Verified by NPC: OpenAI (ChatGPT) plugin package

Status: 1.0.0, package validated, NOT submitted. No developer account has been created, no submission has been made, the demo API bundle is built and locally verified but not yet deployed to api.npclabs.xyz. Everything here runs against the local API or documents the gated production path.

What this is

The "Verified by NPC" ChatGPT plugin answers the question "is this real?" inside the conversation. When verification becomes relevant mid-chat, the plugin offers five operations:

Design principles (same as the API and the MCP server): read-first, so the plugin verifies and never issues attestations; fail-closed, so an unknown subject returns "unknown" and never a guess; privacy-preserving, so responses carry no PII beyond what the subject made public; and self-verifying, so every API response is Ed25519 signature-checked before it is trusted. explain_verification teaches the why, never the how.

How it maps to API v1

This package is a thin client over the NPC Verification API v1 (../../api/). The mapping lives in three places:

• openapi-plugin.yaml: the machine-readable action surface. The four verification operations map 1:1 onto API v1 endpoints (paths are checked against the API's own openapi.yaml by the test suite). explain_verification is marked x-npc-client-side: true: a static operation served by the plugin, never sent to the API. Issuance (POST /v1/attestations) is deliberately absent from this spec.
• skills/verified-by-npc/SKILL.md: the onboarding skill the plugin manifest points at. Human- and model-readable description of the five operations, the fail-closed rules, and the HTTP details.
• npc_verify_openai/client.py: the thin Python helper (stdlib only). It reuses the Phase 2a MCP package's verified HTTP client and tool handlers, which reuse the API's own signing primitives. The reuse chain is npc_verify_openai → npc_verify_mcp.api_client → npc_verify.*. No verification logic is duplicated, and no crypto is reimplemented. Every operation returns (is_error, text); tampered or untrusted responses surface errors, and unknown subjects surface exactly "verdict": "unknown".

The 2026 submission path (researched 2026-10-06)

OpenAI retired the 2023-era ai-plugin.json manifest: the official quickstart repo is archived, the manifest docs are gone, and the legacy file fails the current submission scan. This package therefore does NOT ship ai-plugin.json. The live 2026 format is the portable Agent Plugins manifest, which is what plugin.json is:

• Root plugin.json with $schema from agent-plugins.org, package identity (verified-by-npc), and OpenAI listing metadata under extensions.com.openai.interface (display name, descriptions written for the mid-conversation suggestion trigger, starter prompts, icons).
• skills/ auto-discovered by the platform; the manifest's onboardingSkill points at the bundled skill.
• Root mcp.json declares the hosted production MCP server in the documented portable shape: type streamable-http, the endpoint URL https://api.npclabs.xyz/v1/mcp, and a no-authentication declaration (extensions.com.openai.auth type none) that the portal reads. Connecting the server in the portal also involves a domain verification challenge. The platform does not support adding an MCP server to a plugin that was created skills-only, so a package with an MCP server must carry mcp.json in its initial upload.
• Logo and composer-icon fields point at internal placeholder SVGs in assets/. They are NOT the "Verified by NPC" mark; public mark use awaits counsel clearance. Replace them with approved artwork before any public submission.
• Review metadata ships in the manifest: five positive and three negative test cases, commerce declarations, and the demo video walkthrough URL (review.demo_recording_url, required for MCP review per the submission field reference). Reviewer credentials are entered only in the portal's secure form; the package format rejects credential fields outright.

Submission flow (portal, not done): create an OpenAI developer organization with Apps Management write access, complete verified individual or business identity, upload the plugin ZIP, resolve the automated metadata findings, fill in review details, submit for review, then publish after approval. Each of those steps needs Bayo's explicit go-ahead.

Validate

cd plugins/openai
python3 validate_package.py   # full end-to-end check; nonzero exit on failure
python3 -m unittest discover -s tests   # the 52-test suite on its own

validate_package.py runs the tests, checks that every manifest reference resolves inside the package, enforces the copy rules (no em-dashes, no invented live URLs, no money or earnings claims), confirms the retired ai-plugin.json is absent, and assembles a trial submission ZIP.

What remains gated (explicitly NOT done)

1. OpenAI developer account + verified identity (Bayo's decision; NPC Labs business identity approved 2026-10-06).
2. Store listing copy approval (Bayo reviews every public word; copy approved 2026-10-06, release notes rewritten for 1.0.0).
3. Submission + publication in the plugin portal.
4. Public endpoint https://api.npclabs.xyz/v1 - demo API bundle built and locally verified 2026-10-06 (api/serverless/); the Netlify Drop deploy + DNS is Bayo's tap.
5. Approved "Verified by NPC" mark artwork (counsel clearance; placeholders ship for now).
6. Demo video walkthrough link for the review form (2-minute captioned video built 2026-10-06; Bayo to host it as an unlisted video). Country availability: United States (portal entry, Bayo's tap).
7. Mainnet onchain anchoring (testnet until Bayo signs off).

Files

• plugin.json: the 2026 live plugin manifest.
• openapi-plugin.yaml: OpenAPI action spec for the five operations.
• skills/verified-by-npc/SKILL.md: onboarding skill.
• assets/logo.svg, assets/icon.svg: internal placeholders, not the mark.
• npc_verify_openai/client.py: thin stdlib helper reusing the MCP client.
• tests/: 52 tests: manifest, spec, signature, contracts.
• validate_package.py: end-to-end package check.