Your prompts, inside Claude

You're in a Claude chat, and you notice the support prompt your app serves still tells the model to refuse refunds. That policy changed two months ago. You say so: refunds should be escalated, not refused, and publish it.

Claude looks through your projects, finds the prompt, reads the production version, rewrites the rule, saves a new version and publishes it. Your app serves the new wording on its next request.

> Our support prompt still refuses refunds. Make it escalate refund requests
  to a human instead, then publish.
 
  list_projects   -> Customer Support
  list_prompts    -> Support Agent (production 38e8675)
  get_prompt      -> version "production"
  update_prompt   -> version 108e907, publish: true
                     production is now 108e907 (was 38e8675)

Recorded in claude.ai and SuperPrompts against a demo workspace. Unmute for sound. The tool calls above are from a Claude Code run of the same change.

That's SuperPrompts as a Claude connector. It's the same hosted MCP server we shipped with SuperPrompts 2, but you no longer need to copy a key anywhere. You sign in, approve once, and you're done.

Connect in one click

SuperPrompts is listed in Claude's connector directory, so you can also search for it under Customize → Connectors in Claude and click Connect.

Add to Claude opens Claude's custom connector dialog with the name and URL already filled in. Claude asks you to confirm the URL, you press Connect, and you're sent to SuperPrompts to sign in.

Then you get a consent screen. It shows which app is asking, the address you'll be sent back to, the account and organization you're signed in as, and what you're granting: full access to every project in that organization, including ones created later. Read, create, update and publish prompts. Approve or deny.

That scope is wide, and we'd rather you see it before you click than find out afterwards. A connection acts as you, across the whole organization. If you want something narrower, a project API key reaches one project only (more on that below).

If you use a Team or Enterprise plan on Claude, a workspace owner adds the connector first, and then members connect to it.

Prefer a terminal? One line for Claude Code, then run /mcp and pick SuperPrompts to sign in:

claude mcp add --transport http superprompts https://superprompts.app/api/v2/mcp

Cursor, ChatGPT and VS Code take the same URL. Each one's setup is in the MCP docs.

What you can ask

There are six tools: list_projects, list_prompts, get_prompt, create_prompt, update_prompt and publish_prompt. You don't name them. You ask for what you want, and the client picks. A few things that work:

• "What projects and prompts do I have?" Claude lists your projects, then the prompts in each, with the production and latest version of every one.
• "Show me the production version of the onboarding prompt." You get the text, its sections, any tools it defines, and the {{ variables }} it expects. Ask Claude to fill the variables in and it can compile a copy with real values so you can see what the model would receive.
• "Create a prompt for refund emails from this markdown, and don't publish it yet." It's saved as a draft. Production doesn't move until you say so.
• "That last change made things worse. Go back to the previous version." Every save is a version, so rolling back is publish_prompt pointing production at the earlier hash. Nothing is deleted, and the version you rolled back from is still in the history.

Changes made this way land in the same version history as edits you make in the dashboard: same commit message field, same diff, same rollback.

Knowing what happened

Giving a chat window the ability to publish to production deserves a paper trail, so we built one.

Settings → Activity lists everything done in your organization from the dashboard, the API and MCP. Each entry shows who did it, which app they did it through, what it was, the version before and after for prompt changes, and the IP address. You can filter by project, category and source, show only failures, and export to CSV. Entries are kept for 90 days. Prompt content is never stored in the log, so reading it won't expose your prompts a second time.

Settings → Connected apps lists every app you've approved, with a Revoke button next to each. Revoking takes effect within about a minute. Access tokens last an hour, and a revoked connection stops being accepted well before that.

More on how sign-in works, including PKCE and the consent screen, is in the security docs.

API keys haven't gone anywhere

OAuth is for people sitting in a client. For a CI job, a script, or anything headless, a project API key is still the right tool. It lives in Project → Settings → API, goes in an x-api-key header, and reaches exactly one project. The same endpoint accepts either.

Use the connector for the interactive work and keys for code. The REST API, the Node SDK and the MCP server all write to the same place, so a version saved from a chat is the one your CI job reads.

Try it

Add the connector, ask Claude to list your prompts, and have a look at Activity afterwards to see what it recorded. If something looks off, revoke it.

The MCP docs cover each client. If you don't have an account yet, sign up. The free plan includes the MCP server and the connector, and it doesn't need a card.