Sidewise README | AllMCPs

Sidewise

What is the other side reading?

Sidewise pulls 16 feeds from 14 newsrooms across the spectrum, tags each left, center or right, and flags the blindspots: stories only one side is covering. Free. No account, no key, no limit.

Two feeds from one newsroom count as one voice. A publisher can't run an opinion section next to its main feed and call that two sources.

Five ways in: a web reader, native iPhone, iPad and Mac apps, a JSON API, an MCP server, and a terminal client (swift build && ./.build/debug/sidewise-tui 5 — see tui/).

Screenshots

Sidewise

Apps

SwiftUI, one codebase for iOS and macOS, in ios/. It reads the same public API. No account, no tracking. Saved stories and the feed cache stay on the device.

Code

cd ios && xcodegen generate
xcodebuild -scheme Sidewise-iOS -destination 'generic/platform=iOS Simulator' build

MCP

Terminal

claude mcp add --transport http sidewise https://sidewise.heyitsmejosh.com/mcp

Tool	Params	Returns
`get_news`	`view`, `outlet`, `bias`, `developing`, `q`, `limit`	Current headlines, flat or clustered by story
`get_blindspots`	`limit`	Only stories covered by a single political side
`compare_coverage`	`q` (required), `limit`	One story as each side headlines it, plus the words unique to each
`get_feed_health`	:	Which feeds answered, and whether the data served is complete or a stale fallback

Call get_feed_health before you trust an empty or one-sided result. An outage and a quiet news day look the same otherwise.

Stateless streamable HTTP, no auth. Works with any MCP client: Claude Desktop, Claude Code, Cursor.

API

GET https://sidewise.heyitsmejosh.com/api/stories. Every parameter is optional.

Param	Values	Default
`view`	`latest`, `stories`, `both`	`both`
`outlet`	any outlet name, e.g. `Hacker News`	all
`bias`	`left`, `center`, `right`	all
`blindspot`	`true`	off
`developing`	`true`: three or more newsrooms in the last 90 minutes	off
`compare`	`true`: attach the side-by-side breakdown to each story	off
`q`	substring match on headline and summary text	none
`limit`	1–200	60 clusters / 120 headlines

Two more endpoints. GET /api/health reports every feed and answers 503 when more than half are down, so you can point a monitor at it as is. GET /api/sources lists each feed with its bias, its side and its parent newsroom.

Terminal

curl 'https://sidewise.heyitsmejosh.com/api/stories?view=stories&blindspot=true'
curl 'https://sidewise.heyitsmejosh.com/api/stories?view=latest&outlet=Hacker%20News&limit=10'

config.json

{
  "updated": 1754700000000,
  "stories": [
    {
      "title": "Fed holds rates steady",
      "blindspot": false,
      "sources": [
        { "title": "Fed holds rates steady", "link": "https://…", "outlet": "NPR", "bias": -1 },
        { "title": "Fed refuses to cut rates", "link": "https://…", "outlet": "Fox News", "bias": 2 }
      ]
    }
  ],
  "latest": [
    { "title": "Fed holds rates steady", "link": "https://…", "outlet": "NPR", "bias": -1, "ts": 1754699000000 }
  ]
}

CORS is open. Feeds are re-pulled at most every 2 minutes. Responses are no-store (see below). ts is epoch ms, or 0 when the feed gave no date. Those sort to the bottom. Full spec: openapi.yaml · orientation for agents: llms.txt.

How it works

One Cloudflare Worker (worker.js) polls every feed in src/feeds.js and serves the page, the API and MCP from one pooled pull:

latest: every headline, newest first.
stories: headlines clustered by title-keyword overlap. Each source tagged left, center or right. One-sided clusters flagged blindspot.

Only the feed pull is cached, under one constant key. Filtering happens per request in shape(), after it, and responses go out no-store. The zone's CDN cache ignores query strings, so caching them would hand one caller's ?outlet= to everyone. Nothing is refetched either way. Only the cheap filtering runs again.

Clusters are filtered after clustering. The flat feed is filtered before its 120-item cap. Both orders matter. Narrow the input first and you lose the cross-outlet comparison. Filter after the cap and small outlets vanish.

architecture

Bias scores

Each source has a score from -2 (left) to +2 (right). 0 is center, or non-political for tech outlets like Hacker News and Daring Fireball. I assigned these by hand in FEEDS. They are a rough lean, not a rating.

CBC · The Guardian · NPR · BBC · Global News · National Post · Fox News · NY Post · Daily Wire · Hacker News · Daring Fireball · NBC News · Wall Street Journal · NY Post Opinion · Vancouver Sun · The Province

Add one by appending [outlet, bias, url] to FEEDS at the top of worker.js. Any RSS 2.0 or Atom feed works.

Develop

Terminal

npm test         # 95 checks, no network
npm run feeds    # check every feed for freshness, not just a 200
npm run deploy   # wrangler deploy

Run npm run feeds after touching FEEDS. It checks recency, the only way to catch a zombie: a feed that still answers 200 with clean XML whose newest item is two years old. CNN did that for three years. An item count never noticed.

Cloudflare Workers plus Static Assets. One deploy serves the page, /api/stories and /mcp.

Tests

test/ covers the four modules in src/ with no network and no Worker runtime:

File	What it holds down
`parse.test.mjs`	Entity decoding, double-escaped summaries, CDATA and Atom shapes, and the rejection of `javascript:` and `data:` links
`stories.test.mjs`	Clustering, the newsroom-not-feed counting rule behind blindspots, the developing window, and filter-after-cluster ordering
`load.test.mjs`	Feed failure reporting, timeouts, the degraded/stale fallback, and the rule that a bad pull never overwrites the last good one
`worker.test.mjs`	Query parsing and clamping, and that everything under `/api/` answers JSON with CORS, including 404s, 405s and 500s
`mcp.test.mjs`	JSON-RPC framing, malformed input, and each tool's contract

Security: see SECURITY.md.

License

MIT 2026, Joshua Trommel. Headlines and links belong to their publishers. Sidewise stores nothing and links out to the original.

Whitepaper

Technical whitepaper

API and agent tools

REST (/api/*), the POST /mcp server, and the reader's in-page WebMCP tools (public/webmcp.js) expose the same four operations. They are tested against each other so they can't drift. See docs/API.md.