Let coding agents resume without starting over.

Agent Watchtower is a tiny local CLI that records goals, next safe actions, run receipts, and artifacts so interrupted AI coding work can continue cleanly.

View on GitHub Install

Local files only. No background service. No external account required. Current version: 0.1.4.

$ agent-watchtower init
initialized: 1 goal, 1 open task

$ agent-watchtower worker-status
next_safe_action:
  write one local artifact

$ agent-watchtower worker-run --result "Recorded the actual work result."
completed_task
artifact:
  work-artifacts/task-watchtower-demo-first-step.md

What it keeps after the session stops

Watchtower does not replace the agent. It gives the next session a small work notebook that is easy for humans and agents to inspect.

Goal

The work the agent was trying to finish.

Next safe action

The next bounded step that can be resumed without guessing.

Artifact path

The last useful markdown output written by the worker run.

Before

The terminal closed. The next agent window has no local trail and asks the human to explain again.

After

The next session checks Watchtower state, finds the latest artifact, and continues from the recorded next action.

Run a local demo

Install from PyPI or Homebrew, then run a disposable local demo. The generated runtime can be deleted whenever you are done testing.

Runs as a Python 3.11+ CLI on macOS, Linux, and Windows. macOS is tested in the current release flow; Linux and Windows are supported targets, with CI coverage planned.

python3 -m pip install agent-watchtower-core

# Or install with Homebrew:
brew tap un-n-smith/tap
brew install agent-watchtower

# Or install from GitHub with pip:
python3 -m pip install "agent-watchtower-core @ git+https://github.com/un-n-smith/agent-watchtower.git"

ROOT="$(mktemp -d)"
agent-watchtower --root "$ROOT" init
agent-watchtower --root "$ROOT" worker-status
agent-watchtower --root "$ROOT" worker-run --result "Recorded the actual work result."
agent-watchtower --root "$ROOT" artifact-path