Request a demo · Read the setup guide
Who is using it. What it touches. What it costs.
Security wants the first two answered. Finance wants the third. Without them, a Claude Code rollout ends as a pilot for ten people and a spreadsheet. APERION Shield and Smartflow Shield answer all three on every call.
Four stops between the developer and the model.
- The laptop. Each laptop enrolls once and gets its own device certificate. Enrollment writes Claude Code's managed settings, and Shield runs inside Claude Code's hooks. Every prompt and every tool call is checked on the laptop before it runs. The organization's provider key is never stored on the machine.
- Sign-in. The person on the record is the person from your corporate directory: Microsoft Entra ID, Okta, Ping or any OIDC issuer, with SCIM for people and groups.
- The Smartflow gateway. Calls go to the gateway, never straight to the model provider. The gateway runs the policy and personal-data checks, enforces the spend cap for each person and keeps one log.
- The model. What passed reaches Anthropic's models.
The check runs before the action.
- What Shield sees. User prompts, shell commands, file edits and MCP calls.
- What it does. It applies your policy to each one. A command that wipes a directory stops, or goes back to the developer as a question.
- The device. Claude Code will not reach the gateway without the device certificate. A key copied to another machine does not produce a working setup.
- Settings a developer cannot change. The Claude Code settings file is owned by root. It allows managed hooks alone and turns Auto mode and bypass mode off.
One place for identity, policy, budget and the record.
- Identity. Every trace row carries the person from the directory and the laptop, in separate columns.
- Key custody. Provider keys stay on the gateway, in a sealed vault or your own key vault.
- Policy and personal-data checks. Applied to every call before it leaves the gateway.
- Spend cap. Each person has a cap, and a ceiling set on a directory group applies to each member. The cap is checked before every call. At the cap, the next call is refused with the amount spent and the limit. Finance reads spend by person and by group on a dashboard.
- One log. Claude Desktop goes through the same gateway and lands in the same log once the laptop is enrolled.
What one row answers.
Each call writes one row: the person, the laptop, the model, the provider, the cost and the policy that applied. Security reads who did what. Finance reads what it cost and against whose cap.
What it does not do.
- Commands a developer types in their own terminal are outside Claude Code.
- A tool called by its full path goes around Shield's command shims.
- Shield does not inspect an editor extension's startup code.
- A personal Claude account is allowed by default. Its calls go through the gateway, run the same policy and are recorded against the person from the directory. They do not count against the spend cap. An administrator can refuse personal accounts.
- A question from Shield goes to the developer at the keyboard. Approval by a second person applies to tool calls that pass through the gateway.
- Cursor sends most of its model traffic to Cursor's own service. From Cursor, the log carries Shield's decisions and tool events.
- This page states no savings, latency or uptime figure.
What a first rollout takes.
- Smartflow Shield, on one Smartflow gateway host inside your environment.
- Ports 443 and 444 from the laptops to the gateway, or two hostnames on 443 where the network allows 443 alone.
- Directory sign-in.
- Shield 1.26.0 or later on each laptop, delivered by your device management tool, and a one-time enrollment token from the dashboard.
Start with one developer. Enroll the fleet when you are ready.
Shield is free and local for an individual developer: local rules, hooks, a scan, a sandbox and a local approval inbox. Enrollment, the device certificate, directory sign-in, the spend cap and the central log come with Smartflow Shield.
Frequently asked questions
Does it matter which Claude account is on the laptop?
Every call is recorded against the person from your directory. Calls on the organization's key count against that person's cap. A personal Claude account is allowed by default, is marked on the record and does not count against the cap. An administrator can refuse personal accounts.
What happens when someone reaches their cap?
The next call is refused, and the message shows the amount spent and the limit. A response already streaming finishes. Finance sees it on the dashboard, before the invoice.
Does it cover Claude Desktop, Cursor and Codex?
Claude Desktop goes through the same gateway and the same log once the laptop is enrolled. Enrollment also writes machine-level hooks for Cursor and Codex. Most of Cursor's model traffic goes to Cursor's own service, so from Cursor the log carries Shield's decisions and tool events.
Can a developer turn it off?
The settings file is owned by root, allows managed hooks alone and turns Auto mode off. Claude Code will not reach the gateway without the device certificate.
Does it run on a network that allows 443 alone?
Yes, with two hostnames on 443. The default setup uses 443 and 444.
Claude Code and Claude Desktop are products of Anthropic. Cursor is a product of Anysphere. Codex is a product of OpenAI. APERION is an independent company.
Put this in the path of your own agents.
Policy enforced inline between your agents and every model and tool they reach, with a record bound to the human who owns it.
Request a Demo Read the docs