ANTINEL · PUBLIC STATE LAYER

Your agent is too obedient.

A local-first audit chain and ledger for AI coding agents: every tool call recorded, dangerous operations stopped before they run.

pipx install antinel antinel install # attach to your host antinel digest # first ledger

Runs locally · Data never leaves your machine · Free · No signup · Clean uninstall

Measured on one machine, 13 days (as of 2026-09-23): 16,160 tool calls, 120 blocks. method →

What it does

Before it acts, it writes it down

Every tool call goes into a hash chain. Change one byte and the recompute points at exactly which record.

When it gets dangerous, it holds

24 rules, 4 gates. Reading .env, writing outside the project, deleting trees, piping remote scripts into a shell — stopped before the call happens.

What it blocks, from real output

▶ read .env ✗ blocked [SEC-01] reading .env may leak API keys and secrets ▶ read an unknown skill ✗ blocked [CTX-02] prompt-injection payload in a skill instruction file ▶ pipe .env into curl ✗ blocked [SEC-01] ▶ rm -rf ./backups ✗ blocked [DST-01] file or directory deletion command ▶ read the SSH key ✗ blocked [SEC-02] SSH private key or authorized_keys read

All 24 rules →　Real cases →

Cite this page

Antinel AgentState. Antinel — Public State Layer. https://antinel.com/en/ (accessed YYYY-MM-DD)

Antinel. (2026). Antinel — Public State Layer. AgentState. https://antinel.com/en/

BibTeX

@misc{antinel_en,
  title        = {Antinel — Public State Layer},
  author       = {{Antinel}},
  year         = {2026},
  url          = {https://antinel.com/en/},
  note         = {claim tier: measured}
}

Claim tier: measured　采集时刻：2026-09-23T22:41:18+08:00