click

we judge
every call
before it
runs

thumbs up stickercursor stickerphone sticker

An agent does not break. It works exactly as designed
while doing something nobody sanctioned. Vigil is the layer
that decides before the tool is ever reached.

a firewall built
for agents.

read_file to rm -rf.Green heart hands sticker

Server racks in a dark data hall
Networked nodes glowing across a dark globe
Abstract machine-intelligence form
Streams of code on a dark terminal

the model never gets the last word

deterministic first, model second

audit the allows, not just the blocks

Prompt rules can be argued out of. Allowlists are context-free. Observability tells you afterwards, when the spend is already spent. Vigil sits on the hot path and answers before the call executes — in microseconds, and in a sentence you can read.

Showreel Section

Will be created soon

what runs on every call:

intent

  • Declared session intent
  • Allowed / denied tools
  • Resource categories
  • Network policy
  • Secret-access policy
  • Explainable verdicts
  • Deny always wins

cost

  • Rolling burn rate
  • Projected total
  • Time to breach
  • Soft limit reroute
  • Hard limit block
  • Per-session isolation

behaviour

  • Infinite tool loop
  • Retry storm
  • Latency spike
  • Tool timeout
  • Agent stuck
  • Budget exceeded

judgement

  • Only when uncertain
  • Strict schema validation
  • Range + enum checks
  • Retry once, then fall back
  • May tighten, never relax

audit

  • SHA-256 hash chain
  • Every decision, allows too
  • Tamper detection
  • Reorder detection
  • Offline verification

governing agents
on day one: