ArcadeDB MCP Server: Query Graph, Document and Vector Data with AI

• Lists the databases you can open

  Returns the names of the databases on the ArcadeDB server that your sign-in is allowed to use.

  Try: “Which databases are on this server?”

  Read · list_databases

• Shows the types, properties and indexes

  Returns the full schema of one database: vertex, edge and document types, their properties, indexes and inheritance.

  Try: “Show me the schema of the mydb database”

  Read · get_schema

• Runs a read-only query

  Runs a SELECT, MATCH or other read query in SQL, Cypher, Gremlin, GraphQL or MongoDB syntax. The server checks the query and rejects it if it would change data. Results stop at 1,000 records unless you set a limit.

  Try: “List the 10 newest Person records in mydb”

  Read · query

• Runs a command that changes data

  Runs INSERT, UPDATE, DELETE, schema changes and other commands. Each kind needs its own permission in the MCP config, and all of them are off until you turn them on.

  Try: “Add a Person named Charlie to mydb”

  Delete · execute_command

• Shows server version and cluster state

  Returns the ArcadeDB version, server name, supported query languages, your databases and, with admin rights on a clustered setup, which node leads.

  Try: “What ArcadeDB version is running?”

  Read · server_status

• Peeks at the first records of each type

  Returns the first few stored records from up to 20 types, so the assistant can see real values before it writes a query. It is not a random sample.

  Try: “Show me 3 sample records from each type in mydb”

  Read · sample_records

• Finds nearest neighbours in a vector index

  Searches a dense or sparse vector index using a vector you supply. ArcadeDB does not create embeddings, so the assistant must pass one made by an embedding model. An optional filter narrows candidates.

  Try: “Find the 20 papers closest to this embedding in Paper[embedding]”

  Read · vector_search

• Searches a text index and ranks matches

  Searches a full-text index by index name or by type and properties. It supports required and excluded terms, phrases, prefixes, fuzzy terms and field filters, and returns records with a relevance score.

  Try: “Search Paper[abstract] for +graph +retrieval -sql”

  Read · full_text_search

• Combines vector, text and graph results

  Merges a vector search with an optional full-text search and an optional walk along graph edges into one ranked list. Each result says which of the three found it.

  Try: “Search papers for "graph retrieval" using this embedding and follow CITES edges two hops”

  Read · hybrid_search

• Creates or updates one vertex without duplicates

  Finds a vertex by the match keys you give, creates it if missing, then sets the properties you pass. Needs insert and update permission. Missing types are created automatically.

  Try: “Upsert a Company with name Acme and country DE”

  Write · upsert_entity

• Creates or updates one edge without duplicates

  Finds or creates the two end vertices by their match keys, then the edge of the type you name between them, and sets edge properties. Needs insert and update permission.

  Try: “Link Alice to Acme with a WORKS_AT edge since 2024”

  Write · upsert_relationship

• Starts the query profiler

  Starts recording every query with its run time and plan for up to 3,600 seconds (60 by default). It needs admin permission, and a second start leaves the first session running.

  Try: “Profile the server for two minutes”

  Write · profiler_start

• Stops the profiler and returns the results

  Ends the session and returns per-query counts, minimum, maximum, average and p99 times and plan step costs. The results are also saved to disk on the server. Needs admin permission.

  Try: “Stop the profiler and show the slowest queries”

  Write · profiler_stop

• Checks the profiler without stopping it

  Reports whether the profiler is recording and returns the current or latest results, including server metric snapshots. Needs admin permission.

  Try: “Is the profiler still running?”

  Read · profiler_status

• Lists server settings and their defaults

  Returns the server-level settings with current value, default and description. Passwords and the cluster token come back masked.

  Try: “What memory settings is the server running with?”

  Read · get_server_settings

• Changes a server setting while it runs

  Sets one server configuration key to a new value immediately and returns the old value. Whether it survives a restart depends on the setting. Needs admin permission.

  Try: “Set arcadedb.asyncWorkerThreads to 8”

  Write · set_server_setting

Watch out for

• Writes and deletes are one config change away

  execute_command can insert, update and delete data and change schema once the matching permission is switched on. Turn on only the permissions you need and use allowedUsers and per-database limits to narrow who and where. docs.arcadedb.com

• Password is sent as plain Basic auth over HTTP

  The documented HTTP setup uses a Basic Authorization header against a port that listens on all network interfaces by default. Put it behind TLS or a firewall, and prefer a limited user over root. github.com

• Stdio mode always signs in as root

  The stdio script needs the root password and runs MCP as root, though the permission switches still apply. It also turns MCP on for that process even if the config says disabled. docs.arcadedb.com

• Admin tools can change server settings

  set_server_setting and the profiler tools need the admin permission, which is off by default. Leave it off unless you want the assistant tuning the server. docs.arcadedb.com

• Tool descriptions prefer Cypher

  The query and command tools tell the assistant to use OpenCypher unless you ask for SQL. Say "use SQL" in your prompt if you want SQL. github.com

Security advisories