AI assistants (MCP)
Every workspace is a remote MCP server. Add it to Claude, sign in through the workspace, and ask about your projects.
Every workspace is a Model Context Protocol server, at the workspace's own address: on shpyrd cloud, https://acme.shpyrd.app/mcp. Add it to an assistant such as Claude as a connector, sign in through the workspace, and ask about your projects in plain language: "check on Acme the metrics of the shop project for the last day". The assistant sees what you see and changes nothing.
Connecting Claude
- On Workspace › Overview › AI assistants, copy the Server URL (
https://<workspace>/mcp- on shpyrd cloud,https://acme.shpyrd.app/mcp). - In Claude: Settings › Connectors › Add custom connector, paste the URL, Connect.
- Claude sends you to your workspace's sign-in; sign in as usual and Allow the connection. The consent page names the assistant and what it may do.
- In a conversation, enable the connector and ask. The server's name — what Claude shows — is set by a workspace admin on the same card; by default it is "<workspace> on shpyrd".
Any MCP client that speaks Streamable HTTP with OAuth 2.1 works the same way: the workspace publishes its authorization server (/.well-known/oauth-authorization-server), accepts dynamic client registration and requires PKCE.
What the assistant can do
| Tool | Answers |
|---|---|
list_projects | the projects you can see, with status and URL |
get_project | one project's status: phase, URL, processes and their instances, the current release, access mode, domains |
get_logs | the most recent log lines (how many, which process) |
get_metrics | requests, latency, errors, CPU and memory over a range, summarised (latest, average, peak) |
Everything is read-only today and within your roles: a project you cannot open does not exist to the assistant; an owner's connection is a viewer's. Tools that change things (deploy, scale, config vars) come in a later release behind an explicit permission the consent page will name.
Connections and revocation
Your connected assistants are listed on the same card with what they may do and when they were last used; disconnect revokes the assistant's access at once (a token it already holds stops within the hour). The assistant asks you to sign in again the next time. Admins see everyone's connections (?all=true on the API).
For the platform operator
Self-hosted only: on shpyrd cloud there is nothing for you to set up. On a cluster you run, there is nothing to configure either: the OAuth server and the MCP endpoint answer at every workspace's address, with the workspace's dashboard as the token issuer and the platform's signing keys. Scripts that already hold a personal API token may call /mcp with it as a bearer instead of going through OAuth.