How it works

Four steps between an agent and your production.

ForgeOS is not a report you read tomorrow. It is a decision made now, on the path the action has to take.

execution pathexample
intent
agent:ops-copilot → db.migrate
intercept
boundary: mcp-proxy
policy
production.schema.write
verdict
BLOCKED · <50ms p99 · recorded
01

Discover

ForgeOS builds a live inventory of every agent, copilot and script touching production — including the ones nobody declared.

02

Intercept

The action is captured at the execution boundary: an SDK call, an HTTP gate or an MCP proxy in front of your tools.

03

Decide

Policy is evaluated within a <50ms p99 budget and returns ALLOW, HOLD until you approve, or BLOCK. Unknown means blocked.

04

Record

Every verdict, payload hash and approver is written to a hash-chained ledger you can hand to an auditor.

Questions

What teams ask first.

Is this another gateway?

No. A gateway sees the request an agent makes. ForgeOS sees the action it executes, at the boundary where it lands.

What happens if ForgeOS is unreachable?

Fail-closed by default. Sensitive actions do not run without a verdict; you can scope which classes fail open.

Does anything leave our network?

Agent payloads stay on your infrastructure — VPC, on-prem or fully air-gapped.

How long does a pilot take?

A design-partner pilot is typically scoped in one call and instrumented on one agent within days.

Where does enforcement sit?

Policy is evaluated before the action reaches your tools, then enforced on the host, with an optional OS-level deny below that. One path, fail-closed.