Questions we hear
on every call.

Cortega is software that runs on your own servers. It sits between your apps and AI agents and the models they call — like GPT-4, Claude, or Gemini — and decides in real time whether each request should go through, get blocked, get edited, or wait for a person to approve it. It does this before the request goes out, not after the fact.

It checks for sensitive information — things like Social Security numbers, medical records, or classified government data — using fixed rules instead of relying only on an AI model's best guess. And it manages the AI gateways you already use, like LiteLLM, Portkey, or Kong, from one place, instead of replacing them.

No. Cortega installs inside the cloud environment you already use — AWS, Azure, or wherever your team already runs things — so you don't have to move anything or change how your systems are set up. It works with the tools you already have, like Harvey, ChatGPT Enterprise, or Claude, instead of asking you to rip them out and start over.

Yes. ChatGPT, Claude, Harvey, and Copilot are all just apps that send requests to an AI model behind the scenes. Cortega sits in that path for every one of them, so it can keep a single record of what happened — who asked what, when, and with which tool — instead of each tool keeping its own separate log that nobody can piece together. That matters most when different teams are using different tools and no one can see the whole picture.

Yes — we call this Cortega Verifier (patent pending). It checks whether a model's answer is actually backed up by your own documents and data, not just whether the person asking was allowed to ask the question. It's built to catch a made-up or wrong answer before it reaches a customer, employee, or regulator.

Cortega AI Bench lets you test how your governed AI stack behaves, before and after a policy change. It runs industry-specific safety suites through a chosen gateway and model, then scores whether your active guardrails actually allowed or blocked the samples they were supposed to, and keeps a history so you can see the effect of a change over time.

Benchmark traffic is tagged and kept separate from production traffic, so testing doesn't skew your real observability data. More benchmark types, covering agent behavior, MCP tool-result safety, and performance under load, are planned. AI Bench is part of the PRO tier and above.

No. Cortega runs inside your own network, not a Cortega-hosted cloud, and can run fully disconnected from the internet if you need that. Your requests never get routed through cortega.ai — think of it as a firewall you install inside your own network, not an outside service you send data to. Your data stays on your systems, and we don't make money from the traffic or tokens that pass through it.

Cortega finds and blocks or masks sensitive information before a request goes out — things like Social Security numbers and credit card numbers (PII and payment data). It uses fixed rules to do this, not just an AI model's guess, so the results are consistent. It also controls which AI models different people are allowed to use — for example, letting only senior staff use a more powerful (and more expensive or higher-risk) model, while everyone else uses a lower-risk one.

Yes. Cortega ships a prompt injection guardrail that watches requests and responses, including MCP tool calls, for attempts to override a model's instructions or pull out its system prompt — phrasing like "ignore previous instructions" or "reveal your system prompt" — and blocks the request before it goes through.

It's pattern-based, built to catch known injection phrasing rather than every possible obfuscated attempt, and you can add stronger detection, like Azure Content Safety's jailbreak model, as an additional guardrail provider if you need broader coverage.

Yes, as a guardrail on your AI traffic. Cortega can watch prompts, model responses, and MCP tool calls for SQL injection patterns — things like UNION SELECT, DROP TABLE, or SQL comment markers — and block them before they reach a downstream tool or database.

This is aimed at a model or agent being tricked into forwarding a malicious query through a tool call, not a general security audit of every system it touches.

Starting at the Enterprise tier, Cortega includes what you need to show GDPR and HIPAA compliance, plus help preparing for a formal certification audit. It's built to produce the paperwork and logs an auditor will actually ask for, not just an internal dashboard.

Yes. Cortega works with local models you run yourself, custom models your team built, and the big commercial providers like OpenAI and Anthropic. That includes fully disconnected setups where nothing ever leaves your network.

On LiteLLM's own benchmark (AIGatewayBench), Cortega added about 0.33 milliseconds of delay per request — about half of LiteLLM's own Rust gateway (0.66 ms), and faster than Portkey (2.29 ms) and Bifrost (4.54 ms). It does this while also enforcing policy, protecting sensitive data, and governing MCP tool calls — for example, when an agent looks something up in a database or runs a piece of code — on that same request.

The bigger difference is what's included. LiteLLM charges extra for single sign-on (SSO), audit logs, and full control over MCP tool calls, and it doesn't cover AI tools people use outside of your production systems — things like ChatGPT or Claude opened in a browser. Cortega includes all of that as part of the platform, and a separate product called Cortega EdgeSafe extends coverage to that everyday employee AI use, not just your production traffic.

See the full side-by-side comparison for the detailed feature and cost-of-ownership breakdown.

LiteLLM is a solid gateway for connecting to AI models and routing requests — we're not asking you to replace it. What LiteLLM either doesn't do, or only offers on its paid tier, is full enterprise governance: single sign-on (SSO) tied to policy and audit logs, complete control over MCP tool calls (not just some of them), coverage for AI tools employees use outside your production systems, and keeping all that governance data out of your main database so it doesn't slow down as it grows. Cortega is built to run alongside LiteLLM and add that layer, not to be a second system you have to operate in its place.

Copilot only governs how employees use Microsoft's own AI inside Microsoft 365 — it can't see or control traffic to ChatGPT, Claude, Gemini, or any custom models and agents your teams build and call directly. Cortega sits in front of all of it: every AI provider, every internal gateway, and every tool employees use, not just the Microsoft ones. If someone opens ChatGPT in a browser, or an engineer calls Anthropic's API directly, Copilot has no idea it happened — Cortega does.

That agreement only controls what OpenAI does with your data after it gets to them — it doesn't stop sensitive data from leaving your network in the first place, control who's allowed to send it, or do anything once your team also starts using Claude, Gemini, a local model, or an internal agent. Cortega checks and enforces policy before the request goes out: blocking things like Social Security numbers or medical records from leaving, requiring a person's approval for high-risk actions, and keeping a record of what happened — the same way, across every model and provider, not just the one you have a training opt-out with.

Traditional DLP scans email, file transfers, and general network traffic for sensitive data using pattern-matching rules — built for a world of attachments and uploads, not prompts and agents. Cortega applies that same kind of rule-based sensitive-data detection, but directly inside the AI request path: the prompt someone sends, the model's answer, and every MCP tool call an agent makes in between, not just a file being emailed or uploaded. That's the gap a general DLP tool has no visibility into at all.

Cortega also does things DLP was never built to do: controlling which AI models different people are allowed to use, catching prompt injection and SQL injection in AI traffic specifically, requiring a person's approval for high-risk agent actions before they execute, and keeping an audit record built around AI-specific detail — identity, policy decision, model used, data involved, outcome.

If you already run DLP for email and endpoints, keep it. Cortega isn't a replacement for that; it covers the AI request path specifically.

Barely. In independent testing, one 8 vCPU gateway handled more than 52,000 requests per second with a 100% success rate, adding about 5 milliseconds of typical delay at that load. Streaming responses still show up word by word — Cortega doesn't hold them back to check them first. Full test details are on the performance statistics page.

Cortega runs as many small gateways spread across your network, managed together as one system — not one giant server everything has to pass through. As you grow, you add more gateways: the cost per request stays about the same, you get backup if one gateway goes down, and you can place a gateway closer to wherever your traffic is coming from, without redesigning anything.

Because no single model or provider is the right answer forever. Providers have outages, rate limits, and price changes of their own, and different models are better suited to different tasks. If every application calls one provider directly, a single outage or deprecation becomes your outage.

Cortega lets admins configure routing across providers and models, with automatic failover and configurable failover sensitivity, from one place — so a vendor problem doesn't take your AI features down with it, and switching or adding a model doesn't mean rewriting every application that calls it.

Not sure which models belong in that rotation? The free Model Recommender ranks models against your actual workload, scale, and priorities before you configure routing.

Yes — Foundation is free: one gateway, one server, up to 8 vCPUs, standard guardrails, plus reporting, usage insights, and cost tracking. It's built for a real technical trial, not just a demo sandbox.

Foundation is free to download from our GitHub releases page and runs with Docker Compose on infrastructure you control. You'll need Docker Engine and Compose v2, roughly 2 or more vCPUs and 4 or more GB of RAM to start, and outbound network access to pull the container images.

Our download page has the full system requirements and the three supported install paths: a Docker host you already manage, a scripted AWS install, or Terraform for AWS, Azure, or GCP.

Every design partner and integration partner who has installed Cortega says the same thing: it's straightforward and painless, typically under 15 minutes from pulling the images to a running gateway. There's no source build and no bespoke provisioning — Cortega ships as prebuilt Docker images, configured through a single runtime .env file.

See the download page for the full pull-configure-run walkthrough.

Guardrails are standard on every tier, including Foundation. PRO adds unlimited gateways, advanced insights, spending limits, benchmarks, and standard support. Both PRO and Enterprise are priced to your deployment — talk to our team for a quote. Enterprise adds single sign-on (SSO), a full audit log, GDPR/HIPAA readiness, help with certification, fully disconnected (air-gapped) deployment, and a control plane that spans multiple regions, with production or dedicated support depending on the size of the deal. See the full pricing and feature comparison for every tier, side by side.

Each call that passes through the Cortega gateway counts as one request — one exchange with a model, one step an agent takes, or one tool call. Foundation gives you a full gateway to test with, and PRO and Enterprise aren't priced by how much traffic you send either — what you pay for past Foundation is which controls you need, not how many requests you make.