ChatGPT directory release package
Edit this configuration when changing the public listing, connection, review
cases or brand assets. Add an @trizum/mcp changeset for these changes so the
next MCP release contains an updated submission ZIP.
- plugin.json owns the listing name, publisher, descriptions, prompts, public URLs, five positive and three negative review cases, country availability and Spanish listing translation. The empty country list allows all supported countries; set an explicit allowlist if needed.
- mcp.json connects the portable package to the production HTTPS MCP endpoint with no authentication. Use the permanent production URL.
- assets.json maps ZIP paths to selected PNG files, relative to the MCP package directory. The defaults reuse the PWA's existing brand images.
The builder derives version from packages/mcp/package.json and
publication.release_notes from that version's CHANGELOG.md section. Do not
duplicate those values in the configuration. There are no build timestamps or
credentials in the ZIP. Only plugin.json, mcp.json and explicitly mapped image
files are included; source code and unrelated files are excluded.
Build and inspect locally
From the workspace root:
vp run --filter @trizum/mcp package:submission
This builds the MCP package first, validates the configuration and referenced PNG
assets, and writes these ignored files under packages/mcp/artifacts/:
trizum-chatgpt-<mcp-version>.ziptrizum-chatgpt-<mcp-version>.zip.sha256
The MCP check task also builds this archive, so regular CI validates the actual release configuration, version's changelog entry and selected images.
The ZIP has portable plugin.json and mcp.json at its root, plus assets/.
It contains the hosted connection settings and listing materials; the MCP server
and widget continue to run on the existing Worker.
GitHub release attachment
The Release workflow's Attach MCP ChatGPT Submission job runs when a stable
@trizum/mcp@<version> GitHub release is published. It checks out that exact tag,
builds the archive and attaches the ZIP and checksum to that MCP release using
the job's GitHub token. The tag must match the checked-out package version.
Retries replace assets with the same names. PWA and mobile release events do not
run this job, and prereleases are excluded.
The first MCP release containing this packaging workflow will have the assets. Existing releases are not backfilled. Packaging is independent of the production deployment, so check that the production server is running the intended tools before uploading its ZIP to OpenAI.
Complete the submission
Download the ZIP from the MCP GitHub release and follow the official submission guide. This automation creates the package; it does not upload it to OpenAI or publish the listing.
Before submitting, run the configured review prompts against the production app
and confirm the expected behavior. Add a real, reviewer-accessible video URL as
extensions.com.openai.review.demo_recording_url in plugin.json; it is omitted
until a recording exists. The package can be generated and uploaded without that
optional ZIP field, but OpenAI requires the recording for MCP review. Omitting it
preserves an existing portal value. The build prints a reminder while it is absent.
Complete publisher identity verification and the domain challenge in the portal, review its automated tool/UI checks, and submit the draft. Keep credentials and reviewer access details in the portal rather than these public files. The local validator checks our configured package fields, not every OpenAI review rule; portal checks and approval are still required.
The portable manifest and MCP connection follow OpenAI's package format.