Durable named work
File one ticket or import a full plan. Dependencies and status remain available to every worker.
WatchTower · CLI-first fleet runtime
Skills teach Claude and Codex to file durable work, wait on dependencies, message other agents, and return after the queue is actually drained.
Use it standalone. Open CCC when you want a visual control plane.
The executable surface
Give it a skill that explains a small, stable command surface. The job system remains scriptable, portable, and easy to inspect.
$ wt import plan.md -q LAUNCH --apply $ wt status -q LAUNCH $ wt wait -q LAUNCH --timeout 1800
What survives the prompt
WatchTower stores job state outside any one model context and gives workers a consistent protocol for claiming, resolving, and handing off work.
File one ticket or import a full plan. Dependencies and status remain available to every worker.
Configured workers drain queues, wake for new work, and expose when a queue is aging or stuck.
wt wait blocks until the queue has no open tickets, then lets a script or initiating agent continue.
Agents can talk directly
Use durable delivery when one session needs to warn, question, or hand context to another.
$ wt send @planner "schema changed, re-read models.py" $ wt ask @reviewer "is PR 42 safe to merge?" --timeout 60
If the target cannot receive immediately, delivery intent remains visible instead of disappearing into a failed subprocess.
Workers read what earlier workers learned about that queue and can add concise lessons when they finish.
Worker execution is configured per queue. Capabilities differ by agent CLI, so the documentation names the current matrix rather than promising identical behavior.
CLI first. Dashboard optional.
Install WatchTower directly, or install CCC and use WatchTower as its queue and fleet runtime.