---

What is this?

"What changed?" is a question agents and developers ask constantly — before an upgrade, after an incident, when a tool suddenly behaves differently. The answer is scattered across GitHub releases, RSS feeds, changelog pages, and blog posts in a hundred different formats.

Releases collects those into one registry. It watches hundreds of sources across the vendors developers actually depend on, and for every release stores the original content plus an AI-generated title, summary, and breaking-change classification. Each organization gets a maintained overview of what it shipped recently. The web page a human reads and the JSON an agent fetches are the same content.

This repo is the source of the canonical deployment at releases.sh: the API worker (the authoritative data plane), the MCP server, the web frontend, and the ingest pipeline + agent harness that keep the registry fresh. The user-facing CLI ships separately from buildinternet/releases-cli (npm + Homebrew).

Use it

No account or API key needed for reads — all four surfaces are public.

MCP — hosted at mcp.releases.sh, listed in the MCP Registry as sh.releases/mcp:

claude mcp add --transport http releases https://mcp.releases.sh/mcp   # Claude Code
codex mcp add releases --url https://mcp.releases.sh/mcp               # Codex
npx -y mcp-remote https://mcp.releases.sh/mcp                          # stdio bridge (VS Code, Zed, …)

CLI — one-off via npx, or brew install buildinternet/tap/releases:

npx @buildinternet/releases get anthropic     # what did Anthropic ship lately?
npx @buildinternet/releases search "MCP"      # search across every vendor

REST API — public GET endpoints on api.releases.sh, spec at /v1/openapi.json:

curl "https://api.releases.sh/v1/releases/latest?limit=5"

Agent skills — auto-triggering playbooks, installable into any agent (Claude Code / Codex / Cursor / OpenCode) without checking out anything. Start with the reader skills — search, MCP lookups, and release analysis — which is what almost everyone wants:

npx skills add buildinternet/releases-cli   # reader skills — search, MCP, release analysis

# just the skill that writes a releases.json manifest for your own product
npx skills add buildinternet/releases --skill creating-releases-json

On Claude Code, the CLI repo also installs as a plugin — the reader skills plus a bundled MCP connection and a /releases command:

/plugin marketplace add buildinternet/releases-cli
/plugin install releases@releases

_{Running or maintaining the registry itself? This repo also ships operator skills (source onboarding, parsing, bulk maintenance) — most need an admin key: npx skills add buildinternet/releases. See releases.sh/docs/skills.}

Docs — user-facing documentation is served from the web app at releases.sh/docs (source in web/src/content/docs/): installation · skills · REST API · MCP · webhooks · listing your product.

Agents consuming the product start from releases.sh/llms.txt; the MCP tool catalog and auth model are also covered in docs/architecture/mcp.md. Creating a free account unlocks higher rate limits (mint an API key at releases.sh/account), plus follows, personalized feeds, webhooks, and email digests.

What's in this repo

Path	What
workers/api/	Hono API on Cloudflare D1 — the authoritative data plane
workers/mcp/	Remote MCP server at mcp.releases.sh
workers/discovery/	Durable-Object agent-session orchestrator
workers/webhooks/	Signs + delivers release.created events (HMAC-SHA256, retry/DLQ) — docs
web/	Next.js frontend, deploys on Vercel
packages/	Shared code — core + api-types publish to npm; the rest are private workspaces
src/agent/	Managed-agents discovery + worker harness (prompt builder + shared types)
.claude/	Claude Code config — skills/ (canonical skill home), agents/, commands/, workflows/

How it fits together:

• Storage — Cloudflare D1 (FTS5 + Vectorize). The API worker is the sole data plane.
• Ingest — adapters for GitHub Releases, RSS/Atom/JSON feeds, and a browser-rendering fallback for feed-less pages (packages/adapters/). The crawler signs outbound fetches (RFC 9421) as a Cloudflare Verified Bot.
• AI — changelog parsing, summarization, grouping, and org overviews run in the API worker as direct Anthropic SDK calls.
• Agents — discovery + worker run as Anthropic-hosted managed agents; definitions auto-deploy on merge when their source changes.

Contributor mental model

This repo has four kinds of code:

Area	Paths	What to know
Product surfaces	workers/api/, workers/mcp/, workers/webhooks/, web/	The API worker is the data plane. Web, MCP, CLI, and webhooks read from or route through it.
Shared packages	packages/core/, packages/api-types/, private packages/*	core owns schema and pure helpers; api-types owns wire shapes. Change these first when shared contracts move.
Hosted operations	workers/discovery/, managed-agents/, .claude/skills/, .claude/workflows/	These power the canonical releases.sh ingest and operator loop. Some paths need hosted credentials, but local work does not.
Historical context	docs/architecture/, docs/plans/, docs/superpowers/	Architecture docs are maintained references. Plans and specs are point-in-time design history.

For normal contributions, start with bun run bootstrap, bun run check, and bun test. You do not need production Cloudflare, Anthropic, Vercel, email, or Firecrawl access to work on the core loop. If you are trying to fork or self-host the full service, read deploy-coupling.md.

Per-package detail and project conventions live in AGENTS.md — the agent entry point for working in this repo. Architecture deep-dives are in docs/architecture/, with a reader's guide at docs/README.md.

Local development

Prerequisites: Bun. No external accounts needed for the core loop:

bun run bootstrap        # one-command setup: tooling, deps, env files, local D1
bun run doctor           # diagnose the setup — reports what's missing and how to fix it

bun run check            # lint + type-check + format (the CI gate)
bun test                 # full test suite, secret-free
bun run dev:api          # API worker on local D1
bun run dev:web          # Next.js frontend

bootstrap is idempotent (safe to re-run; never overwrites your env files or wipes the local DB) and doctor is read-only. Prefer the manual steps? They're bun install, copy each *.example env file, then bun run db:reset:local.

AI passes, scrape fetches, and semantic search take your own keys (ANTHROPIC_API_KEY, Cloudflare Browser Rendering, VOYAGE_API_KEY) and degrade gracefully without them. Setup detail, environment variables, testing, deployment, and the full no-accounts / bring-your-own-keys / hosted-only breakdown live in CONTRIBUTING.md.

License

Apache-2.0. The published npm packages (releases-core, api-types) are deliberately MIT for maximum reuse.