Loveliness

Loveliness

Loveliness is a clustered graph database with a Cypher query layer. This skill drives it through the loveliness MCP server (run by loveliness-mcp or loveliness mcp).

Tools

Read-only:

• schema — node + edge tables with property names, types, and primary keys. Cached 30s.
• cypher_read — run a Cypher read query. Rejects writes (CREATE/MERGE/SET/DELETE/DROP/REMOVE/LOAD/COPY/ALTER/DETACH) and mutating CALL procedures.
• cluster_status — leader, peers, per-shard assignment, registered schema.
• list_databases — databases in the cluster catalog with state, shard count, and creation time.
• describe_schema — schema joined with annotations in one payload. Use this first when you need to plan a query — it shows what each table holds and the example queries the operator left behind.
• list_annotations / get_annotation — read schema annotations (descriptions, query examples, tags) for an element.

Write (skipped when the server is launched with --readonly):

• cypher_write — Cypher writes / DDL.
• create_node_table, create_edge_table, drop_table — typed schema management. Builds the DDL from structured input and runs it; bust the schema cache so the next schema call is fresh. Prefer these over hand-built cypher_write DDL.
• create_database, drop_database, start_database, stop_database — database lifecycle. Wraps /admin/cypher. Must hit the leader (call cluster_status if you get NOT_LEADER).
• admin_cypher — escape hatch for raw admin commands.
• bulk_nodes, bulk_edges — synchronous CSV load. Provide either inline csv_data or a host-readable csv_path.
• ingest_nodes, ingest_edges — async CSV load. Returns a job_id to poll with ingest_status.
• set_annotation, delete_annotation — attach or remove a description, query examples, and tags on a schema target. Latest-wins (a SET replaces the previous body). Replicated through Raft alongside the schema, so annotations survive restart and follow the data.

Resources (read-only blobs the client can fetch directly):

• loveliness://schema — same payload as schema.
• loveliness://cluster — same payload as cluster_status.

Working pattern

1. Always start with describe_schema. It returns the schema and the cluster's annotations (descriptions, query examples) joined. If describe_schema is missing, fall back to schema then list_annotations.
2. Use params, not string interpolation. The query field accepts $name placeholders that map to the params object. Loveliness inlines parameters as escaped Cypher literals, so user-controlled strings stay safely quoted. Don't build queries with string concatenation.
3. Pick the right write tool:
   • Schema (DDL): create_node_table / create_edge_table / drop_table over cypher_write. They validate identifiers and bust the schema cache.
   • Database lifecycle: list_databases, create_database, drop_database. Skip if the cluster is single-DB.
   • One-off data mutation: cypher_write.
   • Synchronous CSV ≤100K rows: bulk_nodes / bulk_edges.
   • Async / large CSV: ingest_nodes / ingest_edges, then poll ingest_status every few seconds until the job is done.
4. When something fails: call cluster_status first. A 503 or "shard unavailable" error usually means a peer is down or the shard map is mid-rebalance.

Common shapes

Read by primary key:

MATCH (p:Person {name: $name}) RETURN p

With params: {"name": "Alice"}.

Write a node:

CREATE (p:Person {name: $name, age: $age}) RETURN p

Define a node table:

create_node_table(table="Person", properties=[
  {"name": "name", "type": "STRING", "primary_key": true},
  {"name": "age",  "type": "INT64"}
])

Define an edge table:

create_edge_table(table="KNOWS", from="Person", to="Person",
                  properties=[{"name": "since", "type": "DATE"}])

Bulk-load nodes (synchronous):

bulk_nodes(table="Person", csv_data="name,age\nAlice,30\nBob,25\n")

Async edge load:

ingest_edges(rel_table="KNOWS", from_table="Person", to_table="Person", csv_path="/data/knows.csv")
→ {"job_id": "abc123", "status": "queued"}

ingest_status(job_id="abc123")
→ {"job": {"status": "running", "loaded": 1200000, ...}}

Annotate a table so future LLM turns know what it holds:

set_annotation(
  target="db:default/table:Person",
  description="People in the social graph. Each row is a unique human. PII.",
  examples=[
    {"title": "by name",  "query": "MATCH (p:Person {name: $name}) RETURN p", "params": {"name": "Alice"}},
    {"title": "top by friends", "query": "MATCH (p:Person)-[:KNOWS]->(f) RETURN p, count(f) AS n ORDER BY n DESC LIMIT $k", "params": {"k": 10}}
  ],
  tags=["core", "pii"]
)

Annotation targets

set_annotation takes a hierarchical target string:

• cluster — top-level "what is this cluster" description.
• db:<name> — what a database holds.
• db:<name>/table:<name> — node table description + example queries.
• db:<name>/table:<name>/property:<name> — column-level note (units, format, deprecated, etc.).
• db:<name>/edge:<name> — relationship table.
• db:<name>/edge:<name>/property:<name> — edge column.
• db:<name>/saved_query:<id> — a named, parameterized query template.

For single-database clusters use db:default. Annotations are latest-wins (a SET replaces the body) and replicated through Raft like the schema and catalog — they survive restarts and follow snapshots.

Common errors

• cypher parse error — Cypher didn't parse. Re-read schema to confirm table/column names; LadybugDB is case-sensitive.
• cypher_read rejects write statements — switch to cypher_write, or rephrase as a MATCH ... RETURN if you really meant a read.
• schema propagation in progress, retry — another node is broadcasting a DDL change. Wait a moment and retry.
• shard unavailable — call cluster_status; the peer for that shard is down or replaying.
• query exceeded timeout — narrow the query (add LIMIT, tighter WHERE) or relaunch the server with a longer --timeout.

Don't

• Don't paste user-controlled input into the query string. Use params.
• Don't loop ingest_status faster than every 1–2 seconds — async ingest is throughput-oriented, not real-time.
• Don't call cypher_write for bulk data. The bulk endpoints are an order of magnitude faster.
• Don't call mutating CALL procedures (drop_table, export_database, …) through cypher_read; the server rejects them. Use cypher_write if you really intend a mutation.

Configuration

The server is configured by env vars / flags at launch:

• LOVELINESS_URL (default http://localhost:8080)
• LOVELINESS_TOKEN — bearer token for auth-enabled clusters
• LOVELINESS_TIMEOUT (default 30s)
• LOVELINESS_READONLY=true — register read-only tools only

These are set in the MCP client config (e.g. ~/.claude.json), not at tool-call time.