This page is for the people who make the Spendesk assistant available to a team: the Spendesk Admin or Account Owner, and the administrator of the AI workspace (Claude, ChatGPT, Dust, Langdock or Copilot Studio). It covers what each of them does, in order, and what users do next. For what the assistant can do once connected, see Connect an AI assistant (MCP).

Assistant	Plans	Client created in Spendesk?	Who adds Spendesk	Where users connect
Claude	Free, Pro, Max, Team, Enterprise	Yes	Team and Enterprise: an Owner or Primary Owner. Other plans: each user	Customize → Connectors
ChatGPT	Business, Enterprise, Edu (web only)	Yes	A workspace admin or owner, who publishes the app	Apps menu of a chat
GitHub Copilot	VS Code with a Copilot plan	Yes	Each user, in their own VS Code	Copilot Chat, Agent mode
Dust	All workspaces	No — Dust registers itself	A workspace admin	The agents that use the tool
Langdock	All workspaces	No — Langdock registers itself	A workspace admin	The agents that use the server
Microsoft Copilot	Copilot Studio	No — Copilot Studio registers itself	The agent's author	The agent, in Copilot Studio, Microsoft 365 Copilot or Teams

1. In Spendesk: create the connection

For Claude, ChatGPT and GitHub Copilot, a Spendesk Admin or Account Owner does this once for the organisation, for each assistant. Dust, Langdock and Copilot Studio skip this step: go to step 2.

Open Settings → Integrations → MCP (for an organisation with several entities, the MCP tab of the multi-entity hub).
Click New MCP, choose Claude, ChatGPT or GitHub Copilot, and confirm with your second factor.
Spendesk shows the connection details: the server URL https://public-api.spendesk.com/v1/mcp, a client ID, a client secret, and the authorization and token endpoints.
Send them to the administrator of the AI workspace through a secure channel — the client secret is a password for your organisation's connector, not something to paste in a chat or a ticket.

The connection's Permissions tab sets what the connector may ask for, and changing it needs a new second-factor check; each user still picks their own permissions when they connect. Write permissions are in Beta and off by default: turn on only the ones your team needs. When you enable a permission later, each user has to disconnect and connect again, and tick it — refreshing the tool list is not enough. Before allowing write permissions, read Assistant actions and safety.

2. In the assistant: add Spendesk

Claude Team and Enterprise

An Owner or Primary Owner of the Claude organisation:

Opens Organization settings → Connectors and clicks Add.
Chooses Custom → Web, and enters the server URL.
Opens Advanced settings and pastes the client ID and client secret.
Clicks Add.

Users then open Customize → Connectors, find Spendesk (labelled Custom), click Connect and approve the access in Spendesk. The first time Claude uses a Spendesk tool, it asks whether to allow it once, always, or not.

On Free, Pro and Max plans, each user adds the connector themselves from Customize → Connectors → Add custom connector, with the same URL, client ID and client secret.

Source: Claude Help Center, "Get started with custom connectors using remote MCP".

ChatGPT Business, Enterprise and Edu

Custom MCP apps in ChatGPT are in beta and web-only. Only admins and owners can publish an app to the workspace.

Turn on developer mode. On Business, each admin or owner turns it on for themselves, in Settings → Apps → Advanced settings, or when creating the app. On Enterprise and Edu, grant it under Workspace settings → Permissions & Roles → Connected Data, then turn it on in Settings → Apps → Advanced settings.
Create the app. In Workspace settings → Apps → Create, enter the Spendesk server URL, choose OAuth and fill in the client ID and client secret. Click Scan Tools, approve the access in Spendesk, then Create. The app is saved as a draft.
Test it from a chat, with the draft app selected.
Publish it from Workspace settings → Apps → Drafts → Publish. On Enterprise and Edu, the publish dialog lets you choose which Spendesk actions the app may use (Configure Actions) and which groups get it (Configure Access).

Users then select Spendesk in a chat and approve the access in Spendesk the first time. ChatGPT may ask for confirmation before a write action.

❗️
New Spendesk tools do not appear on their own
ChatGPT keeps a snapshot of the app's tools from when it was published. When Spendesk adds or changes tools — as when write actions arrived on 25 September 2026 — an admin has to update the app: on Enterprise and Edu, Workspace settings → Apps → ⋯ → Action control, then Refresh (new actions start disabled); on Business, recreate and republish the app.

Source: OpenAI Help Center, "Developer mode and MCP apps in ChatGPT".

GitHub Copilot (VS Code)

Each user adds Spendesk in their own VS Code, with the details from step 1:

Pastes the configuration shown on the last step of New MCP (server URL and client ID) in .vscode/mcp.json.
Clicks Add Secret and pastes the client secret.
Clicks Start Server, approves the access in Spendesk, and asks questions in Copilot Chat, in Agent mode.

If several people use the same connection, share the client secret through a password manager.

Dust

Dust registers itself: there is no client to create in Spendesk. A Dust workspace admin:

Opens Spaces → Tools → Add tool, searches for Spendesk and selects it.
Chooses Personal accounts, so that each member uses their own Spendesk account, and clicks Setup connection.
Approves the access in Spendesk (Authorise), and is sent back to Dust.
Adds the Spendesk tool to the agents that should use it.

Users approve the access in Spendesk the first time an agent uses it.

A connection registered this way gets default permissions. An Admin or Account Owner enables the others from the connection's Permissions tab in Spendesk. Dust can still use a client created in Spendesk instead (Add MCP server, static OAuth, with the client ID, client secret, authorization and token endpoints).

Source: Dust documentation, "Adding an MCP Server", and the Spendesk app's set-up steps.

Langdock

Langdock registers itself: add a remote MCP server with the Spendesk server URL, and Langdock opens the approval page. As for Dust, such a connection gets default permissions, and an Admin or Account Owner enables the others from its Permissions tab.

Microsoft Copilot (Copilot Studio)

Copilot Studio registers itself: there is no client to create in Spendesk. The agent's author, with a Copilot Studio licence (or the trial):

Opens the agent in Copilot Studio, then Tools → Add a tool → New tool → Model Context Protocol.
Enters a name, a description and the Spendesk server URL. Under Authentication, chooses OAuth 2.0 and keeps Dynamic discovery, then clicks Create.
Creates a connection, approves the access in Spendesk, then clicks Add and configure.
Tests the agent in its Test pane, then publishes it to Microsoft 365 Copilot or Teams.

If the test pane asks to open the connection manager, another tool of the agent has a missing or outdated connection: fix it or turn that tool off, then start a new test.

Source: Microsoft Learn, "Connect your agent to an existing Model Context Protocol (MCP) server", and a Spendesk test on 29 October 2026.

3. Check the rollout

Ask a user to connect and try the questions in Check the connection. If a company is missing, the user is not a Controller or Account Owner there (a Requester only sees their own card spend); if a tool is missing, its permission is not enabled on the connection, or the user did not tick it — they disconnect, connect again and tick it — or, in ChatGPT, the app needs a refresh.

Assistant steps checked against each vendor's documentation on 29 September 2026. Their menus change often: if a step no longer matches, follow the linked source and tell us at [email protected].