WYRE MCP Gateway Troubleshooting

WYRE MCP Gateway Troubleshooting

This skill walks through the most common failure modes in the WYRE MCP Gateway and how to resolve them. Use it whenever a user is in the gateway dashboard or in Claude Desktop / Claude Code and something isn't behaving as expected.

The gateway request path has five layers — identify which layer is failing first:

Claude Desktop / Code  →  mcp-remote  →  WYRE Gateway  →  Vendor Container  →  Vendor API
        (1)                  (2)             (3)              (4)                (5)

Most user-reported issues are at layer 1 (Claude's tool registration cache) or layer 3 (gateway permissions / credentials). Layers 4 and 5 generally surface as a tool call returning an error after a tool has already been registered.

Anti-triggers

This is a shared skill and it owns layers 1–3: the tool never appeared, or it fails before it reaches the vendor. Once a tool has registered and is returning vendor responses, the problem is at layer 4 or 5 and belongs to that vendor's own api-patterns skill, which knows its auth model, rate limits, and error codes.

• A tool that runs but returns an error from the vendor — a 4xx with a vendor-shaped error body is not a gateway fault. Use ninjaone-api-patterns, hudu-api-patterns, qbo-api-patterns, pax8-api-patterns, sherweb-api-patterns, clio-api-patterns, or the equivalent for that vendor.
• A vendor-specific permission model — Hudu's per-API-key password toggle, Clio's decision-tree navigation, Sherweb's progressive tool disclosure, and QuickBooks' realm scoping all look like gateway problems and are not. Each is documented in that vendor's skill.
• Rate limiting or partial results mid-task — that is the vendor's throttle, not the gateway's.

Symptom: Newly added vendor's tools don't appear in Claude

Most common cause: Claude has not refreshed its tool list.

Claude registers the gateway's tool catalog at the moment its MCP client connects, and does not refresh while a session is live. Adding a vendor in the Wyre dashboard will not push new tools into a running Claude session.

Fix:

1. Verify the vendor connection is saved at conduit.wyre.ai — open Plugins and confirm the vendor shows as connected with a recent test result.
2. Fully quit Claude Desktop (Cmd+Q on macOS) and reopen it. Closing the window is not enough — the MCP client process keeps running.
3. In Claude Code, run /mcp, find the gateway entry, and reconnect it. (Restarting the whole CLI also works.)
4. Start a new conversation. The first tool list request will trigger the gateway to re-aggregate vendor tools and register them with Claude.

If tools still don't appear after a full restart:

• The vendor's credentials may have failed validation. Open Plugins, click the vendor, and run the connection test. Re-enter credentials if the test fails.
• For OAuth-based vendors (Xero, QuickBooks, HubSpot, M365), the OAuth token may have expired. Click Reconnect to re-authorize.
• The unified endpoint silently skips vendors with invalid credentials, so a single bad vendor will not block the others — but you also will not see an error toast for the failure.

Symptom: "Failed to update tool access" when saving the Tool Allowlists page

Cause: Saving or clearing a tool allowlist requires the org owner role. Admins can view the Tool Allowlists page and will see every vendor's tools displayed as "checked" (the allow-all default), but any Save or Reset action returns 403 Requires owner role or higher. The UI surfaces this as the generic "Failed to update tool access" alert.

Fix:

1. Open Organization → Members and identify the org owner. The owner is typically the user who originally created the organization.
2. If you are not the owner, ask the owner to make the allowlist change for you, or have them transfer ownership.
3. If you are signed in as the owner and the error persists, escalate to hello@wyre.ai with your org name and the vendor you were editing.

Important context to share with users:

• A fresh org will show every vendor tool as checked. That is expected and means "allow all" — there is no saved allowlist row, so the gateway permits every tool for that role. Don't read it as "an allowlist already exists."
• Members and admins are subject to allowlists if configured; owners are never restricted by allowlists.

Symptom: OAuth fails — "Invalid OAuth error response: Unexpected token '<'"

Cause: The OAuth discovery or token endpoint returned HTML (typically a Cloudflare error page) instead of JSON. The gateway is briefly unreachable.

Fix:

1. Check the gateway health: curl https://conduit.wyre.ai/health should return {"status":"ok"}.
2. Check vendor health at https://conduit.wyre.ai/health/vendors — admin-gated, so it needs a gateway admin session or the ops admin key. Without one, open Plugins in the dashboard and run Test Connection for the vendor instead.
3. Wait 60–120 seconds and retry — most deployments complete within a minute.
4. If /health is failing for more than 2 minutes, contact hello@wyre.ai.

Symptom: A tool call returns a credential or authentication error

This means the tool was registered (so layers 1–3 are fine) but layer 4 or 5 rejected the call.

Walk through:

1. Confirm credentials in the dashboard. Open Plugins, click the vendor, run Test Connection. If the test fails, re-enter credentials.
2. Vendor OAuth tokens have limited lifespans. If the dashboard shows a "Reconnect" prompt for the vendor, the OAuth token expired — click Reconnect to re-authorize.
3. Some vendors (e.g., Autotask, Datto) silently drop requests from non-allowlisted IPs. If the test connection works from the dashboard but tool calls time out, ask WYRE support to confirm the gateway's egress IP is on the vendor's allowlist.

Symptom: Connection refused on the OAuth callback (localhost:NNNNN)

Cause: mcp-remote opens a local listener for the OAuth callback. With multiple per-vendor entries in claude_desktop_config.json, the listeners can race and time out.

Fix:

1. Use the unified endpoint — replace per-vendor mcp-remote entries with a single entry pointing to https://conduit.wyre.ai/v1/mcp. One OAuth flow instead of many.
2. Kill stale mcp-remote processes: pkill -f mcp-remote.
3. Restart Claude Desktop.

Diagnostic Quick Reference

What to escalate to WYRE support

Before escalating, capture:

1. Org name and approximate time of the failure (UTC if possible).
2. Vendor slug (e.g., itglue, autotask, cipp).
3. Exact error message as it appeared (screenshot is best).
4. What you have already tried — restarting Claude, reconnecting the vendor, etc.

Send to hello@wyre.ai. The gateway logs every request with a reqId, so a precise timestamp lets us correlate.