Hey HN, I built Claude Extender because I wanted a way to run multiple instances of Claude Code (e.g. agents). The core idea: agents are markdown files. The YAML frontmatter defines schedule, tools, cost limits, and notifications. The markdown body is the instruction Claude follows. A daemon runs in the background, spawning claude -p when agents are due.

There are three execution modes:

* Scheduled, cron-based. For example, my daily email triage runs at 6 AM, processes ~50 emails, archives the noise, and sends me a Telegram summary before I wake up.

* Watcher, event-driven. A cheap check script (JS/Python) polls for changes. Only triggers Claude when something actually happened.

* Persistent, long-running sessions with heartbeats, checkpoints, and restart policies.

Tools use MCP (Model Context Protocol) servers that you write and control. No plugin marketplace, no third-party code touching your credentials. Cost tracking is first-class. Every run logs its cost in USD. Per-run limits, daily budgets, monthly caps. The daemon pauses everything if you hit your budget. This matters when agents run unsupervised.

Everything is plain text, agent definitions, run logs, memory, cost ledgers. git diff your agents, grep your run history. I'm currently running this on a $5/mo Hetzner VPS.

Requirements: Node.js 20+ and the Claude Code CLI.

Anyhow, hope it's helpful and happy to answer any questions.