Slack-native intake. Autonomous workflow routing. Every run observable on a Mac your team actually controls.
The Loop
@watchtower can you review this PR before lunch? Also the toast bug if it is obvious.
Got it. Starting PR review on newton-web. Forking a bug-fix lane for the toast issue.
routing: PR_REVIEW -> codex/review
routing: BUG_FIX -> codex/fix (forked)
guardrails: allowlist=pass channel=pass
41s
duration
$0.08
cost
2
files
Why Watchtower
Not just a bot that replies in Slack. A system for turning fast-moving engineering requests into visible, governed work.
The thesis
If Slack is where work arrives, Watchtower is the calm machine that catches it, chooses the right move, and leaves the lights on so leadership can inspect what happened.
A tighter loop between request, execution, and proof — without asking the team to learn a brand-new ritual just to get help.
Slack-native intake
Mentions become missions. Intent detection routes the right workflow without forcing people into another dashboard.
Visible autonomy
Traces, sidecar output, and failure diagnostics stay inspectable. Leaders see evidence, not a black box.
Mac-local control
Watchtower sits on a Mac, supervises locally, and keeps the control plane close to the people shipping code.
Learning loop
Corrections, personality memory, and failure patterns feed back so the next run is less noisy and more aligned.
How It Works
Inputs are captured, the right workflow acts, and the whole system leaves enough exhaust for supervision and trust.
A mention lands. Watchtower parses the ask, detects intent, and frames the job before chaos can spread.
INPUT: REAL TEAM LANGUAGE, NOT CEREMONY.
PR review, bug fixing, owner autopilot treated like different instruments, not one giant hammer.
OUTPUT: ACTION MATCHED TO INTENT.
Runs stream into traces, recommendations, and learning signals so leadership can inspect the system.
PROOF: OBSERVABLE AUTOMATION.
Workflows
Watchtower detects what you're asking and routes it to the right workflow automatically. Each one built for a specific kind of engineering work.
Someone drops a review request in Slack. Watchtower pulls the diff, runs a multi-agent review pipeline, and posts findings back to the same thread.
A bug report lands in a configured channel. Watchtower classifies the repo, attempts a fix, runs tests, and opens a PR when confident enough.
Trusted operators bypass normal guardrails. Tag the bot directly and it executes broader tasks across your workspace without the usual restrictions.
Type wt commands in Slack and it becomes an operator console. Status checks, failure diagnosis, job replays, mission tracking — all from the thread.
Requests that don't map to a workflow still get a personality-aware reply, a thread reaction, and a desktop ping. Nothing silently disappears.
Observability
Traces, live sidecar output, recommendations, heat signals, and learned corrections.
Active
Streak
Chaos
Success
intent_correction: BUG_FIX reclassified
tone_feedback: less formal in #eng
trust_update: autopilot granted #deploy
Codex exit code 1 — context window exceeded
cause: repo too large for single-pass review
suggestion: split into file-scoped passes
action: auto-retry with chunked strategy
FAQ
For leaders who want autonomous developer workflows without giving up inspection, intentionality, or taste.