Table of Contents
About The Project
Every host-read promo a podcast network runs for its new show is an ad slot it didn't sell, yet launch plans are drawn up by topic: the new true-crime show goes on the other true-crime shows. Crossfeed plans by who listens instead. It looks the launch and the network's shows up in Qloo, ranks the roster by how much each show's listeners overlap the launch's, asks a person at each fork, and drafts an insertion order that code prices from the network's own downloads and CPM. A member approves it; the agent never can.
The same engine works the booking inbox. Crossfeed reads the pitches publicists send, looks up the public work each guest is promoting, a book, film, TV show, or podcast, ranks those works against a show's listeners, and drafts a reply to every pitch for the booker to approve, edit, or discard.
Beside every Qloo ranking sits the plan an agent without Qloo would make: Claude's ranking from show names and descriptions, or from a pitch's own words, with no Qloo data. So what Qloo adds is on the screen, not asserted.
It's a product a network's members sign in to, with passkeys, shared networks, a REST API, a TypeScript SDK, an MCP server, and a Claude Code plugin. Anyone can try it as a guest, in one click, on a sample network built from The Ringer's public show names, with labelled sample figures. Crossfeed isn't affiliated with The Ringer.
Built for the Qloo Agentic Hackathon: the entry is on Devpost, and docs/SUBMISSION.md is its full text.
Screenshots
The live app on October 30, 2026, signed in as a guest, on the guest network's sample data.
How It Works
-
Start a run. Open a launch, such as the sample network's The Watch, and select Start run. The agent looks the launch up in Qloo, ranks the roster against its listeners in one call, and asks you for a slot budget, then which shows to leave out.
-
See who shares its listeners. The fit chart ranks the roster by listener overlap, with each show's rank in a topic-only plan beside it. A show that climbs three or more places is marked ▲: select it for why it moved, and select a call number, such as C-53, for the Qloo request and raw scores behind it.
-
Approve the plan. The agent drafts slots and a host-read angle for each finalist from the tags its listeners share with the launch's in Qloo. Code prices each slot, holds the plan to the budget, and names any close pairs among its shows. Only a member approves it, and only then exports the CSV.
-
Hand it on. An approved order has a launch report to print or save as PDF, with an index of every Qloo call it cites.
-
Run it again. Answer differently, and the launch's Changes tab says what changed since the earlier run: the inputs, the finalists, each show's move, and the plan's slots and value.
-
Work the inbox. On the inbox's Rank tab, select Rank pitches. Each type of pitched work is ranked against the show by who listens, beside its rank by the pitch's own words, and every pitch gets a drafted reply under Book, Later, or Decline.
Features
- Ranked by who listens. One Qloo insights call ranks the network's closed list, its own roster, against a launch, or against up to three comparable shows for a launch that hasn't aired.
- The counterfactual on screen. Claude's topic-only ranking sits beside Qloo's in every run, and the chart's lens moves the rows between the two.
- An agent that asks. It asks a person for the budget, the shows to leave out, and any ambiguous Qloo match, and it has no tool to approve or export.
- Priced by code. Each slot is worth downloads per episode × CPM ÷ 1,000, from the network's own figures, and code refuses a plan over the budget.
- Evidence for every figure. Each ranking, tag, and flag cites its Qloo call, which opens the request, its status, latency, and cache hit, and the raw scores.
- Analysis where it belongs. The roster's overlap and close pairs, its listeners by age band and gender, what changed between runs, the promo inventory, and a printable launch report.
- The booking inbox. Pitches as
.eml, .mbox, or text. Only each work's title and type reach Qloo, never a name, an address, or the pitch's text.
- A product to sign in to. Passkeys with no password or email, networks shared with invited teammates who see each other's changes live, roster CSV import, each network's own Qloo key, and daily caps.
- For scripts and assistants. A REST API with an OpenAPI document, the
@crossfeed/sdk TypeScript client, an MCP server at /mcp, two Agent Skills, and a Claude Code plugin. None of them can approve anything.
- Honest limits. Past the day's limits, or while Qloo fails, the sample launch replays its stored run, labelled "Replayed", and every other run stops and says why.
Architecture
One Worker serves every surface. Each network and each launch is its own Agents SDK Durable Object, which runs Claude's tool loop, and every Qloo call goes through one gateway Durable Object that caches, paces, retries, and caps them. A D1 database holds only what has to be looked up across them: members, networks, sessions, invites, and API keys. docs/TRD.md has the whole design.
Made with Archify from architecture.json, in Crossfeed's colours by export-architecture.mjs.
Tech Stack
- Languages: TypeScript for the Worker, the client, the SDK, and the tests.
- Frontend: React 19 islands in server-rendered Hono JSX shells, Tailwind CSS 4, and shadcn/ui on Base UI, with Lucide icons and Schibsted Grotesk, built by Vite.
- Backend: Cloudflare Workers, the Cloudflare Agents SDK on Durable Objects, D1, the Workers rate limiting binding, and cron triggers.
- Taste data: the Qloo REST API: search, insights, and compare.
- AI: Anthropic's Claude Haiku 5.5 (
claude-haiku-5-5), through the Anthropic TypeScript SDK.
- Accounts: passkeys with SimpleWebAuthn.
- For programs: an OpenAPI document from Zod schemas, the Model Context Protocol TypeScript SDK's server, and SDK types from openapi-typescript.
Getting Started
Crossfeed runs on your computer under Vite's dev server, which runs the Worker, its Durable Objects, and a local D1 database as Cloudflare would, with no Cloudflare account. It calls the real Qloo and Anthropic APIs, so it needs a key for each. The tests need neither.
Prerequisites
Installation
-
Clone and install. The app is in apps/.
git clone https://github.com/m1kuorg/qloo.git
cd qloo/apps
bun install
-
Add your keys. Write them to .dev.vars, which is git-ignored, with two random values: VISITOR_SALT salts the visitor hash behind the daily limits, and QLOO_KEY_SECRET encrypts each network's own Qloo key.
printf 'QLOO_API_KEY=%s\nANTHROPIC_API_KEY=%s\nVISITOR_SALT=%s\nQLOO_KEY_SECRET=%s\nGUEST_PITCHES=on\nRP_ID=localhost\nAPP_ORIGINS=http://localhost:5128\n' \
"$QLOO_API_KEY" "$ANTHROPIC_API_KEY" "$(openssl rand -hex 16)" "$(openssl rand -base64 32)" > .dev.vars
-
Create the local database. This applies the D1 migrations to a local copy under .wrangler/state.
bunx wrangler d1 migrations apply DB --local
-
Start the app. Open http://localhost:5128, then Open the sample network to sign in as a guest, or Create your network to sign up with a passkey. The local guest network holds the sample roster, inbox, and The Watch; the four sample launches are Qloo data, kept only in each deployment's database.
bun run dev -- --port 5128 --strictPort
-
Run the checks. The tests run the Worker in the Workers runtime, with Qloo and Anthropic faked at the network boundary, so they need no keys.
bun run typecheck
bun run test
-
Optionally, deploy to Cloudflare. Replace account_id and the D1 database_id in wrangler.jsonc with your own, set the four secrets with bunx wrangler secret put, and run bun run deploy. The apps README has every step, the preview environment, and the live checks.
Roadmap
See open issues for a full list of proposed features (and known issues).
Team
Made with contrib.rocks.
License
Apache 2.0. See LICENSE for more information.
Third-party files keep their own licenses:
- Matt Pocock's skills, MIT, in
.agents/skills/ and linked from .claude/skills/, apart from graphify and impeccable.
- The graphify skill, Apache 2.0, in
.agents/skills/graphify/.
- The impeccable skill and its agents, Apache 2.0, in
.agents/skills/impeccable/, .claude/skills/impeccable/, and .claude/agents/.
- shadcn/ui's components, MIT, the starting point of
apps/src/components/ui/.
- Lucide's icons, ISC, whose paths
apps/src/icons.tsx draws.
- Schibsted Grotesk, SIL Open Font License 1.1, which the app serves from
@fontsource-variable/schibsted-grotesk.
- docs/references/markdown-style.md, adapted from Google's documentation guide, CC BY 3.0.
Acknowledgments
