Praxsuite

AI Connectors

Vincent Depassier · August 30, 2026

An AI connector lets a hosted AI platform — Claude.ai, ChatGPT — reach your workspace directly, without you running anything.

It is one of two ways to expose Praxsuite over MCP, and picking the wrong one is the most common early mistake here.

You will find it at API Gateway → AI / MCP.

Connector or API key?

Both end at the same MCP surface. They differ in who holds the credential.

AI connector

API key (`sk_live_`)

Auth

OAuth2, registered with our identity server

A bearer secret you paste

Who holds it

The AI platform

You, in a config file

Set up in

The AI platform's connector UI

claude_desktop_config.json, Cursor settings, your own code

Right for

Claude.ai and ChatGPT on the web

Claude Desktop, Claude Code, Cursor, your own MCP client

Revoking

Revoke the connector

Revoke the key

The rule: if the AI runs on somebody else's servers, use a connector. If it runs on your machine or in your code, use an API key.

A hosted platform cannot be handed a long-lived bearer secret responsibly, which is exactly the problem OAuth2 exists to solve. A local client has no OAuth redirect to complete, which is exactly why a key is simpler there.

Setting up the key path is covered in MCP — Getting Started.

Creating one

Registering a connector creates an OAuth2 client scoped to your workspace and returns a client id and client secret.

The client secret is shown exactly once and is never stored on our side. Copy it into the AI platform when you create it; if you lose it, delete the connector and register a new one.

What a connector can reach

A connector carries the same three permission axes as a credential, and they are checked the same way:

Axis

Governs

Table scopes

Which tables, at what access level, and whether it may introspect their structure

Docs scopes

Which folders, spaces and documents it may read or edit

Tool permissions

Which MCP tool groups it may call at all

None of them widens another. A connector with every tool group enabled and no table scopes can call every tool and find nothing.

One default differs from a credential's: a connector's table scope is created with schema introspection on, because an assistant that cannot see column names cannot write a sensible query. That is the right default and it is worth knowing you have made it — a connector sees the shape of every table you grant it.

Scopes can be replaced at any time and take effect on the connector's next call.

Guidance

  • One connector per audience, not per person. "Claude – Marketing" with the marketing tables beats one connector everybody shares.

  • Start read-only. Give read on the tables the assistant should reason about, then widen when you see from the Logs what it actually calls. MCP entries there show every tool call.

  • Keep `Security` tools off. An assistant that can mint API keys can outlive its own revocation.

  • Deactivating the connector stops it immediately. That is the lever if an assistant starts doing something you did not expect — no need to go and change scopes first.

Next