Troubleshooting MCP

The MCP server is mature, but a few situations crop up repeatedly. This page collects the fixes in one place.

"My assistant can't see Jasning"

Most common cause: the URL is wrong.

The endpoint is exactly https://jasning.com/mcp. With the trailing /mcp. Without a trailing slash. Without /api. Without /v1.

Second cause: the OAuth flow didn't complete. On the first connection you should see a Jasning consent screen open in a new tab. If you closed it before clicking Grant access, the connector exists on the assistant side but has no token. Remove and re-add it.

Third cause: your assistant doesn't support the streamable HTTP transport. Claude (web, iOS, desktop), ChatGPT (Plus+), Cursor (0.40+), and Claude Code all do. Older or self-hosted clients may not.

"Calls fail with 401"

The OAuth token expired, was revoked, or was never granted. Reconnect from your assistant's settings. New tokens take effect immediately.

If you keep getting 401s after reconnecting, check Jasning Settings → Tokens. If the token shows there but still fails, email hello@jasning.com with the token's short ID; we can dig in server-side.

"Calls fail with 429"

Rate limit. The MCP server allows several calls per second per token in normal usage. Bulk operations that fan out to 50+ rapid calls will hit the limiter.

Fix: ask the assistant to batch. Instead of "create each of these 50 customers", which fires 50 parallel calls, ask "create these 50 customers one at a time". Claude and ChatGPT both respect that hint.

"The AI made up an appointment that doesn't exist"

Classic hallucination. The assistant generated plausible details without looking them up.

Fix: tell it to use the tools first.

Use appointments-list before saying anything specific.

Once you say this in a session, the model usually maintains the habit for the rest of the chat. Some assistants forget after long pauses; repeat the instruction in long sessions.

"First call after a quiet stretch is slow"

The streamable HTTP connection goes cold after inactivity. The first call after a long pause incurs a TCP + TLS handshake. Subsequent calls reuse the connection.

This is normal. If the first call routinely takes more than 5 seconds, something else is wrong; tell us.

"The assistant booked something I didn't want"

Two patterns to avoid this:

State the constraint up front. Don't trust the assistant to remember a constraint from three turns back. Say it again when it matters.

Ask for a preview.

Don't change anything yet. Tell me what you'd do first.

You can then OK it, tweak it, or reject it. Most assistants default to confirming before mutating, but the explicit hint makes it bulletproof.

"Cursor / ChatGPT lost the Jasning tools mid-chat"

Long chats can evict MCP tools from the model's context, especially if you switched topics. Start a fresh chat for a fresh dispatch session.

In Claude, the connector persists across chats. In ChatGPT it's also persistent, but the visibility of the tools can drop within a long conversation. New chat fixes it.

"I can't tell what the assistant actually did"

Every mutating call shows up in the dispatch board within a second or two. If you don't see the change there, it didn't happen. Don't trust the chat summary alone; check the board.

For higher-stakes work, ask the assistant to print what it did:

After each change, show me the appointment id and the new time.

"How do I revoke an assistant's access?"

Two paths, either works:

• In the assistant: Settings → Connectors (Claude, ChatGPT) or Settings → MCP (Cursor) → remove the Jasning entry.
• In Jasning: Settings → Tokens → revoke.

The token stops working immediately. The assistant will need to re-authorize before it can call any tool again.

Still stuck?

Email hello@jasning.com with:

1. Which assistant you're using
2. The exact prompt you sent
3. What the assistant said back

We read every message and a human will reply.