Claude routines vs Desktop tasks vs /loop for recurring Sume renders
Pick between a Claude cloud routine, a Desktop task and /loop for scheduled Sume generation, or a Sume cron schedule when the work is only generation.

For a recurring Sume render, use a Claude cloud routine when it must run with your computer off, a Desktop task when it needs local files, and /loop only for polling during a session. If the work is only media generation on a cadence, a Sume cron schedule needs no Claude session at all. Anthropic's own comparison table is the source for the Claude side (Claude Code docs, read 2026-10-04).
The three Claude options, from Anthropic's table
The page compares cloud routines, Desktop scheduled tasks and the /loop command on where they run, whether the machine must be on, MCP access, permission prompts and minimum interval.
| Cloud routine | Desktop task | /loop | |
|---|---|---|---|
| Runs on | Cloud | Your machine | Your machine |
| Machine must be on | No | Yes | Yes |
| Needs an open session | No | No | Yes |
| Local files | No, fresh clone | Yes | Yes |
| MCP servers | Connectors configured per task | Config files and connectors | Inherits from the session |
| Permission prompts | None, runs autonomously | Configurable per task | Inherits from the session |
| Minimum interval | 1 hour | 1 minute | 1 minute |
What that means for paid Sume calls
A cloud routine runs autonomously with no permission prompts, so nothing stops a paid call except the gates you set. On the hosted MCP at https://mcp.sume.com/mcp, idempotency_key is required on write and paid tools, dry_run previews cost, and max_spend_usd caps a call only when you send it (MCP tools and gates). Put all three in the routine's prompt.
A Desktop task can use a permission mode instead, but a run in a mode that needs approval stalls until you answer. /loop is session-scoped, so it suits checking a job's status, not producing a day's renders.
When a Sume schedule is simpler
A Sume schedule stores instructions, a model, a 5-field cron expression with a timezone and a spend cap, and Sume runs it as an Agent in a fresh thread. You read the run receipt over the API, and an optional webhook delivers it. It is the plain choice when nothing else needs the run, but you author it in the dashboard and cannot create it over the API (Scheduled).
- Needs your repository or local files: Desktop task.
- Must run with your computer off and call tools from Claude: cloud routine, with gates in the prompt.
- Only generation on a cadence, bounded by a per-run cap: Sume Scheduled.
- Checking a job while you are in a session:
/loop.
One more difference to plan for
A cloud routine's minimum interval is one hour, and a Desktop task or a loop can run every minute. A cadence tighter than the routine floor points to Desktop or a Sume schedule. Whichever you choose, send a stable idempotency_key per intended output so a retry cannot render twice.
Sources
Related posts
More in Comparisons
- Code2Video: 168 briefs and how to score your own
HeyGen's Code2Video benchmark uses 168 briefs from real launch videos. Here is how to build a smaller scoring set and render it from structured input on Sume.
- Edits First Draft is iOS only: the Android and web route
Instagram's Edits First Draft is iOS only, per a secondary source. On Android or web, cut clips with Sume video trim at $0.02 per job and join them in Timeline.
- ElevenLabs TTS on fal at $0.10 per 1,000 characters vs direct and Sume
fal lists ElevenLabs TTS at $0.10 per 1,000 characters, ElevenLabs lists $0.08 for v3, Sume is $0.0475. A 2,000-character script: $0.20, $0.16, $0.095.
- fal image model list vs the Sume catalog: how to check by query
fal.ai lists Seedream 5.0, GPT Image 2.5, Flux 2, Nano Banana 2, Ideogram 4 and Krea 2. How to see which ones a Sume key can call, with a short Python script.
Written by Sume