A real public MCP endpoint now exists
The endpoint supports JSON-RPC initialization, tool discovery, tool calls, notifications, ping, and batched requests over HTTP POST. GET correctly declines server-sent events because the current server has no server-initiated notifications.
Every result comes from the same public synthetic fixture used by the browser lab. The endpoint cannot access Stripe, collect keys, authenticate a customer, or change external state.
Eleven focused tools create an AI-native contract
Seven founder-metric tools cover revenue, customers, products, subscriptions, transactions, refunds, and forecasts. The compare-founder-scenario tool adds deterministic counterfactual analysis without claiming prediction.
The standard search and fetch tools make the public evidence corpus compatible with citation-oriented knowledge workflows. The Civic Proof tool screens AI systems and policies through public rules and returns exact constitutional citations.
All tools set readOnlyHint true, destructiveHint false, openWorldHint false, and idempotentHint true because they only retrieve public evidence or compute from published rules and synthetic data.
Resources and prompts expose more than tool calls
Twenty MCP resources publish the highest-value identity, methodology, benchmark, civic, project, evidence-network, discovery-research, and distribution records with audience, priority, and last-modified annotations.
Three user-controlled prompt templates guide source-backed founder analysis, metric-claim challenges, and civic proposal review. The server instructions tell compatible clients to search, fetch, calculate, screen, cite, and preserve each boundary.
The OpenAI review package is reproducible
The submission dossier contains listing copy, policy URLs, annotation justifications, starter prompts, eleven positive cases, five negative cases, availability, release notes, and the exact owner actions still required.
FounderMetricBench adds twenty-four gold cases. CivicProofBench adds one hundred CC0 cases spanning eight constitutional principles and cross-article review, with exact expected scores and dispositions.
The MCP Registry record is prepared
The public server.json uses the official remote-server schema, a GitHub-auth-compatible namespace, and the production Streamable HTTP URL.
Publishing still requires Marco to authenticate with the Registry under the matching GitHub identity. The Registry entry must not be described as published until that account-owned step succeeds.
Four account-owned actions remain
Marco must complete or select a verified OpenAI developer identity, confirm Apps Management write permission, create the plugin draft and scan the production tools, then provide the generated domain-verification token for this site to host.
After the scan and test set pass, Marco can submit the plugin. The application must remain described as pending until OpenAI approves and publishes it.
SOURCE HIERARCHY
Verify at the origin.
- OpenAI plugin submission dossierMachine-readable prepared materials↗
- MCP evaluation setEleven positive and five negative cases↗
- FounderMetricBenchTwenty-four gold evaluation cases↗
- CivicProofBenchThirty-two CC0 civic evaluation cases↗
- AI evidence graphMachine-readable claim and proof topology↗
- MCP security modelPublic boundary and control record↗
- Official MCP Registry recordPrepared remote-server metadata↗
- OpenAI submission requirementsOfficial current requirements↗
- Official MCP Registry remote-server guideOfficial Registry documentation↗
Prepared submission materials are not proof of review, approval, publication, endorsement, or Registry listing. Those statuses change only after the respective external platform confirms them.