A computer is required to install Sentinel. Sentinel runs as a desktop app (Windows, macOS or Linux) with Python. Feel free to read everything here on your phone — but open this page on your PC when you're ready to download and set up the SDK.
Git tracks users. Sentinel tracks agents. Lock your codebase to a trusted AI, log every prompt and run, and gate untrusted agents behind a single-click approval queue.
Sentinel is a nice-to-have for an individual building their own agent team. But it's a must-have for an organization building a mission-critical team of agents.
Free account · $20/mo to unlock the engine · cancel anytime
# 1. Drop sentinel_sdk.py into your repo, then run one command: python -m sentinel_sdk init --token $SENTINEL_TOKEN # Sentinel creates your project, installs a git hook, # and opens your live dashboard — no other setup. # 2. Commit as normal — every commit is reported automatically: git commit -m "Fix Site Policy crash" # Sentinel: 6e3e3a87 -> approved (risky writes wait for your OK)
Your AI agents are already writing real code — opening files, editing logic, touching configs, shipping changes. So here's the uncomfortable question: who's governing them? For most teams, the honest answer is nobody. The agent has the same keys as the developer who ran it.
We think that's backwards. If an agent can change your code, it should be a first-class member of your team — with an identity, a defined scope of what it's allowed to touch, and a paper trail for everything it does. Not a faceless script with root access. That's why we built Sentinel.
Scoped, auto-expiring API keys — know who did what, rotate or revoke in one click.
Agents write only where you say. Everything else is off-limits by default.
Touch .env, payments, auth, migrations or deploy configs and it won't auto-merge.
Risky or out-of-scope commits wait for a human — with the prompt, model and diff.
One toggle freezes every agent write on a project. Stop the line instantly.
Every grant, block and approval, recorded. Answer “what changed?” without guessing.
Wiring it in takes about a minute with our single-file Python SDK — drop it into your repo and your agents' commits start flowing through Sentinel.
Agents are about to be a permanent part of how software gets built. They should be governed like it.
sci-ficomics.com/sentinel
Git was built in 2005 for human developers. When agents commit, you lose the prompt, the model, the run, and the chain of authorship.
Attribution is broken.
"Committed by Claude" or "Committed by Codex" tells you nothing. Which prompt? Which model? Which run?
There's no lock-down.
Any agent with repo write access can rewrite your production code. There's no concept of a "trusted" agent identity.
Audit is an afterthought.
Reconstructing what a rogue agent did across 200 commits is a Friday-night nightmare. There's no replay.
Watch a commit travel through Sentinel.
Fix BrowsePage.jsx
frontend/src/pages/BrowsePage.jsx
Kestrel · primary
Wire stripe webhook
backend/.env + 1 more
Kestrel · primary
Edit /etc/hosts
/etc/hosts
Codex-Outsider
allow_paths · 4 rules
sensitive_patterns · 14 rules
rate_limit · 100 / hr
kill_switch · OFF
Auto-merged
Primary agent · no risk flags
Human approval queue
Sensitive files — even from a trusted agent
Rejected · logged
Outside allow-list — caught instantly
Every push from your agent gets classified in real time. Clean commits ship instantly; sensitive ones pause for your sign-off; nothing bypasses the kill switch.
Launch it with a single command from our Python SDK and a live, always-on window docks to the corner of your screen. Watch commits land, spot flagged writes the moment they happen, and keep an eye on every agent — without ever opening a browser tab.
Real-time commit counts — approved, pending and rejected — refresh automatically as your agents work.
Flagged commits that touch .env, payments or auth surface immediately, with clean-vs-flagged tallies at a glance.
See which agents are generating the most flags so you know exactly where to tighten the guardrails.
It's a real desktop window — drag to resize, or minimise it out of the way, using your system's window controls.
python -m sentinel_sdk dashboard --token … and it opens in your browser — served right from your machine.
python -m sentinel_sdk dashboard \ --token $SENTINEL_TOKEN
Know exactly what your agents shipped — every month.
Sentinel rolls up every commit into a clean month-over-month report: how many shipped, how many waited for approval, how many were flagged for touching sensitive paths — and how many projects and agents were active. Export it as a CSV in one click for your board deck or compliance file.
The exact desktop window · live interactive demo · click a flagged commit to inspect its diff
Four steps from chaos to a fully audited, lock-down codebase.
Designate a trusted agent as the project's primary. Once locked, no other agent can write without explicit approval.
Your agent uses the Sentinel SDK to push a commit. Every push includes the prompt, run id, and full file diff for audit.
Commits from your trusted primary agent merge automatically. Everything else pauses for a one-click human decision.
Nothing is silent — every write, granted or denied, is recorded so you can always answer who changed what, when, and why.
It's the cloud — not your network.
A common question: does the Sentinel app need to reach across my network to find commits? No. Sentinel is cloud-based. Each commit is reported up to the cloud by a tiny git hook, and your dashboard simply reads it back. Your computer never scans the network or touches anyone else's machine.
On any machine, a developer or AI agent makes a commit in a repo where you ran init.
Sentinel's git hook fires and sends one outbound HTTPS call describing the commit.
Sentinel's servers record it under your account and project — the single source of truth.
Sign in on any device and the desktop app pulls your feed down from the cloud. Read-only.
Installing the hook just writes a file into a repo you already own (.git/hooks). No OS elevation, no domain or network permissions — ever.
Commits flow up to the cloud from each contributor. Your viewer only needs plain internet access — no inbound ports, no firewall changes, no LAN scanning.
Each repo or agent that runs init reports on its own. Add teammates or agents and the feed grows — being on the same network changes nothing by itself.
Who's it for? Teams (or solo devs) who want governance and an audit trail over code changes — especially changes made by AI agents. The owner subscribes; each developer or agent installs the hook in their own repos. Everyone's commits roll up into one governed feed.
We treat your agent as a first-class citizen of your dev team. Not a black box.
Track which AI agent made each change. Co-authored by a human + an agent? Both get attribution.
The prompt that produced each commit is attached as first-class metadata — searchable, auditable, replayable.
Python first, language-agnostic next.
Three commands and your agent is publishing fully-attributed commits. No git plumbing, no auth dance, no YAML.
from sentinel import Sentinel
sentinel = Sentinel(
agent_key=os.environ["SENTINEL_AGENT_KEY"],
project="sci-ficomics",
)
@sentinel.commit(message="Fix Site Policy crash")
def apply_change(diff, prompt, run_id):
# ...your agent's code here...
return diff$20 / month — everything unlocked.
$20 per month per user — every feature unlocked. No tiers, no hidden limits, no surprise overages. See pricing →
Sign up free, name your first project, register an agent, and drop the SDK into your code — most teams are governing live commits in under a minute. Subscribe for $20/mo when you're ready to run.
Free to create an account · no card required to sign up