Aegis Legal Checkpoint · what your agents say

Every communication, checked against the law before it sends.

The last thing an agent's message passes through, and the first thing a regulator will ask about. Allow, block, or hold, with the statute cited, sub-second on the standard path.

Scope: agent communications only · email · SMS · chat · voice scripts · social

Talk to us

e·01 how it decides

Intercept

The agent calls its send tool. Checkpoint sees the action first. Nothing has left the building.

Decide in tiers

Deterministic rules end most traffic in about 3 ms. An exact-match cache catches repeats. A small model reads the ambiguous. The hard 2 to 5 percent gets a frontier model.

Answer with a statute

Every non-allow cites the rule, the version it came from, and a suggested rewrite. Verdicts are explanations, not vetoes.

Chain the record

Verdict, rule version, hash: into your ledger, not ours. Content is already gone from our side.

{
  "verdict":   "QUARANTINE",
  "basis":     ["UWG §7(2) prior consent", "GDPR Art. 6 basis unverified"],
  "tier":      "T3",
  "confidence": 0.87,
  "rules":     "de-comms v2028.31",
  "suggested_rewrite": "available",
  "hash":      "9c4e…a1",
  "chained_to": "a3f0…c9"
}

e·02 two modes

Advisory Recommends, never blocks Ten-minute self-serve install. The agent asks, Checkpoint answers, your team sees every recommendation. The right first step for most teams.

Enforcing Holds, quarantines, blocks The proxy sits in the path and enforces your policy: what holds, what blocks, how long to wait for a human. Per tenant, per action class, recorded.

p50 5 to 10 ms rules tier terminal for most traffic frontier escalation 2 to 5% hold deadline default 10 s, then human content retained zero

// your agent's MCP config. that is the install.
"aegis-checkpoint": {
  "command": "npx",
  "args": ["@aegis/checkpoint"]
}

Talk to us Need internal agents covered? Protect