Connect your AI client

MCP cannot intercept a conversation. The logger is passive: it records only what the model chooses to send it. This experiment measures whether the model voluntarily invokes the logging tool on every turn after being instructed to.

Missed calls are an expected outcome, not a bug. They surface in the dashboard as missed turns and lower the compliance percentage.

The MCP server runs inside this Logbook deployment — there is nothing separate to host. Both vendors connect from their own cloud, so use the public HTTPS URL below, never localhost or a stdio server.

https://llm-log-book.lovable.app/mcp

Fallback mount (identical behaviour, same token auth) if a client cannot reach the primary path: https://llm-log-book.lovable.app/api/public/mcp

Railway or any other standalone host is a fallback only, to be used if this runtime turns out to be incompatible with a client. The Supabase schema, token model and tenant checks would move unchanged.

The endpoint rejects unauthenticated requests. Authentication is a personal bearer token scoped to exactly one user and one workspace.

1. Open MCP credentials while signed in and generate a token.
2. Copy it once — only a SHA-256 hash is stored, so it cannot be shown again.
3. In the client's connector dialog choose bearer / access-token authentication and paste the token. Some clients label this “Authorization header”; the value is Bearer <your token>.
4. Revoke or rotate the token at any time; revocation takes effect immediately.

Your token is shown only to you while signed in, masked by default. It is never embedded in this page, logged, or visible to another workspace.

Paste this as the first message of a fresh conversation, after the connector is enabled for that chat.

Use the LLM Logbook MCP server for this entire conversation. Before answering each user message, call the logging tool to record the user turn. Before completing each assistant turn, call the logging tool to record your full assistant response. Continue doing this for every turn unless I explicitly end the logging experiment. Do not claim a turn was logged unless the tool call succeeded.

• Sign in to Logbook.
• Confirm the MCP server is deployed and healthy.
• Copy your personal connector URL/token if authentication is token-based.
• Add the connector in ChatGPT Business/Enterprise/Edu or Claude.
• Start a NEW conversation.
• Enable/select the connector in that conversation.
• Paste the logging instruction.
• Exchange at least 10–20 turns, including some long responses and topic changes.
• Return to Logbook and inspect compliance and missed turns.