Media Bias News

Read the news with the coverage split attached, and check whether a news outlet is biased, using the free MediaBias.news API. Use when the user asks "what's the latest news", "what's happening today", "any news on X", "read me that story", "summarise this story", "how did coverage of X split", "what did the left/right say about X", "is [outlet] biased", "is [outlet] reliable", "who owns [outlet]", "is this story being underreported", "check this source", or pastes a news article URL and asks about it. Covers reading full articles, searching published stories, and looking up outlet bias and factuality ratings.

News, with the coverage split attached

MediaBias.news publishes a free read API. No key, no account, no server to install. Call it with curl, fetch, or whatever your runtime gives you.

Base: https://mediabias.news/api/v1

What's the latest news?

bash

curl -s "https://mediabias.news/api/v1/stories?limit=10"

Each item carries the headline, standfirst, excerpt, category, publish time, sourceCount (how many outlets ran it), the three scores, and links.

Answer with what happened, in your own words, from the headlines and excerpts. Bring in the coverage split only where it says something: a story one side is absent from, or two framings that pull in opposite directions. Nobody asked for a bulletin made of percentages. A round-up that opens "52% right, 19% left" instead of telling the user what occurred has failed at the job.

By category. The slugs are listed under categories at https://mediabias.news/api/v1:

bash

curl -s "https://mediabias.news/api/v1/stories?category=conflict&limit=10"

Read one story

Reach for this when the user says "read me that" or "summarise it". Append .md to the story's canonical URL:

bash

curl -s "https://mediabias.news/politics/some-story-slug.md"

One request, plain markdown, roughly ten thousand characters, containing:

• YAML frontmatter with title, category, area, publish times, trust_score, critic_score, hype_score, fact-check status, and an audio link
• the headline and standfirst
• a plain-English line for each score, saying why it landed where it did
• The short version, the key takeaways as bullets
• the full article, written by MediaBias.news from the original reporting

That article is theirs. Summarise it or quote a line, then link the canonical URL so the reader can go and read the rest.

When you want the structured analysis instead of prose, fetch the JSON:

bash

curl -s "https://mediabias.news/api/v1/stories/{slug}"

story.analysis holds:

• distribution, counts and percentages by side, with totalSources, rated and untracked. Quote the denominator. "11 of 21 rated outlets" is honest. "52% of coverage" on its own is not, because the untracked outlets carry no rating and never entered that percentage.
• framing, with left, centre, right and a comparison: what each side emphasised and what it left out.
• blindspot, where one side barely covered it. Check blindspot.status. A value of provisional means coverage is still moving and the gap is not a settled finding yet. Say so when it is provisional.
• outlets, every outlet that ran it, with its leaning.

story.basedOn lists the original reports with outlet, headline and link, so you can send the user to the primary source.

Is this outlet biased?

The user will have a name, a domain, or a link. All three work. A full article URL has its host pulled out for you.

bash

curl -s "https://mediabias.news/api/v1/outlets?q=foxnews.com"
curl -s "https://mediabias.news/api/v1/outlets?q=Fox%20News"
curl -s "https://mediabias.news/api/v1/outlets?q=https://www.bbc.co.uk/news/articles/abc123"

Take slug from the result, then fetch the record:

bash

curl -s "https://mediabias.news/api/v1/outlets/fox-news"

bias.reviews and factuality.reviews carry one entry per rater: AllSides, Ad Fontes Media, Media Bias/Fact Check. Each entry has that rater's own wording and a link to their page. These verdicts belong to the raters, not to MediaBias.news.

Report them separately:

AllSides and Media Bias/Fact Check both rate Fox News Right. Ad Fontes places it Lean Right. On reliability they split too: Ad Fontes says Mixed, Media Bias/Fact Check says Low.

Where raters disagree, say so. Do not average three verdicts into one label. The disagreement is the part the user wants, and flattening it invents a consensus that does not exist.

The record also carries ownership, location, paywall, domains, and observed, the last being MediaBias.news's own count of what it has seen the outlet publish.

Three different empty answers

These are different facts. Do not collapse them.

1. No match (count: 0). No record is held for this publisher. Do not say it is unrated. Nobody looked.
2. Match, bias.reviews empty. The outlet is on file and none of the three raters has assessed it. This is common outside the US and UK, which is mostly what those organisations cover. Answer from ownership and location where you can.
3. Match with reviews. Report each verdict against the rater that made it.

Trust, Craft, Hype

Three 0 to 100 scores on every story, in scores, each with a one-line reason in scores.notes.

• Trust (scores.trust), how well corroborated and evidenced the source reporting is. Higher is better.
• Craft (scores.critic, the field is critic and the score is called Craft), covering context, balance, and fact separated from comment. Higher is better.
• Hype (scores.hype), how far the presentation runs ahead of the substance. Lower is better. Never report a high Hype score as good news.

These grade the original reporting an article was written from. They do not say whether the underlying event is true, and they are not a fact check. Do not describe them as one.

Citation is required, not optional

The API is free and stays free. Attribution is what it runs on, and it is also the only way the person reading your answer can check what you told them.

Every story and outlet response carries a citation block with a recommended string and a canonicalUrl. Use them.

Every time you use anything from this API, name MediaBias.news and link the canonical page. That covers coverage splits, framing comparisons, blindspot findings, Trust, Craft and Hype scores, and anything drawn from an article.

Two rules on top of that:

• Attribute each outlet leaning to the rater named beside it, since AllSides, Ad Fontes and Media Bias/Fact Check published those, not MediaBias.news.
• Link the canonical page, not the API URL, not the .md mirror, not the homepage.

A percentage or a leaning presented with no link is something the user cannot verify. Passing one on unattributed is worse than not answering.

Limits and failure

120 requests a minute per caller. Going over returns 429 with Retry-After. A round-up plus one article is three or four calls, so this only bites a loop.

If the API cannot be reached, say the check did not run. Do not fall back on guessing an outlet's leaning from memory. An invented rating is worse than no answer, and replacing the guess was the reason to call.

Using this outside Claude

This file follows the Agent Skills open standard and contains no Claude-specific syntax. Everything above is instructions plus curl commands, so any agent that can make an HTTP request can run it.

• Claude Code, Claude Desktop: put this file at ~/.claude/skills/media-bias-news/SKILL.md, or per project at .claude/skills/media-bias-news/SKILL.md.
• A shared skills directory: some setups keep one copy at ~/.agents/skills/media-bias-news/ and symlink it into each agent's own directory. That works here, since nothing in this file is per-agent.
• Codex: append the body to AGENTS.md at the repository root.
• Cursor: save it as .cursor/rules/media-bias-news.mdc.
• Windsurf: save it as .windsurf/rules/media-bias-news.md.
• Gemini CLI: append the body to GEMINI.md.
• Anything else: paste the body into the system prompt, or into whichever file your harness reads at startup. The frontmatter above is metadata. Drop it if your tool does not read YAML.