[InlinePizza]

Summary

What the page is for now: it gets a Promptless member who uses Claude (web or desktop) or ChatGPT from no connection to a working, tested Promptless connector, and tells them what the connection can access before they approve it.

Why it changed: the Claude connectors directory listing went live on 2026-10-14 and links to this page, but the page only described adding a custom connector. It also opened with GitHub PR review and protocol tables before the reader got to any setup step.

The URL and slug are unchanged. The title is now "Connect Promptless to Claude or ChatGPT". The full audit (chunk inventory, diagnosis, brief, options) is in tmp/page-audit-claude-and-chatgpt.md on the author's machine, and the essentials are below.

Commits:

1. Reorder sections only, with no rewording, so the move is easy to review.
2. Rewrite.
3. Update the inbound link text on MCP triggers.

What moved where

Before	After
Intro: Promptless works through PRs	Cut. How Promptless works covers it
Intro list of five other review surfaces	One routing line for editor users. The sidebar covers the rest
Server details table + chat-vs-editor URL paragraph	Reference > Server details (end of page)
Requirements	Before you start
Data access (after setup)	What the connection can access (before setup)
Claude custom-connector steps	Directory steps first. Custom connector becomes a short fallback that links to Anthropic's steps
Claude org-plan steps	Organization plans: separate Owner and member procedures, plus the Request path
ChatGPT developer-mode steps	Plugins > Add custom MCP server steps. Developer mode kept as an "earlier layout" note
Approve the connection	Kept, using the page's real labels
Tools (between approval and testing)	Reference > Tools
Test prompts	Right after approval. The wrong-URL hint moved to Troubleshooting
Reconnect / disconnect, Troubleshooting	Kept, with current Claude labels

New content and sources

• Claude directory steps (individual, Owner, member, Request). Source: Anthropic's Get started with connectors and Connectors directory, read 2026-10-16.
• Updated Claude labels: Needs approval (was "Ask each time"), Disconnect vs Remove, Reconnect. Same source.
• Custom-connector OAuth client choice (Register automatically). Sources: Anthropic's Add a connector by URL and runtime mcp_oauth/routes.py (dynamic client registration).
• ChatGPT Plugins flow. Source: a search summary of OpenAI help article 12584461. The article itself returns 403, so this needs a check.
• message_promptless_support in the tool table. The chat URL serves 20 tools (runtime/mcp/profiles.py, support_tools.py). The tool's access bullet is new too.
• Approval page labels: Returns to, Organization, Your role, Create organization and authorize as {email}. Source: promptless-dashboard/src/app/mcp/authorize/* on runtime origin/main.
• Next steps and frontmatter brief fields: purpose, audience, out_of_scope, depends_on_external_ui, last_verified.

I verified the token lifetimes, the 30-minute approval expiry, membership checks, and tool annotations against runtime origin/main (77e618c, 2026-10-16). No facts changed.

Needs verification

These are marked as TODO(verify) MDX comments in the page, so readers don't see them.

• The Claude directory listing uses https://api.gopromptless.ai/mcp/with-direct-edits. Check claude.ai/directory/manage. (Eng)
• Directory connectors are available on the Claude Free plan. (Vendor docs or a test account)
• Whether Promptless supports Claude's "Use Claude's published identity" (CIMD), or whether Register automatically is required. Runtime advertises DCR only. (Eng)
• Owners can block specific tools for members in Organization settings > Connectors. This was carried over from the old page. (Claude Team account)
• Every label in the ChatGPT Plugins steps, and whether the Promptless tab opens at Create as a plugin or at install. (ChatGPT account)
• Whether ChatGPT Plus can add custom MCP servers at all. OpenAI's current text names only Pro, Business, Enterprise, and Edu. (ChatGPT account)
• ChatGPT reconnect and disconnect labels in the Plugins layout. (ChatGPT account)

Collateral changes

• connect/triggers/mcp.mdx:25: updated the link text to the new title.
• Regenerated this page's OG image and its manifest entry for the new title. I reverted the build's unrelated OG and route-manifest changes.
• No redirects, because the slug is unchanged. No sidebar change, because the label is still "Claude & ChatGPT".
• Kept the anchors #approve-the-connection, #test-the-connection, #reconnect-or-disconnect, and #troubleshooting. No anchor on this page is known to be linked from outside.

Checks

• Vale: 0 alerts on the page.
• remark-lint auto-fix: only reflowed tables.
• docmeta: 131/131 passed.
• npm run build: passes.
• Internal links and anchors on the built page: all 112 resolve.
• npm run test:smoke: 27/28. The one failure, agent score landing renders its form…, expects the old "How well can…" hero copy from before Trim the Agent Score hero and scan progress copy #1159, so it's unrelated to this PR.
• Length: about 2,950 words, up from about 2,730, because of the directory steps and the separate Owner and member steps. A Claude individual-plan reader now reaches step 1 after about 420 words instead of about 1,000, and their whole path is about 1,080 words. The ChatGPT path is about 1,280.

Decisions made without review (Checkpoints: skip)

• Audited work-the-queue/claude-and-chatgpt.mdx. The Inputs named connect/chat-apps.mdx, which doesn't exist. The published URL matches this file.
• Kept one page instead of splitting by vendor. The URL is locked and linked from the directory. The approval and access sections are shared, and each vendor section is short.
• Kept the tool table on this page, in a Reference section at the end, even though MCP triggers owns the tool details. It's the only list of the 20 chat-URL tools with their read/write type, which Owners and directory reviewers need on the page the listing links to. Tool arguments stay on MCP triggers.
• Couldn't read the KB style guide from this environment. I followed the repo's Vale package and the style of neighboring pages.
• Rewrote the ChatGPT steps to the Plugins flow based on OpenAI search summaries, not a live check. All of them are flagged above.
• Omitted owner from the frontmatter because I don't know who owns the page.

Out of scope, worth fixing

• MCP triggers page:
  • Its Claude section uses /mcp and the old custom-connector steps.
  • It says both URLs carry 11 tools, but the runtime mounts 14 on both.
  • It doesn't cover get_account_status, connect_integration, or message_promptless_support. docs: Document agent sign-up, account claiming, and connecting integrations over MCP #1083 partly covers this.
• Consent page copy (mcp-authorize-views.tsx) tells web-app users to "run the connect command in your editor" and says it's "Returning you to your editor". The page now warns about this in Troubleshooting.
• Dashboard: the Editor connection (MCP) settings section has no Claude.ai or ChatGPT guidance.
• Content strategy: no CUJ covers using Promptless from a chat app.
• Open PR docs: Add a dedicated Promptless MCP server page #1109 edits this page's old intro list, which this PR removes. Rebase docs: Add a dedicated Promptless MCP server page #1109 after this merges.
• Smoke test: the agent-score assertion is stale since Trim the Agent Score hero and scan progress copy #1159.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
promptless-docs	Ready	Preview	Oct 16, 2026 8:10pm UTC