Skip to content

Connecting AI clients

Canvus MCP 26.11.0 was verified with Codex CLI 0.164.0 using normal OAuth without paid model calls. Hosted Claude acceptance is not claimed. Replace the reserved example origin with the HTTPS origin supplied by your administrator. See release qualification; provider plan and organization permissions may control connector availability.

Before connecting

Your administrator must complete installation, confirm Canvus compatibility and supply an endpoint of the form https://mcp.customer.example/mcp. Do not connect to an MT demonstration instance. You need your own Canvus account and any required permission from your AI-provider organisation to add a custom connector.

The service uses Streamable HTTP with OAuth, not a shipped stdio bridge. It supports its own discovery, S256 PKCE and compatible client registration/metadata flows; do not replace those with a shared static token or a generic proxy login screen.

Your chosen AI provider can receive authorised content, images and manual results. Your Canvus account permissions and protected-change approvals still apply. An AI subscription is not a Canvus credential, and an MCP token is not your Canvus password.

Claude Desktop and claude.ai remote connectors

These remote connectors run from Anthropic's cloud, including when configured in Claude Desktop. Your administrator must expose the customer MCP origin over publicly reachable HTTPS on 443, with a publicly trusted certificate. A private LAN hostname or a VPN on your laptop will not work from that cloud. No MT relay is required.

  1. Follow Claude's custom connector instructions. In the connector settings, add a custom connector with your administrator's /mcp endpoint. If the option is absent, check your plan and organisation permissions rather than changing the MCP server.
  2. Start the connector's Connect/Authenticate action. Check that the browser reaches your customer MCP origin. The MCP consent flow asks you to sign into your own Canvus account; never send your password to support or another facilitator.
  3. Review the requested access and complete consent. The hosted callback is https://claude.ai/api/mcp/auth_callback; do not replace it with a laptop callback or add wildcard callback rules. Client registration and browser CORS allowlists are different settings.
  4. Return to Claude and confirm the connector is connected and its tool catalogue is available. OAuth/catalogue acceptance can be checked without a paid conversation. Record cloud acceptance separately from local CLI success.

If the callback is rejected, record only the error code and client version, not the callback URL with its query parameters. Reconnect using the normal client registration flow after your administrator checks the exact registered callback.

Directly configured local Claude Code

Use Claude Code's HTTP MCP setup. The execution host must reach the endpoint; LAN/VPN routing can work when Code runs there. Remote execution or inherited cloud connectors may use a different network path.

Run this on the host where Code executes, after replacing the hostname. This is a client configuration recipe, not a claim of hosted Claude qualification:

claude mcp add --transport http canvus https://mcp.customer.example/mcp

Inside Claude Code, open MCP management and select the Canvus server's authentication action:

/mcp

Complete the browser flow with your own Canvus account, then return to Code. Do not copy access tokens into the command line or enable a wildcard redirect to make authentication pass.

Local Codex

Use the Codex MCP instructions. This recipe is for local execution, not a cloud environment. Reachability is from the executing host, not necessarily from the browser used for sign-in.

For Codex CLI 0.164.0, replace the hostname before adding the server:

codex mcp add canvus --url https://mcp.customer.example/mcp

Then start OAuth:

codex mcp login canvus

Complete the browser sign-in/consent and return to Codex. If you use remote execution, have your administrator verify that environment's routing separately.

Check the connection without making changes

Use your client's tool inspector where available, or an explicitly approved AI session; a provider conversation may incur charges. For acceptance use synthetic content you own, not another user's canvas.

  • Confirm discovery/catalogue succeeds and your own authorised canvas listing/read succeeds. Do not mistake an empty authorised listing for a server-wide lack of canvases.
  • Call canvus_get_knowledge with topic set to start to read the operating guide.
  • Call canvus_search_manual with query set to canvas, then read one returned URI with canvus_read_manual. A valid result includes a corpus revision and citation; manual_unavailable is an installation fault, not zero search matches.
  • Do not test by creating/deleting content or approving a destructive action. Listing tools alone does not prove protected reads, manual availability or all backend capabilities.

Full facilitator functionality remains available under the actor's permissions; this is not a reduced-capability mode. Changes requiring preview/approval must follow their normal flow. Never retry an uncertain write as a new operation merely because the client timed out.

Reconnect and retire access

If the service origin/Canvus authority changes or your session becomes invalid, use your client's normal reconnect/authenticate action. Do not copy tokens between installations. Canvus passwords are not retained by MCP; backend credentials are encrypted in the local installation's state.

When removing access, use the client's documented disconnect/revoke mechanism and verify that local MCP authority is retired. Removing a configuration entry alone may not revoke its server-side grant. MCP's OAuth /revoke retires local MCP authority and stored secret; it does not call Canvus users/logout. Upstream credential retirement must follow your Canvus administrator's lifecycle procedure. There is no universal MCP /logout URL or customer revocation dashboard asserted here.