RESTRICTION LAYER
FOR AI AGENTS

The runtime execution firewall and control plane between AI agents and external APIs.

Use the dashboard or the agen CLI

The restriction layer

Four primitives.
One accountable runtime.

01 / 04

Execution Firewall

Route agent traffic through /proxy/<scoped_key>/<host><path>. Zero-trust upstream credential injection.

02 / 04

Live Human-in-the-Loop

Ask-All mode holds risky proxy calls in mid-air for 25s, pausing for instant web/terminal approval or returning 428 Pending Approval.

03 / 04

Bastion Site Protection

bastion.agen site protection with five permission rungs ([r], [w], [rw], [x], [rwx]), case-sensitive !redlist payload inspection, DNS-verified domains, and Redis DENIED strike-to-ban honeypots.

04 / 04

Agenlink Delegation

Single-use owner connect codes and time-bombed agent proxy URLs for upstream OAuth or credential delegation, with automatic revocation at expiry.

THE WHYs

01 / 04

Why now?

Agents went from drafting text to taking real actions — calling APIs, touching production, and spending money. The layer that decides what they are allowed to do before they do it is now essential.

02 / 04

Why was this not built already?

Until recently, agents mostly suggested actions. Once they could act instead of suggest, the risk became real — and the tooling had not caught up.

03 / 04

Why is this its own product?

Agent frameworks are built to make agents powerful. Agenline is the accountability layer above the framework, regardless of what the agent is built with.

04 / 04

Why not wait?

The first serious incident is too late to begin defining boundaries. Agenline lets teams establish rules before an agent reaches a sensitive resource.

A simple boundary

From open-ended
to under control.

For example

A clear path from intent to enforcement.

Connect the agent

1 / 5

The real upstream key stays encrypted at rest. Use a test key for this walkthrough.

agen agent connect support-bot --upstream-key-env STRIPE_API_KEY

Start with the dashboard or CLI.

Sign up, create a field, grant its state through a policy, then attach the policy to an agent. Inspect its effective access to understand each decision. HTTP fields are enforced through the Agenline proxy; filesystem, command-runner, and database configuration requires a trusted external runtime.