Watcher

Watches your application logs, notices crashes and errors on its own, and explains the root cause with a small self-hosted model (via Ollama). It runs in the background and reports — you never invoke it against an incident by hand.

Install

Linux or macOS — no Go toolchain, compiler, or build step:

curl -fsSL https://github.com/shaheeranser/watcher/releases/latest/download/install.sh | sudo sh

The installer verifies the download against the published checksum and puts watcher in /usr/local/bin. Set WATCHER_INSTALL_DIR="$HOME/.local/bin" for a per-user install that needs no root. To remove it:

curl -fsSL https://github.com/shaheeranser/watcher/releases/latest/download/install.sh | sudo sh -s -- --uninstall

Run

Watcher names no model of its own, so point it at a local Ollama and one log source:

ollama serve
ollama pull qwen2.5:1.5b

watcher run --model qwen2.5:1.5b --file /var/log/app.log

Use watcher --version to report the binary's build version when filing a bug report.

--model is required. Results render for a terminal, or as one JSON object per line when redirected:

watcher run --model qwen2.5:1.5b --file /var/log/app.log > incidents.jsonl

Sources

tail -f app.log | watcher run --model qwen2.5:1.5b               # stdin
watcher run --model qwen2.5:1.5b --file /var/log/app.log         # one file
watcher run --model qwen2.5:1.5b \
  --source api=/var/log/api.log \
  --source worker=/var/log/worker.log                            # several, labelled

Inside Docker Compose, watcher run with no source attaches to the project's sibling containers over the Docker socket; --containers project=NAME scopes it explicitly and --containers none disables it.

A live terminal UI over the running daemon is available with watcher attach.

Configure once (optional)

watcher onboard writes a config file and can install the systemd unit:

watcher onboard
systemctl enable --now watcher

It prompts for the Ollama URL — verifying it is reachable before writing anything — a model, and one or more sources, then writes ~/.config/watcher/config.toml. Settings resolve as flag > environment > config file > default, and every flag has a WATCHER_* equivalent. Watcher never self-daemonizes: it runs in the foreground and leaves backgrounding to your supervisor.

Docker Compose

Runs Watcher alongside its own Ollama, pulling the model on first boot:

WATCHER_MODEL=qwen2.5:1.5b docker compose up --build

Build from source

go build -o bin/watcher ./cmd/watcher

Contributing

See CONTRIBUTING.md; AI coding agents should start at AGENTS.md.

License

MIT — see LICENSE.

Design rationale and per-milestone specs live in specs/.