[FEATURE] Cowork scheduled tasks: per-routine connector scoping to cut cold-start cost

Status Open
Maintainer reply None cached
Activity 0 comments · opened Aug 29, 2026

This concerns Cowork (claude.ai) scheduled tasks ("routines"), not the CLI; CLI-only template fields are not applicable.

The problem

Every firing of a Cowork scheduled task boots a fresh session that loads the tool lists and server instructions of every connector on the account. On an account with ~25 connectors, a routine that needs three or four of them (e.g. an instant message relay that needs Gmail and one chat bridge) still pays the full connector payload on every single firing — for high-frequency routines that is the dominant startup cost, paid dozens of times a day.

What we checked before filing

  • The scheduled-task records returned by GET /api/organizations/{org}/cowork/scheduled_tasks carry no connector field per task — there is no hidden knob to set.
  • Connectors are scoped to the account, not to a project, so binding the routine to a leaner project does not help either.
  • The only trim available today is disconnecting a connector account-wide, which breaks the other routines and interactive sessions that need it.

The ask

Let a scheduled task declare which connectors its runs load — at creation (an optional connectors: [...] on the task) or in settings. Default stays "all", so nothing changes for existing tasks. This would cut cold-start latency and token cost roughly in proportion to the trim for every high-frequency routine, and would also narrow what an unattended run can touch, which is a safety win for the same feature.

Related but distinct: deferred tool loading already avoids loading full schemas, but the per-server name lists and instruction blocks still load on every firing, and that is the part per-routine scoping would remove.

🤖 Filed via Claude Code on behalf of TheAviv

View original on GitHub ↗