Aurabase Logo
aurabasedocs
docs›Reference›MCP Server

MCP Server

Control your Aurabase backend from Claude Desktop, Cursor, Windsurf, Antigravity, Mistral AI, VS Code or any Model Context Protocol compatible client — 86 tools, no manufactured value.

7 min read·Level reference·Revised on Oct 18, 2026
This English text was generated automatically from the French original and has not been reviewed yet.
#
Installation

HTTP mode by default & Local STDIO mode

The official Aurabase MCP server exposes 86 tools to inspect and control your backend. It provides two integration modes:

  1. HTTP Mode (Streamable HTTP / SSE) — Recommended & Default: direct network connection via URL (https://mcp.aurabase.cloud/mcp or http://localhost:3001/mcp) without running a local Node.js process per editor.
  2. STDIO Mode (Local): execution of a subprocess npx @aurabase/mcp-server managed by the editor for local development stacks.

To start a local HTTP server (optional if using the hosted cloud service):

terminalbash
npx @aurabase/mcp-server \
  --transport http \
  --port 3001 \
  --auth-token my_secret_token \
  --url http://localhost:8080 \
  --project-id <uuid> \
  --service-role-key aura_sk_…
#
Clients

Configuration per client and per mode

HTTP Mode (Default & Recommended)

Add the following configuration into your MCP client. The server uses Streamable HTTP with a Bearer token:

Cursor — .cursor/mcp.json or ~/.cursor/mcp.json:

.cursor/mcp.jsonjson
{
  "mcpServers": {
    "aurabase": {
      "url": "https://mcp.aurabase.cloud/mcp",
      "headers": {
        "Authorization": "Bearer aura_sk_…"
      }
    }
  }
}

Claude Desktop — claude_desktop_config.json:

claude_desktop_config.jsonjson
{
  "mcpServers": {
    "aurabase": {
      "url": "https://mcp.aurabase.cloud/mcp",
      "headers": {
        "Authorization": "Bearer aura_sk_…"
      }
    }
  }
}

Windsurf — ~/.codeium/windsurf/mcp_config.json:

mcp_config.jsonjson
{
  "mcpServers": {
    "aurabase": {
      "serverUrl": "https://mcp.aurabase.cloud/mcp",
      "headers": {
        "Authorization": "Bearer aura_sk_…"
      }
    }
  }
}

Google Antigravity — ~/.gemini/config/mcp_config.json:

mcp_config.jsonjson
{
  "mcpServers": {
    "aurabase": {
      "url": "https://mcp.aurabase.cloud/mcp",
      "headers": {
        "Authorization": "Bearer aura_sk_…"
      }
    }
  }
}

VS Code (Cline / Roo Code) — cline_mcp_settings.json:

cline_mcp_settings.jsonjson
{
  "mcpServers": {
    "aurabase": {
      "url": "https://mcp.aurabase.cloud/mcp",
      "headers": {
        "Authorization": "Bearer aura_sk_…"
      }
    }
  }
}

Mistral AI (Agents & Connectors) — mistral_connector.json:

mistral_connector.jsonjson
{
  "name": "aurabase",
  "url": "https://mcp.aurabase.cloud/mcp",
  "headers": {
    "Authorization": "Bearer aura_sk_…"
  }
}

Continue.dev (VS Code & JetBrains with Codestral) — ~/.continue/config.yaml:

config.yamlyaml
models:
  - name: Codestral
    provider: mistral
    model: codestral-latest
    apiKey: MISTRAL_API_KEY

mcpServers:
  - name: aurabase
    url: https://mcp.aurabase.cloud/mcp
    headers:
      Authorization: Bearer aura_sk_…

STDIO Mode (Local via subprocess)

For offline development or against a local Docker/k3d stack (http://localhost:8080), your client executes the npx subprocess:

mcp_config.jsonjson
{
  "mcpServers": {
    "aurabase": {
      "command": "npx",
      "args": ["-y", "@aurabase/mcp-server"],
      "env": {
        "AURABASE_URL": "http://localhost:8080",
        "AURABASE_PROJECT_ID": "<uuid>",
        "AURABASE_SERVICE_ROLE_KEY": "aura_sk_…"
      }
    }
  }
}
#
Principle

A tool returns the response from the backend, or an error

Never a manufactured value. This rule is not an intention: it is checked at each execution of the CI by a bench which launches the server against a dead URL and fails if the slightest tool responds successfully. Each tool also declares an output schema that the SDK validates before emitting the response.

Astuce

Read-only mode (--read-only) removes mutating tools from the catalog and instructs PostgreSQL to execute the SQL in a read-only transaction (37 tools of 86) — the guarantee comes from the engine, not just from client-side filtering.

#
Options

Flags and environment variables

--url``AURABASE_URLAurabase instance URL

--project-id``AURABASE_PROJECT_IDProject ID

--service-role-key``AURABASE_SERVICE_ROLE_KEYAdmin Key

--read-only``AURABASE_READ_ONLYOnly exposes reading tools (37 of 86)

--allowed-tools``AURABASE_ALLOWED_TOOLSWhitelist, separated by commas

--destructive-policy``MCP_DESTRUCTIVE_POLICYInteractive confirmation of destructive SQL operations: elicit (default), token, deny

--state-key``MCP_STATE_KEYKey (≥ 32 bytes) that signs multi-turn confirmations

--transport—http (default) or stdio

--port—HTTP transport port (default 3001)

--host—HTTP listening interface (default 127.0.0.1)

--auth-token``AURABASE_MCP_AUTH_TOKENToken required on /mcp — required in http mode

#
86 tools

Distributed by domain

82 exposed with a simple project key, 4 reserved for the management plan (API keys and list_all_projects — better no tool than a tool that systematically fails without this channel).

Database (23)list_tables, describe_table, get_performance_advisors, list_extensions, execute_sql, query_table, insert_rows, update_rows, delete_rows, call_rpc, generate_types, diff_schema, apply_migration, list_migrations, explain_query, analyze_table, get_db_size, list_sql_functions, create_sql_function, drop_sql_function, list_sql_triggers, create_sql_trigger, drop_sql_trigger
Storage (9)list_buckets, get_bucket, create_bucket, list_files, delete_files, move_file, create_signed_url, upload_file, empty_bucket
Auth (7)list_users, get_user, create_user, update_user, delete_user, ban_user, generate_link
Notifications (10)send_email, send_push_notification, send_sms, list_notification_logs, get_notification_settings, update_notification_settings, list_notification_templates, create_notification_template, delete_notification_template, get_notification_stats
Backups (5)list_backups, create_backup, restore_backup, get_backup_schedule, set_backup_schedule
Edge Functions (6)invoke_function, list_functions, get_function_info, deploy_function, delete_function, set_function_env
RLS (6)list_rls_policies, create_rls_policy, update_rls_policy, drop_rls_policy, toggle_rls, get_security_advisors
AI (5)index_document, rag_query, vector_search, create_vector_index, ai_completion
Observability (4)get_project_metrics, list_slow_queries, list_active_queries, cancel_query
Realtime (3)list_realtime_channels, get_realtime_settings, update_realtime_settings
Project (4)project_info, switch_project, get_active_project, list_all_projects (management channel)
API keys (3)create_api_key, list_api_keys, revoke_api_key (management channel)
Documentation (1)search_docs — search in embedded documentation (English)

The *_sql_function / *_sql_trigger tools create real PostgreSQL functions and triggers stored in the project schema — unrelated to list_functions / deploy_function (Edge Functions WASM, a separate runtime).

#
Transport

HTTP by default for remote, stdio for local use

http is the modern default mode — adhering to MCP protocols 2025-11-25 and 2026-07-28. The http mode:

  • listens on 127.0.0.1 by default in local, or on your deployment URL;
  • requires a token (--auth-token or Bearer token) — without it, the server refuses to start;
  • validates the Origin header and responds 403 to an unauthorized origin (anti-CSRF protection);
  • one server instance per session, with expiration and configurable cap;
  • compares tokens in constant time and bounds request body size.
Last updated · Oct 18, 2026