Metadata-Version: 2.4
Name: ckg-nvidia-nemoclaw
Version: 0.3.0
Summary: NVIDIA NemoClaw as a traversable knowledge graph — sandboxed AI agent infrastructure, OpenShell runtime, inference routing, and network policy — MCP-native
Project-URL: Homepage, https://graphifymd.com
Project-URL: Repository, https://github.com/Yarmoluk/ckg-nvidia-nemoclaw
Project-URL: Bug Tracker, https://github.com/Yarmoluk/ckg-nvidia-nemoclaw/issues
Project-URL: Benchmark, https://github.com/Yarmoluk/ckg-benchmark/blob/main/paper/main.pdf
Author-email: Daniel Yarmoluk <daniel.yarmoluk@gmail.com>
License: MIT
Keywords: ai-agents,ckg,hermes,inference-routing,knowledge-graph,mcp,nemoclaw,network-policy,nvidia,openclaw,openshell,sandbox
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Security
Requires-Python: >=3.11
Requires-Dist: mcp[cli]>=1.0.0
Provides-Extra: serve
Requires-Dist: fastmcp>=0.4.0; extra == 'serve'
Requires-Dist: starlette>=0.41.0; extra == 'serve'
Requires-Dist: uvicorn>=0.30.0; extra == 'serve'
Description-Content-Type: text/markdown

# ckg-nvidia-nemoclaw

NVIDIA NemoClaw as a traversable knowledge graph — MCP-native.

**55 nodes · 74 edges** covering the full NemoClaw stack: OpenShell runtime, OpenClaw, Hermes, LangChain Deep Agents, inference routing, network policy, security architecture (L7 proxy, Landlock LSM, CONNECT proxy), progressive tool disclosure, managed MCP servers, deployment paths (local, Brev), FOX Blueprint, and the Nemotron 3 Ultra agent harness ecosystem.

Instead of retrieving docs, the agent traverses declared typed relationships. Every answer traces to a declared edge — the graph doesn't guess, it traverses.

**Benchmark (KRB v0.6.2 locked):** CKG F1 0.471 · RAG baseline 0.123 · GraphRAG baseline 0.120 · 269 tokens/query vs 2,982 (11× reduction)

## Install

```bash
pip install ckg-nvidia-nemoclaw
```

## Run as MCP Server

```bash
uvx ckg-nvidia-nemoclaw
```

## Claude Desktop Config

```json
{
  "mcpServers": {
    "nemoclaw": {
      "command": "uvx",
      "args": ["ckg-nvidia-nemoclaw"]
    }
  }
}
```

## Tools

| Tool | Description |
|------|-------------|
| `ask_nemoclaw(question)` | Auto-detect and traverse the relevant subgraph |
| `query_ckg(concept, depth)` | Typed subgraph around a specific concept |
| `get_prerequisites(concept)` | Full upstream prerequisite chain |
| `search_concepts(query)` | Fuzzy search concepts by name or keyword |
| `list_domains()` | List available domains and node/edge counts |

## Coverage

- **Agent runtimes:** OpenClaw · Hermes (Nous Research) · LangChain Deep Agents Code
- **Platform:** OpenShell · NVIDIA Agent Toolkit · CONNECT proxy · L7 proxy · Landlock LSM
- **Inference:** inference.local routing · SharedGateway · vLLM · Ollama · Local NIM · ModelRouter
- **Policy:** NetworkPolicy · PolicyTier (Restricted/Balanced/Open) · PolicyPreset bundles
- **Security:** 5-layer architecture (Network/Filesystem/Process/Gateway Auth/Inference)
- **Features:** Progressive Tool Disclosure · Context Compaction · Managed MCP Servers · Snapshots · Shields
- **Deployment:** Local CLI · Brev CLI · Brev Web UI · DGX Spark · macOS Apple Silicon · WSL2
- **Ecosystem:** FOX Blueprint · MoMClaw (Foxconn) · Nemotron 3 Ultra · Agent Harness ecosystem

## Benchmark

[github.com/Yarmoluk/ckg-benchmark/blob/main/paper/main.pdf](https://github.com/Yarmoluk/ckg-benchmark/blob/main/paper/main.pdf)

---

Built by [Graphify.md](https://graphifymd.com) · patent pending
