ChatGPT’s Custom-GPT integration problems

Hi @utsavmadaan823 ,

Hi @utsavmadaan823,

Sorry for the slow reply. Short answers to all three of your points, then
the setup that actually works.

1. OPENAPI VS MCP — WE DO NOT SOLVE THIS, AND YOUR SHIM IS RIGHT

Our plugin exposes a single JSON-RPC endpoint. Every tool is the same URL
and the same HTTP method, so the Actions builder has no way to tell one
tool from another. Your translation layer is not a workaround for
something we forgot — it is the only thing that can work, and it will keep
working. Nothing below removes it.

Your stack stays as it is:

Custom GPT -> your translation layer (yours) -> Easy MCP AI (ours, MCP)

2. THE DESCRIPTION LIMIT IS OPENAI’S, AND IT IS HARD

https://developers.openai.com/api/docs/actions/production

“300 characters max for each API endpoint description/summary field”
“700 characters max for each API parameter description field”

Of our 244 tools, 184 exceed 300 characters (median 404, longest 1546).
We do not expose a filter for tool descriptions, so the shortening has to
happen in your layer.

You can make that job tiny, though. tools/list is filtered per token, so
create a token under Easy MCP AI -> API Tokens with only the tools this
GPT needs ticked, and the server returns only those — three or four
descriptions to hand-write instead of 244. I verified this on a clean
install: a token with only wp_create_post ticked returns exactly one tool,
and calling anything else is refused with “Token does not have permission
to use this tool”.

For the few that remain:

function my_gpt_clip( $text, $limit ) {
$text = trim( preg_replace( ‘/\s+/’, ‘ ‘, wp_strip_all_tags( (string) $text ) ) );
if ( mb_strlen( $text ) <= $limit ) { return $text; } $cut = mb_substr( $text, 0, $limit – 1 ); $space = mb_strrpos( $cut, ‘ ‘ ); if ( false !== $space && $space > 40 ) {
$cut = mb_substr( $cut, 0, $space );
}
return rtrim( $cut, ” \t,.;:-(” ) . ‘.’;
}

function my_gpt_adapt_tool( array $tool ) {
// Hand-written short descriptions for the few tools you expose.
$short = array(
‘wp_create_post’ => ‘Create a new post from the visitor\’s message. Requires title; content is optional HTML. Saved as a draft for review.’,
);

$name = isset( $tool['name'] ) ? $tool['name'] : '';
$desc = isset( $short[ $name ] )
    ? $short[ $name ]
    : ( isset( $tool['description'] ) ? $tool['description'] : '' );

// An empty description makes ChatGPT pick the wrong tool, so never ship one.
$tool['description'] = my_gpt_clip( '' !== trim( $desc ) ? $desc : str_replace( '_', ' ', $name ), 300 );

if ( ! empty( $tool['inputSchema']['properties'] ) && is_array( $tool['inputSchema']['properties'] ) ) {
    foreach ( $tool['inputSchema']['properties'] as $prop => $spec ) {
        if ( isset( $spec['description'] ) ) {
            $tool['inputSchema']['properties'][ $prop ]['description'] = my_gpt_clip( $spec['description'], 700 );
        }
    }
}

return $tool;

}

Run each tools/list entry through my_gpt_adapt_tool() before turning it
into an OpenAPI operation; our definitions use name, description and
inputSchema, so it maps straight across. Against the real wp_create_post
definition: 881 characters in, 116 out with the override, 284 out on the
truncation fallback. Prefer the overrides — truncation cuts mid-sentence,
and ChatGPT picks tools based on that text.

3. YOUR ANONYMOUS USE CASE ALREADY WORKS — NO OAUTH, NO PLUGIN CHANGE

This is the part you may not know about. I set it up on a clean install
and tested it end to end before writing it down:

a. Create a WordPress user with the Author role. The token screen only
offers users who can publish_posts, so Author is the lowest available
and Contributor will not appear. The easy_mcp_ai_oauth_min_capability
filter changes that if you need something weaker.

b. Create a token bound to that user, ticking only wp_create_post plus
any reads you want. Every post from the shared GPT is attributed to
that one user. Your visitors need no account and never see a login.

c. Turn on Settings -> Force Draft on Create. I called wp_create_post
with status=publish and it came back as a draft, confirmed in the
database. Nothing goes live unreviewed — with a public GPT link,
treat this as required, not optional.

d. Tokens take an optional expiry date, the rate limit is counted per
token (default 60 requests a minute), and successful tool calls are
recorded in the audit log. Refused calls are not logged.

4. WILL WE ADJUST THE PLUGIN?

Two separate things, and I will be straight about both:

• A short description per tool, alongside the full one. Worth doing, it
  is on the list, but I am not promising a release.
• A native OpenAPI bridge (per-tool routes plus a generated schema).
  That is a real feature rather than a quick fix, and it is not planned
  right now. Better you hear that than wait for it.

Neither blocks you. Everything in point 3 works on the current release.

5. ONE THING WORTH KNOWING BEFORE YOU INVEST MORE

ChatGPT can now talk to an MCP server directly through Developer mode, and
our endpoint works there as it stands. It does not replace what you built:
it is configured per account rather than published as a shareable link.

Custom GPTs and MCP are still separate systems in OpenAI’s docs, so your
layer stays the right call for a shareable, anonymous, write-capable GPT.

Thanks for the write-up. Your translation layer is sensible engineering
and should keep working regardless of what we do.