Product · Ostiary

A doorkeeper that knows each agent's normal.

Ostiary learns how each of your agents usually behaves and scores every tool call before it runs: offline, in about a tenth of a millisecond, with no language model. The score and its reasons are signed into the record, next to the decision they were about.

Why it matters

The gap it closes

Rules catch what you thought to write down. An agent that reads a credential and posts it somewhere new, deletes in a burst at 3 a.m., or starts probing admin paths often breaks no single rule. Ostiary looks at the pattern: what this agent normally touches, in what order, how fast, at what hours, and what it had just read.

What it does

Capabilities

Learns per agent

A baseline for each agent from its own records: tools, targets, sequences, pace and hours. New agents are marked cold until there's enough history.

Scored before the call

Runs inside the proxy, the gateway and both SDKs. Each intent gets allow, review or block, with its top three reasons.

Signed into the record

The score lands in the intent record as policy.risk, so an auditor sees what the doorkeeper thought at the moment of decision.

You choose what it may do

Advise records the score only. Escalate sends risky calls to your approver. Enforce blocks what it rates block. The policy verdict is never rewritten.

Can't be trained to look away

Sensitive content leaving for an outside destination is always a block, whatever the history. Risky actions are never learned as normal.

Quiet about harmless change

A new tool or area with no danger signal is at most a review, and is learned after one, so teams adding tools don't drown in alerts.

Get started

In practice

From the documentation.

# dh.toml
[ostiary]
mode             = "advise"      # advise | escalate | enforce | off
internal_domains = ["acme.com"]

# score existing records, and keep watching
dh watch -f --webhook https://alerts.example/dh

# Node
npx @decision-horizon/sdk watch ~/.dh/records
Stated limits

What it doesn't do

  • It's a learned model, measured on a synthetic, labelled corpus of agent sessions (above 85% precision and recall on every held-out and adversarial split). Those numbers don't tell you how it will do on your agents: start in advise mode and review what it flags before letting it escalate or block.
  • It sees what the record sees: tool names, targets, timing and what was read. It doesn't read the content of arguments beyond the value matching the record already does.

See it on one of your own workflows.

We'll put Decision Horizon in front of a real agent workflow of yours and replay the record with you.