This page is for the people who run the site. The admin MCP server lets an AI agent such as Claude Code read the admin data and take the same actions you take on the admin pages: review reports and template suggestions, answer support threads, adjust credits and edit the prompt library.

Create a token

  1. Open Admin, then Agents.
  2. Name the token after where it will live, for example "Claude Code on my laptop".
  3. Tick only the scopes the agent needs, then press Create token.
  4. Copy the token right away. It starts with pgadm_ and is shown once; only a fingerprint of it is stored.

Revoke a token on the same page when you no longer need it. A token also stops working when its owner is removed from the admin list.

Connect Claude Code

Put the token in an environment variable named PROMPTGENERATOR_ADMIN_TOKEN, then add the server:

Prompt
claude mcp add --transport http promptgenerator-admin https://trypromptgenerator.com/api/mcp --header "Authorization: Bearer $PROMPTGENERATOR_ADMIN_TOKEN"

In the repository you can instead copy .mcp.json.example to .mcp.json. Any client that speaks the streamable HTTP transport and can send an Authorization: Bearer header works the same way.

Scopes

ScopeWhat it allows
ReadOverview, users, jobs, reports, blocked attempts, template suggestions, gallery shares, support, credits ledger, library, analytics and the audit log
ModerateHide and unhide gallery items, dismiss reports, suspend or resume a user's generation, unpublish a share
ContentApprove or reject template suggestions; create and edit library prompts and hubs on a local copy of the site
SupportReply to support threads and change their status
CreditsGrant or remove credits and refund a job's credits
RendersStart paid library renders under a dollar budget
EmailSend one email to one account
StripeLook up a customer, their subscriptions and invoices, read-only

Confirm before anything that costs or sends

Tools that move credits, send an email, spend money on renders or take something off the gallery do nothing on the first call. They answer with a preview, for example the current balance and the balance after, and run only when called again with confirm: true. Ask your agent to show you the preview before it confirms.

Renders also need a budget_usd. A run whose estimate is over that budget, over $25, or over the daily render cap is refused before anything starts. Renders are paid from your own account's credits.

Library edits

The prompt library lives in the site's code, so the library tools write files only on a local copy of the site. Run them against your local server, check the changes, run pnpm library:audit, then commit and deploy. On the live site they refuse.

Audit log

Every call is recorded on the Agents page: who made it, with which token, the tool, its input with passwords, tokens and keys removed, and whether it ran, was refused or failed. Agents can read the same log with the list_admin_audit tool.