Scheduled jobs
Cron-based repetition — answers land in the chat automatically.
Some requests are not spontaneous questions but recurring tasks:
- "Write a market observation every Monday at 8 a.m."
- "Produce a KPI briefing from the Stripe figures daily at 5 p.m."
- "Summarise all new GitHub issues labelled
urgentevery hour."
That is exactly what scheduled jobs in SovrGPT are for.
Creating a job
Sidebar → Scheduled jobs → "New".- Give it a name.
- Enter a cron expression (
0 8 * * MON= Monday 8 a.m.). - Choose a model (e.g.
qwen3.8-27bfor longer reports). - Write the prompt — with tools (
brave_search,mcp_*) just like in a normal chat. - Choose the target chat (a new chat per run, or append to an existing one).
Cron syntax
Standard 5-field cron, UTC time. Examples:
| Expression | Meaning |
|---|---|
0 8 * * * | Daily at 08:00 UTC |
*/15 * * * * | Every 15 minutes |
0 9 * * MON-FRI | Weekdays at 09:00 UTC |
0 0 1 * * | Monthly on the 1st at midnight UTC |
A helper in the UI renders the next 5 trigger times in local time.
Where does the answer end up?
- New chat per run: every trigger creates a chat titled
<job name> · 2026-05-12 08:00. Visible in the org's chat list. - Existing chat: the answer is appended as a new assistant message.
- Optionally by e-mail: with the
emailtool in the prompt, the answer can additionally be sent to an address (dispatched from an EU region).
Guarantees & limits
- At most once: our scheduler triggers each schedule exactly once per slot.
- Retry: on failure (e.g. a cold-start timeout) the next regular run tries again.
- Idempotency: that is up to your prompt — write it so that repeated runs do no damage.
- Soft limit: 50 jobs per org on the standard plan. More on request.
Reconciler
If a run hangs during model inference (e.g. a worker crashes), a reconciler cron
picks it up every 10 minutes and marks it as failed with a reason. You can see
this in the job log under /orgs/<slug>/scheduled/<jobId>.
API for jobs
There is no REST API for scheduled jobs today. Jobs are created, edited and
deleted exclusively in the UI under /orgs/<slug>/scheduled.
Until recently this section listed three endpoints
(POST/GET/DELETE /api/v1/scheduled-jobs). They never existed — calling
them answers 404 (measured 2026-09-16). For the endpoints the public API
really has, see the
API reference; GET /api/v1/models is authoritative and always
current.
If you need headless management of scheduled jobs, tell us — demand decides whether we build it.