RushDB MCP Server
Official# RushDB MCP Server
### Connect MCP clients to RushDB, the database for AI.
Connect RushDB to any MCP-compatible AI client and your agent gains persistent, structured memory — semantically searchable and graph-traversable, without managing embeddings or schemas.
Works with **ChatGPT**, **Claude.ai**, Claude Desktop, Cursor, Windsurf, Gemini CLI, and any MCP-compatible tool.
---
## Two ways to connect
### Remote OAuth (no installation)
For **ChatGPT**, **Claude.ai**, and other web-based AI clients — connect directly without installing anything locally:
**MCP endpoint:** `https://mcp.rushdb.com/mcp`
Your client redirects to RushDB for OAuth sign-in and receives a scoped token. No API key setup required.
**ChatGPT:**
1. Go to **Settings → Connectors → Add connector**
2. Enter `https://mcp.rushdb.com/mcp`
3. Sign in with your RushDB account
**Claude.ai:**
1. Go to **Settings → Integrations → Add integration**
2. Enter `https://mcp.rushdb.com/mcp`
3. Authorize with your RushDB account
---
### Local (RushDB Cloud API key)
For Claude Desktop, Cursor, Windsurf, VS Code, and other stdio MCP clients:
**1. Get an API key** at [app.rushdb.com](https://app.rushdb.com)
**2. Add to your MCP client config:**
```json
{
"mcpServers": {
"rushdb": {
"command": "npx",
"args": ["@rushdb/mcp-server"],
"env": {
"RUSHDB_API_KEY": "your-api-key-here"
}
}
}
}
```
Place this in:
- **Claude Desktop** — `~/Library/Application Support/Claude/claude_desktop_config.json`
- **Cursor** — MCP settings in the Cursor config panel
- **Windsurf** — `~/.codeium/windsurf/mcp_config.json`
**3. That's it.** The agent now has 43 tools for storing, querying, searching and traversing structured memory.
> `RUSHDB_API_URL` is optional and defaults to `https://api.rushdb.com/api/v1`, the RushDB Cloud API.
### Local standalone engine (no Cloud account)
Run the Rust engine, then point a local **stdio** MCP client at it. This mode calls the
public engine protocol directly; it does not use a RushDB Cloud API key, dashboard, or OAuth.
```bash
RUSHDB_TOKEN=local-development-token rushdb up --data-dir ./rushdb-data
```
```json
{
"mcpServers": {
"rushdb": {
"command": "npx",
"args": ["-y", "@rushdb/mcp-server"],
"env": {
"RUSHDB_MODE": "direct",
"RUSHDB_ENGINE_URL": "http://127.0.0.1:3000",
"RUSHDB_ENGINE_TOKEN": "local-development-token"
}
}
}
}
```
Direct mode offers every tool whose work the engine does: schema discovery, record reads and writes (with
ordering, projections and aggregations), relationships, labels and properties, bulk import and CSV export, text
indexes and keyword search, and external vector indexes with vector search by a `queryVector` you supply.
The RushDB Cloud tools are not listed and are refused if called: `smartSearch` (natural-language search),
the relationship-pattern tools, and anything that needs managed embeddings (a text `query` for vector search,
managed indexes). The MCP HTTP transport is unavailable in direct mode because its OAuth flow belongs to
RushDB Cloud; run the server through stdio on the same machine as the client.
---
## Compatibility
| Package | RushDB 3.x and RushDB Cloud | RushDB 2.x (legacy) |
|---|---|---|
| `@rushdb/javascript-sdk` | 3.x | 2.11.0: `npm i @rushdb/javascript-sdk@2` |
| `rushdb` (PyPI) | 3.x | 2.10.0: `pip install "rushdb<3"` |
| `@rushdb/mcp-server` | 3.x | 2.13.0: `npx @rushdb/mcp-server@2` |
---
## What an agent can do
After connecting, ask Claude or your agent naturally:
> "Remember that we decided to use Clerk for auth, replacing Auth0, on 2026-04-10."
The agent calls `createRecord` to store a structured memory. Later:
> "What auth decisions have we made?"
The agent calls `getSchemaMarkdown` to learn what is stored, then `findRecords`, `textSearch` or `vectorSearch`.
> "Show me everything related to our auth system."
The agent calls `findRelationships` and `findRecords` with relationship filters to traverse the graph of connected records.
---
## Available tools
52 tools, named exactly as below. In direct mode the twelve RushDB Cloud tools are not listed. The full reference, with
inputs and safety levels, is at [docs.rushdb.com/3.x/learn/reference/mcp/tools](https://docs.rushdb.com/3.x/learn/reference/mcp/tools).
### Discovery and properties
| Tool | What it does |
| --- | --- |
| `getSchemaMarkdown` | The whole graph schema as compact Markdown; call once at session start |
| `getSchema` | Structured schema JSON, with property IDs and vector indexes |
| `findLabels` | List or filter record labels and their counts |
| `findProperties` | Property names, types, IDs and record counts |
| `findPropertyById` | Metadata for one property |
| `propertyValues` | Numeric or date ranges, or distinct values; with `query`, the matching values and their counts (`match: similar` for near spellings; `domainOnMiss` lists a small field's values when nothing matches) |
| `searchValues` | Where a value from the question is stored: matches across every property, with label, property and record count |
| `getSchemaSummary` | Every label's properties with distinct counts, ranges and the values of small fields, in one live read |
| `resolveNames` | The real label or property behind a name that is not in the schema (`patients` → `Patient`) |
| `getJoinCandidates` | How labels reference each other, from property names and exact value containment, with evidence (RushDB Cloud only) |
| `deleteProperty` | Remove a property and its values from all records |
| `getSearchQuerySpec` | The complete SearchQuery syntax reference |
| `getQueryBuilderPrompt` | The built-in query builder prompt (for clients without MCP Prompts support) |
### Records
| Tool | What it does |
| --- | --- |
| `findRecords` | Search records; ordering, projections, grouping and aggregations |
| `findOneRecord` | The first matching record |
| `findUniqRecord` | Exactly one matching record; fails on zero or several |
| `getRecord` | One record by ID |
| `getRecordsByIds` | Several records by ID |
| `createRecord` | Insert or upsert one record |
| `updateRecord` | Update some fields of a record |
| `setRecord` | Replace all fields of a record |
| `deleteRecord` / `deleteRecordById` | Delete one record by ID |
| `bulkCreateRecords` | Insert or upsert many records with one label |
| `bulkDeleteRecords` | Delete the records matching a query |
| `exportRecords` | Export matching records as CSV |
### Relationships
| Tool | What it does |
| --- | --- |
| `findRelationships` | Find relationships by type, properties or endpoint records |
| `attachRelation` | Create relationships between existing records |
| `updateRelation` | Change or remove properties on existing relationships; `types` stores an ISO string as a datetime (RushDB 3.0.4+) |
| `detachRelation` | Remove relationships: every type unless `relationType` is given, both ways unless `direction` is |
### Vector, keyword and hybrid search
| Tool | What it does |
| --- | --- |
| `findEmbeddingIndexes` | List embedding indexes and their status |
| `createEmbeddingIndex` | Create an embedding index (on a self-hosted engine, an external one with `dimensions`) |
| `getEmbeddingIndexStats` | How many records an index covers and has vectors for |
| `upsertEmbeddingVectors` | Write your own vectors to an external index |
| `deleteEmbeddingIndex` | Delete an embedding index and its vectors |
| `vectorSearch` | Rank records by vector similarity after exact filtering; add `text` for hybrid search |
| `semanticSearch` | Deprecated alias for `vectorSearch` |
| `listTextIndexes` | List text indexes and their status |
| `createTextIndex` | Declare a BM25 text index over one property, or `*` for every string |
| `deleteTextIndex` | Drop a text index |
| `textSearch` | Keyword (BM25) search after exact filtering |
### RushDB Cloud only
| Tool | What it does |
| --- | --- |
| `smartSearch` | Turn a natural-language request into SQL and run it |
| `listRelationshipPatterns` | List inferred relationship patterns and analysis status |
| `analyzeRelationshipPatterns` | Queue schema analysis for relationship suggestions |
| `approveRelationshipPattern` | Approve and apply a suggested pattern |
| `ignoreRelationshipPattern` | Ignore a suggested pattern |
| `deleteRelationshipPattern` | Delete a pattern and optionally its materialized relationships |
### Connections and imports (RushDB Cloud only)
Data synced from SaaS tools (CRM, billing, support, issue tracking, docs, spreadsheets) and background
file imports. The agent checks freshness before answering about recent data and offers a sync rather
than starting one unasked.
| Tool | What it does |
| --- | --- |
| `listConnections` | The project's connections: status, schedule, last success and error, per-stream label, record count and last sync |
| `getConnection` | One connection by id (credentials are never returned) |
| `syncConnection` | Start a sync now, of every enabled stream or the ones named (`full` re-reads everything); returns the run. Writes |
| `listImports` | Import runs (file imports and connection syncs), filterable by `kind` and `connectionId` |
| `getImport` | One run's status, per-file or per-stream progress, error sample and summary |
### Utilities
| Tool | What it does |
| --- | --- |
| `helpAddToClient` | Setup instructions for your MCP client |
Writes accept an optional `transactionId`.
---
## Built-in query builder prompt
RushDB exposes a system prompt via the MCP Prompts API that teaches the agent to query safely — discover labels and properties first, then construct queries.
**How it works:**
1. Client calls `prompts/list` → finds `rushdb.queryBuilder`
2. Client calls `prompts/get` → injects it as the session system message
3. Agent now does discovery-first queries automatically
Most MCP clients handle this at session start. If your client doesn't support MCP Prompts, call `getQueryBuilderPrompt` and inject the result as your system message.
---
## Environment variables
These apply to local (API key) mode. Remote OAuth connections do not use environment variables.
| Variable | Required | Description |
| ---------------- | -------- | ------------------------------------------------------- |
| `RUSHDB_API_KEY` | yes | Your RushDB API key |
| `RUSHDB_API_URL` | no | API base URL (default: `https://api.rushdb.com/api/v1`) |
For standalone direct mode, use these instead:
| Variable | Required | Description |
| --- | --- | --- |
| `RUSHDB_MODE` | yes | Set to `direct`. |
| `RUSHDB_ENGINE_URL` | yes | Standalone engine base URL, for example `http://127.0.0.1:3000`. |
| `RUSHDB_ENGINE_TOKEN` | yes | Token configured for the standalone engine. |
| `RUSHDB_ENGINE_SCOPE` | no | Tenant on an engine started with `--multitenant`. |
---
## Registry & autodiscovery
- `mcp.yaml` at the repo root enables autodiscovery by MCP registries
- `glama.json` at the repo root contains Glama metadata
- Package: `@rushdb/mcp-server`, command: `npx`
---
## Development
```bash
git clone https://github.com/rush-db/mcp-server
cd mcp-server
npm install
npm run build
```
### Troubleshooting: `rushdb-mcp: command not found` / MCP error `-32000`
This only happens when the MCP client launches the server with a **working directory inside this repo** (e.g. an `.mcp.json` at the repo root used by Claude Code / Cursor). This repo *is* `@rushdb/mcp-server`, so an unversioned `npx @rushdb/mcp-server` (even with `-p`) can resolve to the local package, whose `rushdb-mcp` bin isn't linked in `node_modules/.bin`. The process exits before the MCP handshake, which the client reports as `-32000`.
The published package is unaffected — the config above works as-is **outside** the repo. To run from inside the repo, **pin a version** so npx fetches from the registry instead of resolving the local package:
```json
{
"mcpServers": {
"rushdb": {
"command": "npx",
"args": ["-y", "-p", "@rushdb/mcp-server@latest", "rushdb-mcp"],
"env": { "RUSHDB_API_KEY": "your-api-key-here" }
}
}
}
```
Alternatively, run the local build directly (uses your in-repo code, must be built first with `npm run build`):
```json
{
"mcpServers": {
"rushdb": {
"command": "node",
"args": ["build/index.js"],
"env": { "RUSHDB_API_KEY": "your-api-key-here" }
}
}
}
```
---
## License
Apache 2.0
TDQS
Scored across 48 tools
Many tools overlap heavily: six-plus schema-discovery tools (getSchemaMarkdown, getSchema, getSchemaSummary, findLabels, findProperties, findPropertyById), duplicate deletes (deleteRecord vs deleteRecordById), near-duplicate reads (findRecords/getRecord/getRecordsByIds/findOneRecord/findUniqRecord), overwrite vs partial semantics (setRecord vs updateRecord), and an explicitly deprecated alias (semanticSearch for vectorSearch). The descriptions do try to disambiguate, but an agent must read paragraphs of guidance each time to pick correctly, and redundancy remains.
All names use camelCase and nearly all follow a verb_noun pattern (createRecord, findRelationships, deleteEmbeddingIndex, bulkCreateRecords). A handful (propertyValues, searchValues, resolveNames, smartSearch, helpAddToClient) are noun-first rather than verb-first, which is a minor deviation but still readable and predictable.
48 tools is well past the point where a set stays legible; the schema-discovery and search areas alone account for nearly twenty overlapping entries. Consolidating getSchemaMarkdown/getSchema/getSchemaSummary, deleteRecord/deleteRecordById, and semanticSearch into vectorSearch would cut the surface substantially without losing capability.
Record CRUD, bulk operations, relationship attach/detach/update, schema introspection, pattern inference, embedding and text index lifecycle, vector/keyword/NL search, and CSV export are all present, which is strong coverage for a graph database service. The main gaps are around transactions (tools note they are unavailable inside one) and lack of explicit transaction/session management tools.