Cost-aware automation defaults¶
Ground truth (verified 2026-04-30 by reading the actual code):
Justin does NOT have a separate Anthropic API account. He runs Claude Code Max. Every Claude-powered surface he owns bills against his Claude Code subscription quota, including:
- His Telegram bots (they shell out to
claude -p, not the anthropic SDK; verified via grep offorge_telegram_*brain*.py) /spawnlocal sessions on Console/followupheadlessclaude -pfires (systemd timer triggered)- This interactive Claude Code conversation
- Any worker spawned via
claude -ppipe mode
There is no "Anthropic API charge" surface in his stack. Stop framing automation choices as "subscription vs API" — that axis doesn't exist for him.
The actual cost axis: procedural vs LLM-driven¶
| Surface | Cost |
|---|---|
| Procedural Python via cron / systemd timer / at-job (no LLM) | $0 |
Anything that invokes Claude (bots, /spawn, /followup, interactive) |
Subscription quota |
Decision tree¶
- Procedural Python (cron / systemd timer / n8n) — $0. Use when the work is deterministic (file presence, log scanning, count comparisons, schema validation, HTTP-glue). 90% of "is X healthy" / "ingest from Y" / "rollup Z" checks fit here.
/spawnlocal Claude Code session — subscription quota. Use when the work needs reasoning right now AND local state access (logs, secrets,/notify, Console files)./followup(local systemd +claude -p) — subscription quota. Use when the work needs reasoning later or on a cadence. Forge does not use/schedule(cloud sandbox); everything runs on Console.
Why prefer procedural¶
Subscription quota is finite. Burning it on deterministic checks that a 30-line Python script can do is wasteful. Reserve LLM-driven work for tasks that genuinely need reasoning, classification, drafting, or synthesis.
How to apply¶
Before reaching for any LLM-driven scheduler, ask: "Is this deterministic, or does it need reasoning?" If deterministic, write a Python script and use cron / systemd timer / at-job. If it needs reasoning, pick /spawn (now) or /followup (later/cadence). /schedule is not in the toolkit, Forge runs locally.
[Claude Code, 2026-04-30; revised 2026-05-02 to drop /schedule entirely]