/ai-release-hunter

name: ai-release-hunter
description: Daily digest of AI releases from Anthropic, Meta, Vercel Labs, OpenAI and Google GitHub repos, the Claude Cowork changelog, the claude.dev blog and Hacker News, deduplicated against a local state file and sent by email. Use when the user asks for a daily AI news digest, to scan AI labs for new releases, or to set up a scheduled release-hunter job.
license: MIT
allowed-tools: Bash Read Write WebFetch
metadata:
  author: Daniel Avila
  version: "1.0.0"
  tags: "ai-news, releases, github, hacker-news, digest, email, anthropic, openai, google, meta, vercel"

AI Release Hunter (Anthropic + Meta + Vercel + OpenAI + Google + Hacker News)

Scans the public AI repos of Anthropic, Meta, Vercel Labs, OpenAI and Google on GitHub, plus the Claude Cowork changelog, the claude.dev blog and Hacker News, and emails a digest of the day's relevant movements. Designed to run once a day as a scheduled job (cron, a Claude Code routine, or `/loop`).

Configuration

Set these before the first run. Every path is relative to the working directory unless you choose otherwise.

| Setting | Default | Purpose | |---------|---------|---------| | `RECIPIENT_EMAIL` | *(required)* | Address that receives the digest | | `SEND_COMMAND` | *(required)* | Command that sends an email given a recipient, subject and body file (e.g. a Gmail CLI, `mail` or `sendmail`; to send through an MCP tool instead, add it to `allowed-tools`) | | `STATE_DIR` | `.release-hunter/state/` | Deduplication state (`seen.json`) | | `SNAPSHOT_DIR` | `.release-hunter/snapshots/` | Page snapshots (Vercel Labs, Cowork changelog) | | `OUTPUT_DIR` | `.release-hunter/output/` | Email bodies and run logs | | `EMAIL_LANGUAGE` | Spanish | Language of the email | | `EMAIL_SUBJECT` | `IA: novedades del día (Anthropic + Meta + Vercel + OpenAI + Google + Hacker News)` | Subject line | | `AUDIENCE` | the user's audience | Who the "why it matters" lines are written for (e.g. a YouTube channel, a team) |

Suggested schedule: daily at ~08:00 in the user's time zone.

Add `.release-hunter/` to `.gitignore` so state, snapshots and email bodies are never committed. The scan makes well over 60 GitHub API calls per run, so a `GITHUB_TOKEN` is strongly recommended (see below).

Watched sources

Anthropic (GitHub org `anthropics`)

• anthropics/claude-code
• anthropics/claude-cookbooks
• anthropics/anthropic-sdk-python
• anthropics/anthropic-sdk-typescript
• anthropics/courses

Also list the org's repos (`https://api.github.com/orgs/anthropics/repos?per_page=100`) and consider any other with recent relevant activity (agent SDKs, examples, tooling).

claude.dev blog (https://claude.dev/)

Anthropic's technical blog with tips, techniques and POVs from its developers for building with Claude. Fetch it with WebFetch and take the list of recent posts (date + title + link). Compare against the already-reported ones in `STATE_DIR/seen.json` under the key `claude_dev_posts` (list of slugs or URLs; create it if missing). Report each new post as a highlighted item: title, date, one line on why it matters to `AUDIENCE` + link to the post. Add the new ones at the end.

Meta AI (GitHub orgs `meta-llama` and `facebookresearch`)

• meta-llama/llama-models
• meta-llama/llama-stack
• meta-llama/llama-agentic-system
• meta-llama/llama-toolchain
• meta-llama/llama-cookbook
• meta-llama/purplellama
• meta-llama/codellama
• facebookresearch/segment-anything
• facebookresearch/segment-anything-2
• facebookresearch/dinov2
• facebookresearch/dinov3
• facebookresearch/faiss
• facebookresearch/audiocraft
• facebookresearch/spiritlm
• facebookresearch/vggt
• facebookresearch/vjepa2
• facebookresearch/xformers

If any repo returns 404, skip it and continue. Also list both orgs' repos and add up to ~15 more with a push in the last 30 days whose name or description fits AI: new models, agents, MCP, vision, audio, multimodal, diffusion, segmentation, embeddings, fine-tuning, RL. Ignore internal infra repos, consumer product demos or archived ones.

Vercel Labs (org `vercel-labs`; experiments at https://vercel.com/labs)

• vercel-labs/agent-browser (browser automation CLI for AI agents)
• vercel-labs/deepsec (security harness for coding agents)
• vercel-labs/json-render (Generative UI framework)
• vercel-labs/skills (open agent skills ecosystem CLI)
• vercel-labs/portless
• vercel-labs/fx (unix-like coding agent)
• vercel-labs/vgpu
• vercel-labs/scriptc, vercel-labs/wterm, vercel-labs/emulate, vercel-labs/native, vercel-labs/callscript, vercel-labs/run, vercel-labs/visual-json, vercel-labs/zerolang, vercel-labs/phase, vercel-labs/webreel, vercel-labs/opensrc

If any repo returns 404, skip it and continue. Also list the org's repos and add those with a push in the last 30 days not already in the list.

Labs page updates: fetch https://vercel.com/labs and compare the experiment list (name + category: Labs products / Active experiments / Past experiments) against the snapshot at `SNAPSHOT_DIR/vercel-labs-experiments.json`. Report any new experiment or category change as a highlighted item. Update the snapshot at the end.

OpenAI (GitHub org `openai`)

• openai/codex (coding agent CLI)
• openai/codex-security (Codex Security CLI + TypeScript SDK)
• openai/codex-action (GitHub Action for Codex)
• openai/openai-agents-python (agent framework, Python)
• openai/openai-agents-js (agent framework, JS)
• openai/openai-guardrails-python
• openai/openai-guardrails-js
• openai/tunnel-client (Secure MCP Tunnel client)
• openai/openai-cua-sample-app (computer use via API)
• openai/symphony (autonomous implementation runs)
• openai/openai-cli (official CLI)
• openai/plugins, openai/community-plugins
• openai/fence

If any repo returns 404, skip it and continue. Also list the org's repos and add up to ~10 more with a push in the last 30 days whose name or description fits AI: agents, codex, MCP, skills, computer use, guardrails, plugins, sandboxing. Ignore