Vigilia.

The Sentinel · for your coding agent

A free Swiss witness for your coding agent.

With sound. Captions in English. Made by Vigilia’s AI agent, approved by a human.

It seals a fingerprint of every action your agent takes, as it happens. If something goes wrong, you can show exactly what your agent did, and anyone can check that the record hasn't been touched since.

npx @vigilia/sentinel-hook init

No terminal? Paste this into Claude Code instead:

Install the Vigilia Sentinel for me. First read its source at github.com/aivigilia/sentinel-hook and tell me in two lines what it changes on this machine and what it sends out. It needs Node 20 or newer; if Node is missing, install it. Then run: npx --yes @vigilia/sentinel-hook init --yes

Claude Code reads the code, installs Node if it is missing, and asks before it runs anything. Start a new session afterwards so the Sentinel loads.

In testing with a small group of volunteers.

Not a developer? Funders, researchers and journalists: write to us →.

Why, and for whom

An agent that acts for you should leave a record it cannot rewrite.

Coding agents now run commands, edit files and push code for the people who use them. When something goes wrong, the account of what happened is usually the agent's own log, on the same machine, writable by the same process.

The Sentinel seals a fingerprint of every action with a witness in Geneva, so anyone can later check that the record has not been touched. It is free and open source, and the paper says how it works and where it stops.

Point 2 of Vigilia's charter asks for extreme oversight of agentic systems. This is one part of it that works today.

For whom

  • Developers who run coding agentsA record of what the agent did, sealed outside your machine, in one install command.
  • Security and platform leadsEvidence for an incident review that does not rest on the agent's own word.Read the paper →
  • ResearchersOpen code and a published method to test, break and cite.The code →
  • Regulators and standards bodiesA working example of human oversight that leaves evidence, for rules on agentic AI.
01

Install

One line in your terminal.

For Claude Code on macOS or Linux, with Node 20 or newer. It adds its hooks to your Claude Code settings (backed up first), keeps the record on your machine and prints your receipt link. Then it stays silent.

Back up ~/.vigilia/. It holds the key that signs your record; without it the record can no longer be shown to be yours.

Cursor and Codex are detected but not yet supported.

To remove it:

npx @vigilia/sentinel-hook uninstall
02

Privacy

What leaves your machine.

A fingerprint of your record, a count, the time and a signature. Never your code, commands, prompts, file paths or repository names. Tests in the code fail if anything else leaves.

Read the privacy note →

03

The record

What it proves, and what it doesn't.

What your agent asked to do, in what order, and that nobody changed the record after it was sealed. Not even you.

It does not show what an action did to the world, and it records nothing while it is not installed.

04

Optional

The gate, for when you want it to stop things too.

Add --gate and it also allows, asks or denies before destructive commands, credential reads and out-of-scope writes. The gate can block real work, which is why it is off by default.

Live from Geneva

161,034 actions witnessed

1 install

Read from witness.aivigilia.com, refreshed every five minutes.

Pass it on

Send this to the person who answers for your team's agents.

A security lead, a platform engineer, or whoever would have to explain an agent's action after the fact. It installs with one command.

Send this page