisms. Environment fit: supports multiple frameworks and deployment modes, but no complete containerized deployment instructions.
Information architecture: README well-structured, includes quick start, CLI reference, architecture. Install notes: provides uvx and npx commands, but no manual installation steps. Naming stability: project name and CLI command stable, but no naming convention documentation. Examples and FAQ: multiple examples provided, but no FAQ. Known limitations: README mentions NAMS write path limitations, but not comprehensive. License: Apache-2.0 license file complete. Versioning and changelog: pyproject.toml has version, but no CHANGELOG file. Maintenance responsibility: README states maintained by Neo4j staff, but no contribution guidelines.
Output usability: generated application includes full frontend and backend, ready to run. Marginal value: provides 27 domains and multiple frameworks, high value. Cost-benefit: free and open source, but requires API keys and infrastructure, moderate cost.
Claim traceability: README claims (e.g., test counts) lack specific test file references. Cross-source corroboration: no external verification. Fact-inference separation: README distinguishes feature descriptions and architecture, but does not explicitly label inferences.
- Publisher identity unverified; be cautious about supply chain risks.
- Dependencies not pinned, posing supply chain attack risk.
- Sensitive data handling (e.g., Claude Code connector) redaction implementation not detailed; needs review.
- Generated projects may connect external services; be aware of data exfiltration risks.
What does this agent do, and when should you use it?
Create Context Graph is an interactive CLI that generates domain-specific context-graph applications from a chosen domain and agent framework. Each generated project includes a FastAPI backend, a Next.js and Chakra UI v3 frontend, a Neo4j schema, fixtures, and optionally an MCP configuration. It uses neo4j-agent-memory v0.4 for short-term, long-term, and reasoning memory, with hosted NAMS as the default backend or self-hosted Bolt Neo4j as an alternative. The generated UI includes SSE chat, live tool-call timelines, an expandable graph view, entity details, document browsing, and decision-trace viewing. It fits teams that want an opinionated, runnable graph-agent application scaffold rather than a small embeddable library.
The create-context-graph CLI collects choices through its wizard or options such as --domain and --framework, then generates a domain ontology, project scaffold, agent tools, and system prompt. Generated backend/app code includes main.py, agent.py, routes.py, context_graph_client.py, memory.py, gds_client.py, and vector_client.py; the frontend includes ChatInterface.tsx, ContextGraphView.tsx, DecisionTracePanel.tsx, and DocumentBrowser.tsx. In self-hosted mode it produces Neo4j schema and GDS projection files plus fixtures.json, and make seed loads demo data; in NAMS mode, conversation-driven extraction populates hosted memory. Optional connectors import data from GitHub, Slack, Notion, Linear, Google Workspace, Claude Code, Claude AI, ChatGPT, and other listed services, with make import available for re-imports. The --with-mcp option generates MCP server configuration for Claude Desktop to query the same graph.
- A healthcare product team needs a runnable Neo4j demonstration with Patient, Diagnosis, and Treatment entities plus a chat interface.
- A software engineering team wants to import GitHub, Slack, or Linear data and generate domain tools and a graph UI around it.
- A team that needs native graph edges, GDS, or arbitrary Cypher writes wants to run a --self-hosted --demo project.
- A developer wants to construct a personal context graph from local Claude Code sessions or Claude AI and ChatGPT data exports.
- A prototype team has an unlisted industry and wants to describe the domain in plain English to generate an ontology, relationships, and agent tools.
What are this agent's strengths and limitations?
- One domain YAML drives schema, Pydantic models, agent tools, and visualization through Jinja2, keeping generated layers aligned.
- It offers both hosted NAMS memory and self-hosted Bolt Neo4j; the latter supports native edges, GDS, and arbitrary Cypher writes.
- The generated app includes more than chat: streaming tool-call visibility, graph exploration, document browsing, and decision traces.
- It documents 27 domains, eight agent frameworks, and imports from multiple SaaS services and conversation exports.
- It generates an independent full-stack project, so adoption involves Python, Node.js, environment configuration, and optionally Neo4j operations.
- The NAMS REST write surface is narrower: relationships are encoded as ccg-edges YAML in descriptions, and entity properties collapse into descriptions.
- NAMS does not support preferences or facts through REST, and its GDS endpoints and arbitrary Cypher writes return 501; those needs require self-hosting.
- Frameworks and connectors depend on their respective model credentials, service credentials, or local export files; end-to-end smoke tests additionally need a running Neo4j instance and API keys.
How do you install or deploy this agent?
For a development install, run: git clone https://github.com/neo4j-labs/create-context-graph.git && cd create-context-graph && uv venv && uv pip install -e ".[dev]". The generator requires Python 3.11+ and generated frontends require Node.js 18+. The default NAMS path needs MEMORY_API_KEY or --nams-api-key plus the API key required by the selected agent framework; the self-hosted path also needs Neo4j 5+ connection details.
How do you use this agent?
For hosted memory, run: uvx create-context-graph my-app --domain healthcare --framework strands --nams-api-key sk-nams-.... Then run cd my-app, add ANTHROPIC_API_KEY to .env, run make install and make start, and open http://localhost:3000. For self-hosted Neo4j with demo data, run: uvx create-context-graph my-app --domain healthcare --framework pydanticai --self-hosted --demo. In the generated directory, configure ANTHROPIC_API_KEY, then run make install, make docker-up, make seed, and make start. Use create-context-graph --list-domains to view the 27 built-in domains.
How does this agent compare with similar options?
Compared with simple RAG, the project combines short-term messages, a POLE+O long-term entity graph, and provenance-bearing reasoning memory. Its stated focus is structured graph reasoning and decision traceability rather than text retrieval alone.