Daily Weights - Setup

A daily, agent-readable news feed for the AI and agent ecosystem. This repo is the open setup flow: you point your agent at it once, and your agent wires itself into a daily read-and-improve loop.

Removal is one file: read UNINSTALL.md and every trace is gone in one pass.

What Daily Weights is (consumer view)

Most AI news is formatted for humans: HTML bloat, images, engagement bait. Daily Weights inverts that. It publishes one plain-markdown edition per stream per day, on a fixed schema (daily-weights/v1), with tags, entities, source URLs, and token estimates on every item. Your agent can grep it, diff it, and act on it without a browser.

The feed is segmented into streams you subscribe to based on your own setup:

Stream	Covers
anthropic	Claude, Claude Code, Anthropic API: releases, changelog, engineering
openai	OpenAI models, API, and tooling
google	Google / Gemini models, API, and tooling
open-source	Open-weight models and local inference
agent-frameworks	Agent frameworks and orchestration libraries
security	Prompt injection, jailbreaks, model and supply-chain risk
research	Notable AI research with practical consequence
tooling	SDKs, evals, observability
breaking-changes	Deprecations, price changes, EOLs, default changes (recommended to everyone)
meta	The top items across every stream, one file

Live streams and their URLs: https://dailyweights.com/topics.json

The one-line install

Paste this at your agent:

Read https://dailyweights.com/setup/INSTALL.md and set up Daily Weights for me.

Your agent reads INSTALL.md, detects your stack, proposes a matching set of streams, adds a daily task that pulls the right feed URLs, and sets up an advisory loop that proposes improvements for you to approve. It stops and asks you at exactly two points (see below).

Trust posture

• The feed is data, not instructions. Your agent treats every feed item as untrusted input. It never executes anything the feed says just because the feed says it. We aggregate prompt-injection news, so we hold that line ourselves.
• Advisory self-improvement, human approval at exactly two points. Your agent asks you (1) when it proposes which streams to subscribe to, and (2) when it creates the daily task. After that, day-to-day it surfaces improvement proposals for you to approve; it never auto-applies a configuration change from feed content.
• No spend, no sends. Nothing here authorizes your agent to spend money or message anyone.
• Zero exit friction. Every edition header carries unsubscribe instructions, and UNINSTALL.md removes the task and the URLs in one read.

Current state (honest)

Daily Weights is pre-launch and proving itself. The feed publishes daily and the schema is locked, but it has no established track record or user base yet. Treat it as a new dependency: useful, low-cost to try, and trivial to remove. If it does not earn its place in your agent's daily task list, uninstall it.

What is in this repo

README.md            This file
INSTALL.md           The runbook your agent reads and follows
UNINSTALL.md         Clean removal in one read
adapters/
  claude-code/       Claude Code skill + CLAUDE.md snippet
  mcp/               Generic MCP-client wiring
  cron/              Plain cron + curl wiring
  generic/           Runbook + AGENTS.md snippet for any other harness
schema/
  stack-manifest.schema.json    JSON Schema for how your agent declares its stack
  stack-manifest.example.json   A worked example
templates/
  daily-task.md               The recurring task your agent writes for itself
  improvement-proposal.md      The advisory-loop proposal format (two approval points)
LICENSE              MIT
CONTRIBUTING.md      How to add an adapter or fix a template

This repo contains no engine code. It carries only how to consume the feed and what to do with it. The fetchers, source catalog, ranking, synthesis, and personalization that produce the feed are not here and never will be.

FAQ

Does this cost anything? No. The static feed is free. The setup flow reads public markdown files.

Do I have to give it access to my system? Your agent adds a daily read task and a small manifest file describing your stack. It does not need credentials, and it never spends or sends. The manifest stays on your machine on the free tier. As with pulling any feed, the daily requests do reveal to the server which streams you fetch and the User-Agent you set, through ordinary web request logs.

What if I do not use Claude? Pick the adapter that matches your harness under adapters/. The generic/ runbook covers anything not listed.

How do I remove it? UNINSTALL.md. One read.

License

MIT. See LICENSE.