ANCC is a short contract a command-line tool publishes as docs/SKILL.md next to its human instructions. A person and an agent can read the same facts: how to install it, what its commands do, what they change and where they stop.
Each result is marked observed, declared, inferred or unknown. It states what it establishes and what it cannot show. The source of a fact is recorded separately from the time it was observed, because an observation can go stale.
A safeguard is marked advisory, enforcing or unverified. A passing probe covers one operation, one configuration and one version. It does not make a tool safe.
An agent may narrow its own access within a scope someone authorised. It must never restore or widen that access. Tightening access can also break things, so permission to narrow it has limits.
Output names the failed checks. It says which results are incomplete, which dependency failed and what was changed or withheld. A score cannot replace those facts.
---
name: bulwark
description: "Kernel-boundary file-read gate for AI agent process trees — pause a protected file open at the kernel on Linux or macOS and decide before any bytes reach the agent"
user-invocable: false
metadata: {"requires":{"bins":["bulwark"]},"platform":["linux","macos"]}
---
# bulwark
## Install
## Commands
## What this does NOT do
## Handoffs
## Failure Modes
## Parsing examples
From the Bulwark repository's docs/SKILL.md.
ancc.dev hosted this convention and a catalogue of forty tools from February 2026. The catalogue closed on 30 October 2026 and is not coming back.