MCP Server

Connect AI assistants like Claude or ChatGPT to your Business Platform projects via the Model Context Protocol (MCP)

mcpaiclaudechatgptapiintegrationoauth

The Business Platform exposes a remote Model Context Protocol (MCP) server that lets you connect AI assistants — such as Claude, ChatGPT, or any MCP-compatible client — directly to your projects. Once connected, you can use natural language to query data, create entries, update records, and more.

Quick start

  1. Open your AI client's MCP server settings (e.g. Claude Desktop: Settings > MCP Servers)
  2. Add the server URL: https://app.smallstack.com/mcp
  3. Your browser opens — log in with your usual account
  4. Approve the consent screen ("Allow [client] to access your projects?")
  5. Start talking to your data: "Show me all my projects"

Enabling the MCP feature

The MCP server is gated behind a feature flag. A platform administrator must enable it for your user account before you can connect.

Once enabled, the MCP endpoint accepts authenticated requests from any MCP-compatible client.

Note: If you receive a "MCP Server feature is not enabled" error after authenticating, contact your administrator to enable the feature flag for your account.

Authentication

The MCP server uses OAuth 2.1 for authentication. MCP-compatible clients handle this automatically — you do not need to manage tokens manually.

The flow works like this:

  1. Your AI client connects to https://app.smallstack.com/mcp
  2. The server responds with OAuth discovery metadata
  3. Your client opens a browser window for you to log in
  4. You see a consent screen listing the permissions the client is requesting
  5. After approval, your client receives a token and uses it for all subsequent requests

Supported OAuth scopes:

Scope Description
openid Identify you
profile Read your profile information
email Read your email address

Available tools

Once connected, your AI assistant has access to the following tools. All tools respect your existing permissions — you can only access data you would normally see in the backoffice.

list_projects

Lists all projects you have access to.

Example prompt: "What projects do I have?"

list_types

Lists all data types defined in a specific project.

Parameter Required Description
projectId Yes The project ID

Example prompt: "What data types are in project X?"

query_data

Queries data entries of a specific type. Returns up to 50 entries by default (max 100).

Parameter Required Description
projectId Yes The project ID
collectionName Yes The collection name (from list_types)
limit No Maximum entries to return (default 50, max 100)

Example prompt: "Show me all contacts in my CRM project"

create_data

Creates a new data entry in a project.

Parameter Required Description
projectId Yes The project ID
collectionName Yes The collection name
data Yes The field values for the new entry

Example prompt: "Create a new contact named Jane Doe with email jane@example.com"

update_data

Updates an existing data entry by ID.

Parameter Required Description
projectId Yes The project ID
collectionName Yes The collection name
id Yes The entry ID
modifier Yes MongoDB-style modifier (e.g. { "$set": { "name": "New Name" } })

Supported update operators: $set, $unset, $inc, $push, $pull, $addToSet.

Example prompt: "Update the email of contact abc123 to newemail@example.com"

delete_data

Deletes a data entry by ID (soft delete).

Parameter Required Description
projectId Yes The project ID
collectionName Yes The collection name
id Yes The entry ID

Example prompt: "Delete contact abc123 from my CRM project"

Permissions and security

  • All operations enforce row-level security (RLS) — you can only read, update, or delete data you have permission to access
  • System fields (id, tenantId, createdAt, updatedAt, deletedAt, revision, access) cannot be modified
  • Data creation respects the type's defaultAccess and creationAccess configuration
  • Tokens are short-lived JWTs validated against the platform's JWKS endpoint

Supported clients

Any client that supports the MCP Streamable HTTP transport can connect. Tested clients include:

  • Claude Desktop — Settings > MCP Servers > Add Server
  • Claude Code — add to .claude/settings.json or .vscode/mcp.json
  • ChatGPT — via MCP plugin support
  • Custom integrations — any HTTP client that implements the MCP protocol

Troubleshooting

Issue Solution
401 Unauthorized Your client could not complete the OAuth flow. Check that the server URL is exactly https://app.smallstack.com/mcp and that your browser can reach the login page.
403 MCP Server feature not enabled Ask your administrator to enable the MCP Server feature flag for your account.
Permission denied on a tool call You do not have RLS access to the requested resource. Check your project permissions in the backoffice.
No projects returned Verify you have at least read access to a project.
  • AI Features — enabling AI capabilities per project
  • AI Agents — configurable AI agents within the platform
  • MCP Technical Reference — OAuth 2.1 authentication details for developers integrating directly with the MCP endpoint
Markdown | llm.txt