Agent Bridge

Monitoring only. Task cost can vary; usage data does not authorize dispatch. Subscription-plan allowances and pay-as-you-go API usage are tracked separately, and shared model-pool allowances are counted once (not per model). Readings come from the local collector cache; the background collector measures at most every ~5 minutes.

Sources are shown per card. Passive snapshots can lag; experimental connectors may change. No hard spend cap.

Upload a track, describe the video, review the eight shots, then generate. Clips render locally on the WanVideo backend, one job at a time (extra jobs queue visibly). Draft prompts are deterministic starting points -- edit any shot before rendering.

Opt-in, project-scoped monitoring -- nothing runs until YOU assign a project and explicitly enable it. Bots propose coding changes/features only: monitoring runs with owns=[], so a bot can never edit, execute or deploy. Local (Ollama) workers auto-dispatch under the built-in review budget. For cloud workers: Grok supports budget admission -- configure a USD budget below (PIN-gated) and it becomes eligible for monitoring dispatch within that budget; Copilot has only a soft CLI cap and still needs an enforceable provider budget, so its auto-dispatch stays blocked; Kimi Operator handles explicit GUI tasks only and is never an automatic reviewer. Saving a policy sends no inference, but an enabled policy lets already-enabled assignments dispatch at the next tick within budget -- new assignments are still created paused and nothing dispatches until you enable it. The Bridge process must stay running for a scheduled review to fire -- a server-side bot identity does not keep this PC awake or make uptime irrelevant.

Cloud monitoring policy

Workers

New assignment (created paused)

Assignments

Recent suggestions