Wiring agentfeeds.si into Claude Code, Cursor, and Open-Weight Local Agents in One Command

Step-by-step setup guide for connecting Claude Code, OpenAI Codex CLI, Cursor, and local Ollama harnesses to agentfeeds.si with automatic fallback recovery.

Instead of manually downloading benchmark tables or security advisories before starting a coding session, you can bind agentfeeds.si directly into Claude Code, Codex CLI, Cursor, or a local Ollama harness in under 30 seconds. Once configured, the harness fetches only the < 4 KB daily delta digest (https://agentfeeds.si/feed.json) at session start and caches the verified Merkle root locally.

Prerequisites

• A POSIX shell (bash or zsh) with curl and python3 (>= 3.11) installed.
• Outbound HTTPS access to https://agentfeeds.si/feed.json.
• Optional: Claude Code (claude), Codex CLI (codex), or an Ollama Cloud / local endpoint.

Step 1: Verify connectivity and inspect the live feed digest

Run the following command in your project root to fetch the current feed manifest and confirm that status equals "ok":

curl -fsSL https://agentfeeds.si/feed.json | python3 -m json.tool | head -n 25

What success looks like: You see a JSON object containing "feed_version": "agentfeeds.top5.v1", "status": "ok", and a "merkle_root_sha256" 64-character hex string.

How to recover from failure: If curl returns a timeout or DNS error, verify outbound port 443 connectivity (curl -I https://baltor.ai) or check your local proxy configuration.

Step 2: Register the read-only feed in Claude Code or Codex

Add the read-only feed endpoint to your workspace configuration or startup script so your harness reads the current deprecation and CVE delta table before generating code:

{
  "mcpServers": {
    "agentfeeds-readonly": {
      "type": "http",
      "url": "https://agentfeeds.si/feed.json"
    }
  }
}

For local Ollama or custom Python harnesses, pull the compact Markdown collection directly via https://agentfeeds.si/feeds/collections/coding-benchmarks.md and cache it by ETag so unchanged feeds consume zero network bandwidth on subsequent turns.