Neo4j CLI

neo4j-cli

Allows you to manage Neo4j resources

Allows you to manage Neo4j resources. Write operations require --rw.

Global Flags

Subcommands

Tips & Gotchas

• Example datasets: neo4j-cli dataset list surfaces a curated suggestion set, but any GitHub <owner>/<repo> carrying a relate.project-install.json manifest can be loaded — the suggestions are not a constraint. Load with the per-target verbs: docker load <owner/repo> (local container), desktop dbms load <owner/repo> (Neo4j Desktop), aura instance load <owner/repo> (new Aura instance). Loading defaults to the latest Neo4j; --database targets a database (default neo4j). Loads are writes, so pass --rw under an agent.
• Before generating ANY Cypher: run neo4j-cli query :schema --format toon first to discover the real labels, relationship types, and properties. Do not guess the schema. Read query-additions.md for the full schema-first workflow, parameters, embeddings, and Cypher 25 vs 5.
• The aura subcommand under neo4j-cli does NOT carry a nested skill group — install agent skills via neo4j-cli skill install at the top level.
• credential lives at the top level of neo4j-cli (not nested under aura) so credentials apply across every subcommand that talks to Aura.
• All read commands accept --format json|table|toon. Write commands print confirmation text only; pipe-friendly output requires explicit --format json where supported.
• Always pass --format toon on read commands — toon uses ~40% fewer tokens than JSON while encoding the same data, so default to it for every list/get/show command. Only use --format json when piping into a JSON-aware tool that requires it; only use --format table when the user explicitly asks for a human-readable table.
• Async resource operations (instance create/resize/destroy) accept --wait to block until the resource reaches a terminal state.
• The version: line in an installed SKILL.md reflects the binary (self-skill) or the cached catalog plugin.json.version (catalog skills). Run neo4j-cli skill check to detect drift; re-run skill install to refresh.
• neo4j-cli skill installs either the embedded self-skill OR any skill from the curated catalog at github.com/neo4j-contrib/neo4j-skills (Cypher, modeling, drivers, GraphRAG, GDS, Aura, …). The self-skill is addressable as self (canonical) or by the binary name neo4j-cli (alias) — both are reserved and shadowed catalog entries are rejected. The five leaves take an optional [skill-name] positional plus --agent <name> (case-insensitive) to scope to a single agent. Hard break: a positional that matches an agent name (e.g. claude-code) errors with unknown skill: <name>; did you mean '--agent <name>'? — no silent fall-through to the old positional-as-agent shape.
• install / remove accept --all: install --all writes the self-skill plus every catalog entry into every detected agent; remove --all removes every catalog skill from every detected agent and never touches the self-skill (which is preserved). install, list, and check accept --refresh to force a network fetch of the catalog before the operation; the new neo4j-cli skill refresh leaf does the same as a standalone command.
• Catalog cache lives under os.UserCacheDir()/neo4j-cli/skill-catalog/ with a 24h TTL — list/install/check auto-refresh when stale. Read commands (list, print, check) fall back to the cached copy on network failure with a stderr warning; cold cache plus a needed catalog install fails non-zero with a hint pointing at neo4j-cli skill refresh. print is strictly offline.
• If you pass an HTTP-style URI (e.g. http://host:7474) to query it is auto-rewritten to neo4j://host:7687 (and https:// to neo4j+s://host:7687); this command speaks the Bolt protocol. Use neo4j+ssc:// for self-signed certs.
• Write operations require --rw when running under an agent harness. The CLI auto-detects known agents (Claude Code, Codex, Cursor, Gemini CLI, Replit, …) via environment variables, so agents always need --rw for writes — interactive humans in a terminal do not. neo4j-cli query run additionally runs EXPLAIN over Bolt to detect write cypher when --rw is not set and blocks statements classified as writes before execution.
• Do NOT preemptively add --rw. Run write commands without it by default. If a command fails with this command writes; pass --rw to allow it, surface the error and ask the user once to confirm the write, then re-run with --rw — do not add it on your own.
• neo4j-cli query accepts multiple statements in one string, split on a ; at end of line (a mid-line ; is kept verbatim; the terminating ; is stripped). By default each runs in its own transaction, in order, failing fast on the first error; --atomic runs them all in one transaction that rolls back on any failure. When >1 statement runs, --format json emits a JSON array of result envelopes and --format table/toon print stacked blocks; a single statement renders unchanged.
• Use --param NAME:embed=<text> on neo4j-cli query to inject an embedding vector inline; the text is sent to the configured embedding provider and the resulting []float32 is bound to $NAME for both the EXPLAIN preflight and the real run. The sibling neo4j-cli query :embed [text] leaf computes a vector standalone (no Bolt connection opened).
• Embedding config (--embed-provider, --embed-model, --embed-base-url, --embed-dimensions) resolves with the same precedence as connection config: flag > OS env (NEO4J_EMBED_*) > .env walk-up > stored embed credential. API keys layer per provider: OPENAI_API_KEY (openai), HF_TOKEN (huggingface), GEMINI_API_KEY then GOOGLE_API_KEY (gemini) — all beating NEO4J_EMBED_API_KEY (generic) and the stored credential's api-key. Ollama needs no API key. Vertex AI uses Application Default Credentials (gcloud auth application-default login or GOOGLE_APPLICATION_CREDENTIALS), takes no --api-key, and requires --vertex-project + --vertex-location (stored on the credential — no per-query flag override).
• Linking dbms→embed: credential dbms add --embed-credential <name> or credential dbms set-embed <dbms-name> [embed-name] attaches an embed cred to a dbms cred so query --credential <dbms-name> --param NAME:embed=... picks up both connection and embed config in one selector. Removing an embed cred is non-cascading; stale links surface lazily at query time.
• Updating the CLI: run neo4j-cli update to self-update the binary in place. Use neo4j-cli update check to report whether a newer version is available without downloading (exits 1 when newer), and --pre-releases to opt into alpha/beta/rc tags (default is stable-only). When the binary lives under a known package-manager prefix (Homebrew, npm-global, pipx, uv tool) the command refuses to overwrite and prints the channel-correct upgrade command; pass --force to override. After a successful swap, any installed agent skill bundles are refreshed automatically — no manual skill install follow-up needed.
• Local Neo4j via Docker: neo4j-cli docker (create / list / get / start / stop / delete) shells out to the host docker CLI. Docker itself is the source of truth — managed containers carry the org.neo4j.cli.managed=true label plus a small set of metadata labels (edition, version, bolt-port, http-port, ephemeral); no separate state file is maintained. docker create defaults to the enterprise image with the evaluation license (NEO4J_ACCEPT_LICENSE_AGREEMENT=eval); pass --accept-license for the commercial license, or --edition community for the community image. Host ports default to 7474 (HTTP) and 7687 (Bolt); override with --http-port / --bolt-port. When the requested pair is taken, both ports are auto-incremented by the same offset until a free pair is found. If --name collides with an existing container or stored dbms credential, the chosen name is auto-suffixed (<name>-1, <name>-2, …) and the chosen name is logged to stderr. All --rw, --format json|table|toon, and --wait conventions from the aura tree apply equally here.

# Persistent flow: create stores a dbms credential, then query --credential connects with no further config
neo4j-cli docker create --name dev --wait --rw
neo4j-cli query --credential dev 'RETURN 1 AS n'
neo4j-cli docker delete dev --force --rw

• The generated password appears in docker create's rendered output; redirects and pipes capture it — pass --password <s> to pick your own, --no-print-password to keep the stored credential but omit the password from stdout (the stored credential still connects via --credential <name> — no plaintext needed; the password is not readable through the CLI), or --no-store-credential to suppress storage and rendering. To set a known password, run admin user set-password neo4j --new-password <s> --credential <name> --rw, then resync the stored credential with credential dbms remove <name> plus credential dbms add.

• Persistent volume mounts: --data-dir <host> bind-mounts at /data (persist DB across docker delete), --logs-dir <host> at /logs, --import-dir <host> at /import (for LOAD CSV). Paths support ~ and $VAR expansion; missing dirs are created at mode 0o755. All three are incompatible with --ephemeral. Neo4j adjusts ownership of the mounted dirs at container startup, so they appear under the container's neo4j UID on the host after first start.

# Persist data on the host so it survives delete + recreate
neo4j-cli docker create --name dev --data-dir ~/neo4j-dev/data --rw
neo4j-cli docker delete dev --force --rw
neo4j-cli docker create --name dev --data-dir ~/neo4j-dev/data --rw  # reuses the same data

• Ephemeral runs: neo4j-cli docker create --ephemeral shells docker run --rm, skips credential persistence, and emits a .env blob (NEO4J_URI / NEO4J_USERNAME / NEO4J_PASSWORD / NEO4J_DATABASE) to stdout. With --env-out-file <path> the blob is written to that path (mode 0600) and stdout stays silent so it can be piped into neo4j-cli query --env <path>. The env-file write goes through a temp file in the same directory + atomic rename; a pre-existing symlink at the target path is replaced by a regular file (the symlink is not followed). Docker auto-removes the container when it stops — nothing to delete.

# Ephemeral flow: throwaway container + env-file consumed by query --env
neo4j-cli docker create --name tmp --ephemeral --env-out-file /tmp/n.env --wait --rw
neo4j-cli query --env /tmp/n.env 'RETURN 1 AS n'
neo4j-cli docker stop tmp --rw

• Aura Agents: neo4j-cli aura agent (list / get / create / update / replace / delete / invoke) manages Aura Agents — LLM-backed assistants bound to an Aura database. --organization-id / --project-id honour the default workspace, identical to every other aura command. invoke is dual-mode: --format json returns the full server response verbatim (content blocks, usage, end_reason, errors), while the default table output joins the text content blocks and prints a single stats line Status: <S> | End reason: <ER> | Tool calls: <N> | Tokens: <req> req / <res> res / <total> total. HTTP 403 on invoke surfaces as agent invocation forbidden: agent may be disabled or private; an HTTP 200 body with type: "error" surfaces as agent invocation failed: <message>.

# List, create, and invoke an agent (workspace default already set)
neo4j-cli aura agent list --format toon
neo4j-cli aura agent create --name docs-bot --description "Docs assistant" --dbid <dbid> --tools '[{"type":"text2cypher","name":"ask","description":"Answer questions about the graph"}]' --rw
neo4j-cli aura agent invoke <agent-id> --input "hello" --rw

• --tools is a JSON array of tool objects. Every tool shares the envelope {type, name, description, config} (plus optional enabled) — name ≤64 chars, description ≤2000 chars. The type discriminator is camelCase (NOT snake_case): text2cypher, cypherTemplate, similaritySearch. The config object is type-specific; the four canonical shapes are:

[{"type":"text2cypher","name":"ask-graph","description":"Convert natural-language questions into Cypher and run them against the database."}]

[{"type":"cypherTemplate","name":"top-customers","description":"Return the top N customers by total order value.","config":{"template":"MATCH (c:Customer)-[:PLACED]->(o:Order) RETURN c.name AS name, sum(o.total) AS spent ORDER BY spent DESC LIMIT $limit","parameters":[{"name":"limit","data_type":"integer","description":"Maximum number of customers to return."}]}}]

[{"type":"similaritySearch","name":"doc-search","description":"Find docs most similar to the user question.","config":{"provider":"openai","model":"text-embedding-3-small","index":"docs_embedding_index","top_k":5}}]

[{"type":"similaritySearch","name":"doc-search-enriched","description":"Vector-search docs, then enrich each hit with its parent page.","config":{"provider":"openai","model":"text-embedding-3-small","index":"docs_embedding_index","top_k":5,"post_processing_cypher":"MATCH (node)<-[:HAS_CHUNK]-(page:Page) RETURN page.title AS title, page.url AS url, node.text AS chunk, score ORDER BY score DESC"}}]

Tool-config field reference (v2beta1):

• text2cypher — no config fields required.
• cypherTemplate.config — required template (Cypher string); optional parameters[] of {name, data_type, description} with data_type ∈ {string, number, boolean, integer}.
• similaritySearch.config — required provider (openai | vertexai), model (provider-compatible embedding model), index (vector-index name, ≤100 chars), top_k (1–100); optional dimensions (integer, when the provider/index needs an explicit size); optional post_processing_cypher — read-only Cypher appended to the vector search and run as a single query, with node and score exposed to the post-processing block (write queries are rejected server-side).