EN
Contact sales

AI Gateway

Control data going to AI at the same checkpoint

The gateway decides on, masks and records the AI apps employees use, LLM API calls from in-house apps and developer tools, and even the MCP tools that AI calls. The value patterns you apply to databases work on AI requests as is.

  • OpenAI
  • Anthropic
  • Gemini
  • DeepSeek
  • Azure OpenAI
  • Ollama
  • vLLM
  • MCP

What is an AI gateway?

An AI gateway is a proxy that sits between an organization and AI services to inspect requests, manage keys and record usage.

BeeHind's AI Gateway uses the same gateway, the same accounts and the same audit store as DAC and Server Gateway. The SQL statements, server commands and questions to AI from the same person are found under the same name.

Three kinds of control

The right control for each way you use AI

AI apps on employee PCs reach the gateway through the desktop agent. Server applications and developer tools reach it with a virtual key.

  • AI apps employees use

    Who
    AI apps that employees sign in to with their own accounts on their PCs (for example, the Claude desktop app)
    How
    The desktop agent sends connections to the AI service through the gateway. Employees do not change any app settings.
    How the user is identified
    Signed identity assertion
    • Choose to block, mask or only record requests
    • Responses masked with the same patterns
    • Records only what is new in each request

  • LLM calls from in-house apps and developer tools

    Who
    In-house services and batch jobs that call LLM APIs; Claude Code, Codex, Gemini CLI
    How
    Point the API address at the gateway and replace the API key with a virtual key. The gateway swaps in your company's real key (organization key) when it forwards the request.
    How the user is identified
    Virtual key
    • The organization key is never handed out
    • Usage by user, service and model
    • Daily or monthly usage limits

  • In-house tools called by AI (MCP)

    Who
    Organizations that expose in-house systems as MCP servers
    How
    Register the MCP server as a service and set an allowlist per tool call. Values in arguments are checked against value patterns too.
    How the user is identified
    Signed identity assertion
    • Allow lookup tools, block delete tools
    • Different allowlists per person and group
    • Tool call history recorded by name

Masks personal data and credentials in prompts before they are sent

For each pattern, choose block, mask or detect only. Turn a new pattern on as "detect only" first to see how often it matches, then change it.

Prompt the employee sent

Can you check why this config file isn't working?
aws_access_key_id = AKIAIOSFODNN7EXAMPLE
Customer ticket: Jane Doe, 415-555-0132, SSN 123-45-6789

What the AI service received

Can you check why this config file isn't working?
aws_access_key_id = ********************
Customer ticket: Jane Doe, 415-555-****, SSN ***-**-****
  • AWS access key
  • Phone number
  • US Social Security number

Features

Features that control AI use

Policies are set per user and group in the admin console.

  • Virtual keys

    Only the gateway uses your company's real API keys; users and applications get virtual keys. Virtual keys can be revoked one at a time.

  • Model restrictions

    Block models whose token price exceeds a limit and models you have decided not to use. Set which models each user and group can use.

  • Content inspection

    Apply value patterns to prompts and streamed responses. Credential patterns such as AWS access keys, private keys and database connection strings are built in.

  • Usage limits

    Limit daily or monthly token and request counts per user and service.

  • Tool control

    Restrict MCP server tool calls to an allowlist and check arguments against value patterns.

  • Upstream format conversion

    Converts the request format when sending to cloud model services such as Amazon Bedrock and Google Vertex AI. When you move to another upstream target, application settings stay the same.

  • In-house model servers

    Register in-house model servers such as vLLM and Ollama as OpenAI-compatible services and inspect them with the same rules.

  • Config files for each tool

    The admin console generates config files for Claude Code, Codex, Gemini CLI and SDKs.

Supported AI services

As of October 2026

Content read and controlled
OpenAI-compatible APIs (OpenAI, Google Gemini, DeepSeek and others), Anthropic API, claude.ai web app, MCP servers (Streamable HTTP)
Upstream targets
Amazon Bedrock, Google Vertex AI, Claude Platform on AWS, Microsoft Foundry, Azure OpenAI
In-house model servers
vLLM, Ollama and other OpenAI-compatible servers
Developer tools
Claude Code, Codex, Gemini CLI, SDKs

Known limitations

  • For services not on the list and apps that pin certificates, the gateway cannot read the content and records only that a connection was made.
  • If PCs have a direct route to AI services, users can turn off the agent and bypass the gateway. Close that route at the firewall.
  • Usage limits are tallied every few seconds, so usage between tallies can go over the limit. They are a cost-management tool, not a security control.
  • Masking is not yet applied to results returned by MCP tools.

See BeeHind in your own environment

Tell us which databases, servers and AI services you need to protect and how your network is laid out. We will recommend a setup and a rollout order. You can start with a single server.

Contact salesFAQ