AI Newsroom

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "AI Newsroom": {
      "url": "https://jialdowpnekknmxrwrdq.supabase.co/functions/v1/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (8)

• get_top_picks

  Get the top AI news of the last 24 hours (or a window you choose), as the AI Newsroom morning digest ranks it: stories grouped across outlets and judged by an AI editor, in four sections. (1) Top stories: the most important news, rated 6+. (2) Also notable: rated 4+. (3) For builders: developer tools and releases. (4) Paper of note: at most one notable paper. Present the sections separately, each story with its headline, why it matters and its outlets. Each entry is one story: the AI editor's plain `headline`, `why` it matters, `importance` (1-10), `type`, `audience` (general or builder), `story_key`, `followup`, every outlet that carried it (`outlets`, Hacker News links as "Publisher (via HN)"), the representative `url` and `published_at`, plus the original `title`, `source` and `score`. Editor fields are null for items not rated yet. If most of the window is not rated yet, the response says so in `note` and falls back to an older score-based ranking (mode "fallback").

  Parameters:

  • hours — number — Look back this many hours (default 24)
  • since — string — ISO 8601 timestamp: the window starts here instead (overrides hours)
  • showAll — boolean — If true, no per-section or per-outlet limits: every story that qualifies for a section
  • industryLimit — number — Older ranking only: max industry items
  • officialLimit — number — Older ranking only (when most items are unrated): max official announcement items
  • researchLimit — number — Older ranking only: max research items
  • communityLimit — number — Older ranking only: max community items

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• get_trending

  Get recent AI news (the last 7 days by default), most important first, optionally filtered by source, category, or the AI editor's importance, type and audience. Each entry is one story: the AI editor's plain `headline`, `why` it matters, `importance` (1-10), `type`, `audience` (general or builder), `story_key`, `followup`, every outlet that carried it (`outlets`, Hacker News links as "Publisher (via HN)"), the representative `url` and `published_at`, plus the original `title`, `source` and `score`. Editor fields are null for items not rated yet.

  Parameters:

  • type — string — Only stories of this type, as judged by the AI editor.
  • limit — number — Max items to return (default 50)
  • since — string — ISO 8601 timestamp: only items collected after this (default: 7 days ago)
  • source — string — Filter by source ID (e.g., "anthropic_blog", "techcrunch_ai")
  • audience — string — "general" for news anyone following AI wants; "builder" for developer tools, SDKs, tutorials.
  • category — string — Filter by category ("company_blog", "community", "research", "github", "industry_news")
  • min_importance — integer — Only stories the AI editor rated at least this important (1-10). 7+ is significant news, 9-10 is rare.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• search

  Search AI news by keyword across titles, the AI editor's headlines and summaries. Results are grouped into stories, most important first. Each entry is one story: the AI editor's plain `headline`, `why` it matters, `importance` (1-10), `type`, `audience` (general or builder), `story_key`, `followup`, every outlet that carried it (`outlets`, Hacker News links as "Publisher (via HN)"), the representative `url` and `published_at`, plus the original `title`, `source` and `score`. Editor fields are null for items not rated yet.

  Parameters:

  • query (required) — string — Search keywords
  • type — string — Only stories of this type, as judged by the AI editor.
  • limit — number — Max stories to return (default 30)
  • since — string — ISO 8601 timestamp — only return items after this time
  • audience — string — "general" for news anyone following AI wants; "builder" for developer tools, SDKs, tutorials.
  • min_importance — integer — Only stories the AI editor rated at least this important (1-10). 7+ is significant news, 9-10 is rare.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• get_new_since

  Get everything collected after a specific time, grouped into stories, most important first. Perfect for 'what's new since yesterday?' or 'what did I miss this weekend?' Each entry is one story: the AI editor's plain `headline`, `why` it matters, `importance` (1-10), `type`, `audience` (general or builder), `story_key`, `followup`, every outlet that carried it (`outlets`, Hacker News links as "Publisher (via HN)"), the representative `url` and `published_at`, plus the original `title`, `source` and `score`. Editor fields are null for items not rated yet.

  Parameters:

  • since (required) — string — ISO 8601 timestamp (e.g., "2026-04-05T00:00:00Z")
  • type — string — Only stories of this type, as judged by the AI editor.
  • limit — number — Max stories to return (default 50)
  • audience — string — "general" for news anyone following AI wants; "builder" for developer tools, SDKs, tutorials.
  • min_importance — integer — Only stories the AI editor rated at least this important (1-10). 7+ is significant news, 9-10 is rare.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• get_source_updates

  Get the latest items from a specific source, the most important of its recent items first. Each entry is one story: the AI editor's plain `headline`, `why` it matters, `importance` (1-10), `type`, `audience` (general or builder), `story_key`, `followup`, every outlet that carried it (`outlets`, Hacker News links as "Publisher (via HN)"), the representative `url` and `published_at`, plus the original `title`, `source` and `score`. Editor fields are null for items not rated yet.

  Parameters:

  • source (required) — string — Source ID (e.g., "anthropic_blog", "claude_code_releases")
  • type — string — Only stories of this type, as judged by the AI editor.
  • limit — number — Max items to return (default 20)
  • audience — string — "general" for news anyone following AI wants; "builder" for developer tools, SDKs, tutorials.
  • min_importance — integer — Only stories the AI editor rated at least this important (1-10). 7+ is significant news, 9-10 is rare.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• get_repo_quickstart

  Look up a GitHub repo and return its metadata (stars, language, license, last push) plus the install/usage section pulled out of its README. Use this when the user asks how to get started with, install, or try out a repo they saw in the news.

  Parameters:

  • repo (required) — string — Repo as "owner/name" (e.g. "anthropics/claude-code"). A github.com URL also works.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• get_paper_brief

  Fetch an ArXiv paper by ID and return its title, authors, abstract, categories, and links. Use this when the user asks what a paper is about after seeing it in the research section.

  Parameters:

  • arxiv_id (required) — string — ArXiv identifier (e.g. "2501.12345" or "2501.12345v2"). An arxiv.org abs/pdf URL also works.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• check_status

  Check the health of the news cache: when it was last updated, total items, per-source breakdown, how much of the last 24 hours the AI editor has rated (share judged, newest judgment, rubric versions), and which commit of the server is deployed.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.