MCP Hub

The MCP Hub keeps OAuth and other remote MCP sign-ins in vord Cloud. Connect a server once and sessions started by vord can use it through the Hub. This avoids agents replacing one another’s rotating sign-in tokens.

The feature requires the vord MCP Hub add-on and an active vord base subscription. It does not require Brain or Chat. See vord Cloud pricing and booking details.

For servers that should stay in an agent’s own CLI configuration, use the MCP manager. Local MCP servers should remain direct; see Limits.

Add a server

In vord, open the MCP manager and use its MCP Hub area. Add a server with its name, slug, upstream URL, and authentication method: OAuth, an API-key header, or no authentication. A slug identifies the server to agent sessions. If vord Cloud does not accept the server, vord shows a message with the reason and the error code from vord Cloud; if vord Cloud does not answer within 30 seconds, vord says so and you can try again.

Some OAuth providers do not support dynamic client registration or client ID metadata. For those providers, enter a client ID under the advanced options. A client secret may also be needed.

Connect an account

Choose the server’s connect action. vord shows a six-digit confirmation code while it waits, and your browser opens a confirmation page showing the server, URL, and a partly hidden account email. Enter the code from the app on that page; Copy beside the code puts it on the clipboard without spaces. Then continue to the provider and sign in there. The code ties the browser page to the connection you started in the app; after five wrong codes the link stops working and you connect again from the app. After authorization, the browser returns to vord and the server’s connection status updates. Starting a new connection needs a current version of the app; existing connections are not affected.

If you close the browser before finishing, choose Open browser again beside the code to return to the same confirmation page. Choose Cancel to stop waiting for the connection. The code and confirmation link expire after five minutes; if they expire, choose Connect again to start a new attempt.

Choose one of three symbols in the server’s share control. The person symbol means the server is not shared. The two-person symbol shares the server setup, so each organisation member signs in with their own account. The key symbol shares the owner’s existing sign-in, so members act through the owner’s account. Hover over a symbol to see its full description. The owner can see members’ calls made through a shared login.

The credential option is for cases where members should use a shared account. Make sure its account permissions are appropriate for everyone who will use it.

An organisation member can have a server of their own with the same slug as a shared server. The member’s own server takes precedence and hides the shared one.

The server list shows whether a server is connected, needs sign-in, has an error, or is not connected. When a server needs sign-in again, vord notifies you once for that server. The app checks every 15 minutes, when a session starts, and when you open the MCP Hub area. After you reconnect the server, a later sign-in problem notifies you again.

Move existing remote servers

The apps can offer eligible existing global remote MCP entries from agent CLI configurations for transfer to the MCP Hub. The direct entry is removed only after the Hub connection succeeds and you confirm the transfer. Before removal, the app saves a backup under ~/Library/Application Support/<App>/mcp-backups/.

Entries marked as the older HTTP+SSE transport (Claude, Codex, Kimi and Grok) are not offered for transfer, and they are not switched off in favour of a Hub server with the same URL.

When vord can disable a duplicate — for example, a Claude entry, a claude.ai connector, or a ChatGPT app in Codex sessions — it uses the Hub server in agent sessions and disables the other entry. In Codex sessions, Hub servers always appear as <slug>-hub, for example machbar-hub. vord does not disable direct Codex entries, as with Kimi and Muse, so a direct entry with the same name remains alongside the Hub entry until you remove it. MCP project selection cannot hide direct Codex servers. Names ending in -hub are reserved for vord’s Hub entries. Changes to server availability apply to newly started sessions; running sessions keep their existing configuration. Codex rereads config.toml on every turn, so changes to direct entries take effect on the next turn; Hub entries remain as they were when the session started. In a session, Hub servers appear under their slug.

If the Hub cannot be reached when a session starts, the session starts without Hub servers and the app shows one notification per outage. New sessions get the Hub servers again once the Hub is reachable.

Sessions that start while the app is still connecting to vord Cloud, for example right after launch, wait up to 10 seconds for the connection. If it is not ready by then, the session starts without Hub servers. Once the connection is established and the Hub has servers for you, the app restarts such a session at the next quiet moment (no reply running, no open question) and resumes the same conversation with the Hub servers. You get one notification per restarted session. If a session cannot be resumed yet, the restart waits until it can.

Hub account controls

In the web Hub, open Connections & MCP to disconnect your own server logins or withdraw shares. Server owners can manage their shares. Organisation owners and admins can also withdraw shares across the organisation.

See Connections & MCP for the Hub’s other connection controls.

Use the MCP hub in other AI apps

Use the aggregate MCP URL https://cloud.vord-app.dev/gw/mcp to connect an AI app to the Hub. Each server also has a direct URL: https://cloud.vord-app.dev/gw/{slug}/mcp, replacing {slug} with that server’s slug.

The aggregate URL always exposes only five fixed tools: search_tools, describe_tools, call_read_tool, call_write_tool, and call_destructive_tool. They let an AI app discover tools, inspect their input schemas, and call each tool through the matching read, write, or destructive route. This keeps the model’s context small.

Connect ChatGPT or Grok

Add https://cloud.vord-app.dev/gw/mcp as a custom MCP connector in the app. The MCP Hub supports Dynamic Client Registration, so apps such as Grok and other clients can register themselves without a client metadata document.

Sign in to vord Cloud with your vord account. Each new connection requires a new email code.
If you belong to several workspaces, vord asks you to Choose tenant. Choose the workspace where your MCP Hub servers and their sign-ins are set up, usually your organisation rather than your personal workspace. Servers and sign-ins belong to one workspace. If you choose another one, the app later reports the servers as not connected or as having no valid login.
On the consent page, check Signed in as <email> · <organization>. If the account or workspace is wrong, choose Use another account and sign in again.
Select the server the app should use.
Choose Allow.

Consent is available only when the vord Cloud sign-in is no more than five minutes old. If it is older, vord Cloud signs you out and asks you to sign in again. Self-registered apps appear as Unverified app, together with the host of the redirect address. The signed-in page has a Sign out button.

If the app already has a connection using the wrong account or workspace, remove that connection in the app first, then connect again.

How AI apps discover tools

The Hub discloses connected servers and their tools in stages:

The Hub’s instructions list up to 20 connected server names.
search_tools without arguments lists each server, a one-sentence purpose, and its number of read, write, and destructive tools. A server owner can set a one-sentence Description of up to 140 characters in the Hub under Account → Connections; it appears as the purpose. If no description is set, the Hub uses the first sentence of the server’s instructions when available and marks it as server-provided.
search_tools with server lists that server’s tools in pages, using offset and returning at most 10 tools per page. With query, it searches across all servers.
describe_tools with up to 10 names returns the tools’ input schemas.
Call a tool with call_read_tool, call_write_tool, or call_destructive_tool, matching its type.

Servers with their own compact catalogs, such as Machbar and vord Brain, are unpacked automatically. Their actual tools appear in stage 2. Tool descriptions come from the servers and are not verified by vord. Hub responses have fixed size limits.

This staged disclosure applies to external AI apps. Agents in vord apps continue to use each server’s individual URL and are not affected.

vord Brain and vord Chat in AI apps

When your organisation has booked the vord Brain or vord Chat add-on, the aggregate URL also offers the built-in servers brain and chat. Without the add-on they do not appear, and if the add-on ends they disappear from the next call on. Through brain, an AI app can search, read, and write notes, learnings, skills, and tasks with your rights. With the read-only Brain add-on, only reading tools appear. Through chat, an AI app reads the team chat and posts as you, without a separate agent identity. Encrypted Brain vaults cannot be read through the Hub yet; connect vord Brain directly at https://cloud.vord-app.dev/mcp for them. Personal Hub tokens cannot use brain or chat; connect the AI app with OAuth. The slugs brain and chat are reserved and cannot be used for your own servers. Agents in vord already have vord Brain and vord Chat through vord itself and do not receive the built-in servers a second time.

AI apps may see these errors:

-31006 — Outcome unknown: A write or delete call may have reached the server before the connection broke or the deadline expired, so the result is unclear. Check the server’s state before trying again. The Hub never retries these calls itself.
-31005: The server requires interactive input that the client’s protocol version does not support.

OAuth errors from vord Cloud use the standard OAuth error format.

A tool call can run for up to 10 minutes. If the server reports progress, the Hub also ends the call when no further progress arrives within 2 minutes of the last update; a tool that never reports progress is not cut off before the 10-minute limit. When a limit is reached, write and delete calls return -31006, read calls a normal timeout error.

The Hub renews an expired upstream login once. If the server requires scopes the connection does not have, the Hub marks the server for sign-in again.

Claude and Claude Desktop

Add a custom connector using the aggregate URL, then sign in with your vord account. On vord’s consent page, choose which connected servers Claude may use. Custom connectors are available depending on your Claude plan; check Anthropic’s current availability information. In Team and Enterprise organisations, an Owner or Primary Owner must add the organisation connector; each user then connects individually.

ChatGPT

OpenAI currently documents full MCP support, including write and modify actions, for ChatGPT Business, Enterprise, and Edu on the web. Pro can connect custom MCP servers with read and fetch permissions only. See OpenAI’s current setup and availability information. A workspace admin or owner adds and publishes the connector; on Enterprise and Edu, an admin can grant selected members permission to create and test connections before publication. ChatGPT may ask you to confirm write or modify actions based on the app’s permissions and the action’s context. Follow Connect ChatGPT or Grok for the vord Cloud sign-in and consent steps.

Claude Code and Codex CLI

Both clients support remote MCP over HTTP with OAuth. Add the aggregate URL and complete the sign-in flow with your vord account. The commands below follow each client’s current official setup documentation: Claude Code MCP and Codex MCP.

For Claude Code, add the remote HTTP server and complete OAuth when prompted:

claude mcp add --transport http vord-hub https://cloud.vord-app.dev/gw/mcp

For Codex CLI, add the server and start its OAuth login:

codex mcp add vord-hub --url https://cloud.vord-app.dev/gw/mcp
codex mcp login vord-hub

VS Code, Cursor, and scripts

VS Code and Cursor can use OAuth when they support Dynamic Client Registration. Clients that do not support that flow can use a personal token instead.

For clients that use a personal token, create one in the Hub’s Connections & MCP page and send it in the HTTP Authorization header, for example Authorization: Bearer <personal-token>. Never put a token in the URL or its query string. See the official VS Code MCP configuration reference or Cursor MCP setup for their header fields.

Create personal tokens in the Hub at Connections & MCP (/account/connections). Confirm the request with the email code; the token is shown only once. Choose an expiry of 30, 90, or 365 days, or no expiry; a token without an expiry stays valid until you revoke it. Each member can have up to 20 active personal tokens. You can revoke a token at any time. Use Manage connected apps to review connected apps in the Hub and revoke their access.

Each person sees only servers they can see in vord and connects with their own vord account. If a server owner shares a login with the organisation, calls from external apps use that shared login too. The server operator declares whether each tool reads, writes, or deletes data.

All MCP calls pass through vord Cloud, and the external AI app receives the tool results. Without the vord MCP Hub add-on, tool search returns no results and tool calls return a notice that the add-on is not active. Your connection stays in place, so booking the add-on later makes the tools available again without reconnecting.

Privacy

The MCP Hub does not store the contents of tool calls. It keeps a 90-day call log containing the tool name and a hash of the call parameters. You can see your own calls. The owner of a credential share can also see calls made by members through that owner’s login. Organisation admins do not have access to individual call logs.

Upstream tokens are sealed in vord Cloud. The first connection flow explains that calls pass through vord Cloud.

Limits

Local MCP servers that run as stdio processes stay configured directly and are not routed through the cloud Hub.
The older HTTP+SSE transport is not supported. The Hub supports Streamable HTTP.
A tool call runs for at most 10 minutes and stops after 2 minutes without progress.
Each account can keep up to 64 streaming responses and 64 server-notification streams open at the same time, at most 16 of each per agent session. An agent session opens its notification stream only while it is working and closes it about a minute after its last request. When it starts working again, the agent reloads its tool lists, so changes made in the meantime are not missed.
OAuth providers without dynamic client registration or client ID metadata need a manually entered client ID in the advanced options.
Connectors managed by claude.ai are not affected.