Connect LiquidLM in ChatGPT

Developer-mode MCP setup with OAuth, tool discovery, and search/fetch verification.

Prerequisites

  • ChatGPT web account with developer mode available (Settings → Security and login → Developer mode).
  • A LiquidLM account with at least one indexed vault item you can search for.
  • Use a test org/vault for verification — do not paste real tokens or operator credentials into screenshots.

Turn on developer mode

Open ChatGPT web → Settings → Security and login → Developer mode and enable it. Custom MCP connectors only appear once developer mode is on.

Add the MCP connector

Go to Settings → Apps and Connectors (developer controls), create an MCP-backed app or connector, and paste the LiquidLM endpoint:

  • Complete the LiquidLM browser sign-in and consent screen when it opens.
  • You do not need an API token here — OAuth carries your account and active workspace.
https://mcp.liquidlm.com/v1/mcp

Refresh the tool catalog

After OAuth, run Update actions (or equivalent) so ChatGPT reloads the tool list. LiquidLM exposes ten tools, including the standard search and fetch aliases for knowledge retrieval.

  • Expect ten LiquidLM tools after Update actions, including search and fetch.
  • If tools are missing, disconnect and reconnect once before filing a LiquidLM bug — a lingering partial-permissions banner was observed on 2026-07-26 and should be cleared with a clean reconnect before OpenAI submission.

Verify search and fetch

In a fresh conversation, ask ChatGPT to search LiquidLM for a distinctive vault item you control:

  • Inspect the tool call: LiquidLM.search should run and return grounded results.
  • Follow with a fetch on one result ID — LiquidLM.fetch should return the item body.
  • ChatGPT may not render a visible citation chip even when the server returns an eligible /k/ deep link. That is a client presentation choice, not a LiquidLM contract failure.

Example prompt

Search LiquidLM for the Q3 pricing decision and show me the source item.

Fix common errors

  • Auth errors after org switch: disconnect, reconnect OAuth, and confirm your active LiquidLM workspace in the app.
  • Scope errors: consent must include mcp; LiquidLM still enforces vault permissions per tool call.
  • Empty search: confirm the vault is indexed and your account can see it; try list_knowledge in MCP Inspector if unsure.

Write tools and headless use

  • put_knowledge and forget_knowledge require stronger vault grants. Start read-only for reviewer accounts.
  • For CI or headless automation, create a narrowly scoped API token and send it only in the Authorization header — never in a URL or checked-in config.

Assistants hub

MCP tool reference