Connect ChatGPT

Prerequisites

• The CodeMonster desktop app, signed in with your organization license. It shows the server URL and the token for your session. Desktop builds are being finalized for launch — request access and we send the installer link the day the build for your operating system ships.
• ChatGPT, signed in on your own plan. That traffic goes to your AI provider under your account and does not route through us.
• MCP connectors are enabled per workspace and can sit behind developer mode or an admin approval, so check with whoever administers your workspace before you start. We have not live-tested this path, and the exact setting names are OpenAI’s to change.
• ChatGPT connects from the network rather than your machine, so the desktop app is where you sign in and copy values from.

Configuration

ChatGPT has no configuration file: open its connector settings, add an MCP connector, and enter three values. Both placeholders below come from the desktop app after sign-in.

Name — CodeMonster
MCP server URL — <server URL shown in the desktop app>
Authorization — Bearer <token shown in the desktop app>

That token acts with your permissions on every call, so treat it like a password: it belongs in the connector’s authorization field, never in a prompt, a shared chat or a repository. How ChatGPT presents that field is a detail the live test will settle; if it is missing in your build, stop and tell us rather than pasting the token anywhere else. Enable the connector in a new chat and confirm its tools appear.

First calls

Ask ChatGPT to run these in order. whoami confirms the account; list-tenants and then use-tenant set the tenant context, and no data write goes through without it.

get-domain-knowledge hands the agent the platform’s conventions — build order, object prefixes, groups first, publish to land. list-solutions shows what exists; create-solution starts something new.

Verify it worked

• whoami returns your email and your account. A 401 here means the token is wrong; see below.
• list-tenants returns at least one tenant, and use-tenant echoes the one you chose.
• A data write attempted before use-tenant fails fast with Tenant context required. That failure is the guardrail working.
• Tenant context belongs to the connector session, so a fresh one starts without a tenant — run list-tenants and use-tenant again.
• After get-domain-knowledge, the agent’s next plan uses the platform’s own vocabulary: objects, forms, screens, groups.

Troubleshooting

• The connector shows no tools — the server answered 401 during discovery. A rejected request returns no tool list, no schema and no server info by design. Copy the token again from the desktop app and check the Bearer prefix survived.
• 401 mid-session — the token expired or was revoked. Sign in again in the desktop app and update the connector.
• Tenant, transaction and freeze errors — Tenant context required, concurrent_mutation_blocked, staged edits vanishing after 60 idle seconds, and 403 on a solution frozen against AI edits behave the same on every client. Each one, and what to do about it, is documented once on the control surface.

Next

• Build something real: the field-service inspection blueprint is a starter prompt plus the primitives map.
• Read the control surface for every operation family and the transaction lifecycle.
• The full reference is being published on codemonster.dev; /docs tracks its status.