@graph-engineering-kits/graph-kit

Install a ready-to-run graph pattern kit into any project. One command, no clone, no course.

npx @graph-engineering-kits/graph-kit list
npx @graph-engineering-kits/graph-kit document-to-facts

Each kit is a working implementation of one pattern for building and operating knowledge graphs that agents can actually use — extraction, resolution, provenance, bounded retrieval, grounded checking, governance, and storage. Kits run against a small worked example that ships with them. No API keys and no external services are needed to try one.

From the Graph Engineering Crash Course.

Usage

graph-kit list                      show all 23 kits, grouped by tier
graph-kit <kit-name>                install that kit into the current directory
graph-kit new <name>                scaffold a blank kit from the template

| Option | | | --- | --- | | --dir <path> | Where to install (default: current directory) | | --tool <t> | claude, opencode, or both (default: whatever the kit ships) | | --force | Overwrite an existing directory | | --dry-run | Print what would be written, write nothing | | --help | Full help | | --version | Print the version |

The two tiers

Core kits (7) ship a Claude Code implementation, an OpenCode implementation, and a verifier subagent:

document-to-facts · alias-merge-with-trail · receipt-per-edge · task-scoped-retrieval · grounded-triple-checker · counter-metric-loop · sqlite-backed-graph

Extended kits (16) ship a Claude Code implementation plus a PORTING.md documenting what a port to another tool needs — config shape, file inputs, and how the skill-invocation model differs.

graph-kit list marks which is which. Asking for --tool opencode on an extended kit tells you so and points at the porting notes rather than failing silently.

What you get

document-to-facts/
├── SPEC.md                 what the pattern is for (category, stage, cost)
├── PATTERN.md              inputs, outputs, failure mode if you skip it
├── README.md               quickstart, expected output, troubleshooting
├── schema.example.json     the fixed type list
├── sample-input.md         the worked example's source document
├── .claude/
│   ├── skills/…/SKILL.md   the Claude Code skill
│   └── agents/…            the verifier subagent
└── opencode/               core kits only
    ├── opencode.json.example
    └── skills/…/SKILL.md

Read SPEC.md, then PATTERN.md, then README.md. The READMEs list expected output including what should be rejected — for extraction kits, seeing the right things dropped is the behavior worth checking.

The 23 patterns

| Category | Patterns | | --- | --- | | A · Extraction | document-to-facts code-change-to-graph conversation-to-claims | | B · Resolution | alias-merge-with-trail confidence-scored-dedup reversible-merge-audit | | C · Provenance | receipt-per-edge supersession-chain versioned-schema-log | | D · Subgraph | task-scoped-retrieval budget-capped-subgraph conflict-aware-bundle | | E · Checker | grounded-triple-checker contradiction-detector early-victory-guard | | F · Governance | counter-metric-loop arbitration-edge audit-loop anchor-and-freeze | | G · Storage | sqlite-backed-graph file-graph-for-small-teams postgres-backed-graph neo4j-at-scale |

postgres-backed-graph and neo4j-at-scale are reasoning walkthroughs — they do not connect to a live database. sqlite-backed-graph ships a real schema.sql that loads as-is.

Requirements

Node 18+. Claude Code or OpenCode to run a kit's skill.

Contributing

Kits live at starters/ in the course repo — that is the single source of truth. This package bundles a copy at publish time. Fixes and new patterns go through CONTRIBUTING.md.

MIT.