Graphify.md is a community-built project, not affiliated with NVIDIA, that turns NemoClaw into a traversable knowledge graph for source-grounded lookup and dependency traversal. Built for platform engineers, agent developers, and docs teams who need deterministic answers about runtimes, policy, security, and deployment — with every result traceable to declared relationships rather than model inference.
"The graph doesn't guess — it traverses. Every answer traces to a declared edge."
55 nodes · 74 edges · hover to explore · colored by layer
Signal from 123 GitHub issues, HN threads 47427027 + 47435066, and hands-on walkthroughs.
Agents forget tool schemas after loop iterations. The model re-infers NemoClaw's architecture on every query instead of reading declared structure.
OpenShell makes token spend visible per agent for the first time. The next question is how to reduce it. CKG is that answer.
NVIDIA's own OpenShell graph names this: the missing layer between the runtime policy engine and the structured knowledge agents need. We filled it.
The full NemoClaw stack as a typed dependency graph. Pre-structured, traversable, deterministic. Served over MCP. No inference at query time.
get_prerequisites("ManagedMCPServer") → ManagedMCPServer ├─ [ENABLES] NemoClaw ├─ [REQUIRES] NetworkPolicy ← root └─ [REQUIRES] L7Proxy ├─ [IMPLEMENTS] OpenShell └─ [REQUIRES] SharedGateway └─ [IMPLEMENTS] InferenceProvider 269 tokens · declared edges only RAG: ~2,982 tokens · probabilistic
query_ckg("ProgressiveToolDisclosure") ← [IMPLEMENTS] OpenClaw ← [IMPLEMENTS] Hermes ← [IMPLEMENTS] LangChain_Deep_Agents All three runtimes share this mechanism. RAG returns three separate docs. The graph knows — it's a declared edge.
Every edge in the graph was extracted from a source document and given a type. There are no probabilistic weights, no cosine similarity scores, no confidence intervals. An edge either exists — declared, typed, sourced — or it doesn't. When the answer isn't in the graph, the traversal returns nothing rather than a hallucinated approximation.
Hard prerequisite. A cannot function without B. Used for build-time and runtime dependencies — e.g., OpenShell REQUIRES L7Proxy.
Capability unlock. A makes B possible but B can exist without A in degraded form — e.g., ManagedMCPServer ENABLES NetworkPolicy.
Concrete instantiation of an abstract concept — e.g., OpenClaw IMPLEMENTS ProgressiveToolDisclosure. Useful for finding all runtimes sharing a mechanism.
Conceptual proximity without a dependency direction. Used for cross-layer concepts — e.g., SecurityHardening RELATES_TO Sandbox.
Confidence scores exist to compensate for retrieval uncertainty. RAG returns chunks that might be relevant, scored by vector distance. A 0.87 cosine score doesn't tell you if the relationship is real — it tells you the embedding is similar. CKG has no retrieval step, so there's nothing to score.
The edge type is the confidence signal. REQUIRES means the relationship is load-bearing and sourced. RELATES_TO means it's real but weaker. Absent means the graph makes no claim — which is different from "low confidence." A missing edge is not a soft no; it's silence from a source-grounded system.
"CorporateCA is probably used for identity management... (similarity: 0.81)"
The score is on the chunk retrieval, not the factual claim. The claim itself is unverified.
"CorporateCA is anchored at image build for TLS interception proxy traversal."
No score. Declared edge. Traces to security hardening source doc. Ask for the edge and it returns it verbatim.
30 questions from real GitHub issues · CPU only · Ollama · temperature 0 · seed 42
What are the three agent runtimes supported by NemoClaw and how is each activated?
How does CorporateCA integrate into NemoClaw's security chain?
What enterprise manufacturing deployment uses NemoClaw via the FOX Blueprint?
Settings → Connectors → Add connector → paste URL above.
// claude_desktop_config.json { "mcpServers": { "nemoclaw": { "command": "uvx", "args": ["ckg-nvidia-nemoclaw"] } }}
| Tool | Description |
|---|---|
| ask_nemoclaw(question) | Natural language query — auto-detects concept, traverses the relevant subgraph |
| query_ckg(concept, depth) | Typed subgraph around a specific concept (1–5 hops) |
| get_prerequisites(concept) | Full upstream prerequisite chain — every dependency in order |
| search_concepts(query) | Fuzzy search across all 55 concepts |
| list_domains() | Available domains and node/edge counts |
55 nodes · 74 edges · edge types: REQUIRES · ENABLES · IMPLEMENTS · RELATES_TO
OpenClaw · Hermes · LangChain Deep Agents
OpenShell · NVIDIA Agent Toolkit · OpenShell TUI · CLI
SharedGateway · vLLM · Ollama · NIM Local · ModelRouter
NetworkPolicy · PolicyTier (Restricted/Balanced/Open) · PolicyPreset
L7Proxy · Landlock LSM · CONNECT Proxy · CorporateCA · SecurityHardening
Progressive Tool Disclosure · Context Compaction · Heartbeat · Snapshots · Shields
DGX Spark · DGX Station · macOS Apple Silicon · WSL2 · Brev
FOX Blueprint · MoMClaw (Foxconn) · Nemotron 3 Ultra · Agent Harness
Signal from 123 open GitHub issues · HN threads 47427027 + 47435066 · hands-on walkthroughs.
OpenShell surfaced per-agent token spend for the first time. Developers immediately ask how to reduce it — not just see it. CKG is the answer to that follow-up.
Dominant skepticism: "even local mode demands an API key." HN thread 47427027 has 83 upvotes on this complaint. CKG runs on pure Python BFS — no model, no API, no cloud.
Three runtimes (OpenClaw/Hermes/LangChain) share it, docs cover each separately. Developers don't realize it's the same mechanism. The graph makes the shared edge visible.
Every node and edge traces to one of these. No probabilistic inference — declared relationships only.
| Type | Source | Coverage |
|---|---|---|
| Official | docs.nvidia.com/nemoclaw/latest/ | Core platform — agent runtimes, OpenShell, security, inference routing, policy |
| Official | FOX Blueprint docs | MoMClaw (Foxconn) manufacturing deployment, enterprise patterns |
| Official | Nemotron 3 Ultra ecosystem | DGX Spark/Station, ModelRouter, NIM Local integration |
| Official | Security hardening guide | LandlockLSM, L7Proxy, CorporateCA, CONNECT Proxy, Sandbox chain |
| Official | Managed MCP Server docs | ManagedMCPServer, NetworkPolicy, PolicyTier, PolicyPreset, messaging bindings |
| Official | Agent features reference | ProgressiveToolDisclosure, Context Compaction, Heartbeat, Snapshots, Shields |
| Community | github.com/Yarmoluk/ckg-nvidia-nemoclaw/issues | 123 issues — phantom-ready, local API key, SharedGateway conflicts |
| Community | HN 47427027 · HN 47435066 | Local inference gap (83↑) · token burn visibility · multi-sandbox use cases |
| Dataset | huggingface.co/datasets/danyarm/ckg-benchmark | KRB v0.6.2 — 7,928 queries, 30 NemoClaw-domain questions |
| Benchmark | github.com/Yarmoluk/ckg-benchmark/paper/main.pdf | Full methodology, F1 0.471, RAG/GraphRAG baselines, reproduction |
| A/B report | ckg-ab-test/results/REPORT_nemoclaw.md | 30-question NemoClaw run · phi4-mini + nemotron-mini · CPU · full raw results |
7,928 queries · 5-hop F1: 0.772 (CKG) vs 0.170 (RAG) · dataset public on HuggingFace · run it yourself
No API key. No GPU. No inference at query time. Add the URL connector in claude.ai or run locally with one command.
97 domains available · ckg-mcp serves the full Graphify.md ecosystem · Free tier included