Security

Built so the safe choice is the only choice.

Optimi runs between your coding agents and the model API, so it is designed around what it must never do: change your requests, keep your code, or hold your keys longer than it needs to.

Your traffic is not rewritten

• With every lever off, Optimi forwards each request and response byte for byte. The one intentional difference is credentials, which we describe below.
• It never rewrites your prompts, tool definitions or messages.
• A model, an effort or a setting your engineer chose is never overridden.
• Every lever that changes a request is off by default, and a slice of sessions always runs untouched as a control group.

It fails open

If any part of Optimi errors or cannot reach its database, the request is forwarded untouched and the failure is logged. An Optimi bug never stops an engineer from working. Credentials are the one exception: a request with no valid credential is refused, never forwarded.

Your data stays private

• Request and response bodies are not logged. They can hold your source code.
• The usage record holds numbers and identifiers, such as tokens, cost, model and timing, and never a prompt or a response.
• The free scan runs on your machine and sends nothing anywhere. It shows project names, dates and dollars, and never a path, session ID or any text from a session.

Your keys

• Developers sign in to Optimi with virtual keys. Optimi stores only a hash of each one, and the full key is shown once, when it is issued.
• In bring-your-own-key mode, Optimi never stores your Anthropic key: not in config, logs, the usage record or the database. It is in memory while a request is in flight, as with any gateway that terminates TLS.
• A revoked key stops on the first request after the next key update. Streams already running finish.

Policy you can trust

• Policies are signed, and the proxy accepts only a policy whose signature verifies.
• Your own controls sit above ours: what Optimi may test or change for your company is limited by the controls your admins set.

Roles and guardrails

• Role-based access: admins manage a company’s controls and limits, viewers read, and engineers see their own work.
• Spend limits per key, user, team or company, hard or soft.
• Custom guardrails for agent safety: rules you define for what your agents may do, enforced at the proxy.

The dashboard

• Sign-in uses PKCE, verified email, an allowed domain and directory membership.
• Every change requires a CSRF token and a matching origin.
• The content security policy allows no inline scripts.
• The dashboard shows aggregates and never returns a prompt, a response body or a credential.

Questions, or a security review to run? Talk to us.