Desktop: option to auto-start Claude-suggested tasks locally (skip the "start locally" confirmation chip)
Problem
When Claude suggests spinning off a background task mid-conversation, the desktop app drops a chip that I have to click ("start locally") before it runs. For a long-running on-call triage loop that spawns one dedicated local thread per incident, every incident needs a manual click. That defeats the point of an automated responder.
Request
Add a setting (Settings > General, and/or a per-session toggle) to auto-start Claude-initiated suggested tasks locally without the per-chip confirmation. Something like "Always start suggested tasks locally". Keep the current chip behavior as the default; this would be opt-in.
Why the existing alternatives do not work
There is a real capability gap, because the three execution surfaces are mutually exclusive on what they can do:
- Desktop "start locally" chip: has my managed connectors (Datadog, Slack, runbooks) and local DB access, but needs a manual click each time.
- Headless
claude -pfrom a script: auto-starts, but runs on an inference-only token and loads none of my managed OAuth connectors (only locally-configuredmcpServers). Triage that depends on Datadog and Slack breaks. - A
claude.aicloud routine: has managed connectors and auto-starts, but cannot reach my local database or local creds, which the remediation step needs.
So today there is no way to get auto-start, managed connectors, and local access at the same time. The desktop spawned-task surface is the only one with connectors plus local access, and it is exactly the one that forces a manual click. An auto-start option on that surface would close the gap.
Use case
An on-call PagerDuty responder running as a recurring desktop session that, per incident, spawns a dedicated local triage thread (Datadog and runbook driven, stages a fix against the local DB).
Environment
- Surface: Claude Code desktop app (Mac)
- Related:
spawn_tasksuggested-task chips, managed MCP connectors vs localmcpServers
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗