ClaudeCodeCron

Run Claude Code on a schedule — without burning your context window, and without leaving your subscription bucket.

Heads up — June 15, 2026. Anthropic moves claude -p, the Agent SDK, and Claude Code GitHub Actions off subscription limits onto a separate monthly credit pool ($20 Pro / $100 Max 5x / $200 Max 20x) at API list prices. Interactive Claude Code stays on the subscription — and so does claude-code-cron, since it drives the interactive TUI. (Anthropic announcement)

claude-cron drives Claude in three modes:

• restart — fire a one-shot prompt, kill the process, sleep, repeat. No context carries between runs.
• clear — spawn one long-lived Claude, send a prompt, and every N minutes type /clear and re-send the prompt to wipe accumulated context without losing the loop.
• compact — same shape as clear, but sends /compact instead of /clear so Claude summarises the context rather than wiping it. Use when you DO want runs to remember each other.

Why

If you've tried scheduling a recurring task inside Claude Code itself — e.g. /loop 1m "check the deploy" — you've probably hit two problems:

• Token cost climbs every tick. Each iteration keeps appending to the same session, so the context (and the per-tick cost) grows without bound. After an hour of /loop 1m, every run is paying for a huge cached prefix it doesn't actually need.
• It eventually breaks. Once the context gets large enough that auto-compaction can't keep up — or fails outright — the loop stalls, derails, or stops responding.

claude-cron sidesteps both by managing the loop outside the conversation:

• restart mode spawns a brand-new Claude session every tick. Context starts empty each time, so per-run token cost is flat no matter how long the loop has been running.
• clear mode keeps /loop itself but resets the session's context every N minutes via /clear, so the inner loop never gets the chance to blow up.
• compact mode sits between the two — keeps the loop running and the session alive, but every N minutes sends /compact so Claude summarises rather than discards what came before. Pick this when later runs genuinely need to remember earlier ones.

The outer scheduler is a tiny Node process that won't degrade under context pressure — it just keeps ticking. Stick with raw /loop only when you specifically need the model to track every single iteration verbatim.

Install

npm install -g claude-code-cron

Requires Node 20+ and claude on your PATH.

Usage

claude-cron restart <interval> "<prompt>" [idle-seconds]
claude-cron clear   <interval> "<prompt>"
claude-cron compact <interval> "<prompt>"

Arg	Meaning
interval	30s, 10m, 2h, or bare seconds.
prompt	The prompt to type into Claude. Quote it.
idle-seconds	(restart only, default 5) How long the output must be silent before Claude is considered done.

Stop the loop with Ctrl+C.

restart mode

Every interval, ClaudeCodeCron spawns a fresh Claude, sends the prompt, waits for the response to finish, kills the process, and sleeps. Use this when each run is a discrete task and conversation context should NOT carry across runs.

# Every 30 minutes, summarize the market in a fresh session
claude-cron restart 30m "summarize today's market news"

# Every 2 hours, longer idle threshold for slow tool calls
claude-cron restart 2h "check for new SEC filings on AAPL and summarize" 15

# Quick smoke test
claude-cron restart 30s "say hi in three words"

clear mode

ClaudeCodeCron spawns one Claude session, sends the prompt once, and every interval types /clear followed by the prompt again. The Claude process stays alive between cycles; only the conversation history is dropped.

Best paired with /loop, where the prompt itself is the long-running task and you only want to prevent its context from growing unbounded across hours of runtime.

# Run /improve-code once a minute forever, with a context reset every 30 minutes
claude-cron clear 30m "/loop 1m /improve-code"

# Same idea, hourly reset
claude-cron clear 1h "/loop 5m /add-test"

compact mode

Identical to clear mode in structure — one long-lived Claude session, prompt sent on startup, a slash command fired every interval — except it sends /compact instead of /clear. Claude summarises the conversation history into a shorter form rather than throwing it away, so later runs still see the gist of earlier ones.

Use this when the inner /loop builds up useful state across iterations (a running list, a tracked target, an evolving plan) and you only want to keep the token count manageable — not reset the loop's memory.

# Add tests in a /loop forever, summarise context every hour
claude-cron compact 1h "/loop 5m /add-test"

# Watch a deploy queue, summarise every two hours to keep history but stop bloat
claude-cron compact 2h "/loop 5m check the deploy queue and tell me what changed"

License

MIT