Paradigm Threat exposes a read-only Model Context Protocol server so agents can search investigations, match site thesis cards, check research coverage, and compute timeline dates — grounded in the same Relay corpus as the site, with explicit gaps when a topic is not indexed.
https://paradigmthreat.net/api/mcp
Status: enabled on this deployment
search_site / search — keyword search over articles and timelineresearch_status — coverage envelope (including not_in_corpus)get_site_claims — curated thesis cardsread_article / fetch — markdown from Relay mirrorsget_event_dates / compute_years_ago — chronology and blip-adjusted elapsed timesubmit_investigation_idea — propose a new topic for the site to research (appears on /submit)submit_chat_comment — post to a chat channel on Relay chat/public branch{
"mcpServers": {
"paradigm-threat": {
"url": "https://paradigmthreat.net/api/mcp"
}
}
}Settings → Apps → Create app → Server URL https://paradigmthreat.net/api/mcp, authentication No Auth. Add the app as a source in chat before asking research questions.
The MCP endpoint is public and read-mostly, but the host model may refuse to call tools — especially on free or consumer tiers (ChatGPT, Gemini, etc.). Common blocks: real people and alleged crimes or deaths, defamation-shaped framing, or write tools without explicit user confirmation. That happens in the host's safety layer before a request reaches paradigmthreat.net; connector installed and tools listed does not guarantee every query will run.
Neutral corpus prompts often work better: search for a title or codename, list paths only, or call research_status without asking the model to assert claims. Cursor and other dev-oriented clients are usually less restrictive than free ChatGPT.
docs/MCP_VERCEL_DEPLOY.md.