Skip to main content
Your agents need two things: models to reason with, and tools to act with. Left unmanaged, every agent holds its own provider keys, its own MCP server configs, and its own log — or none at all. Atlan puts one front door in front of both. Agents connect once with one Atlan key, and Atlan handles credentials, access, and the audit trail. There are two gateways, together called AI Gateway. Use one, or both. Each handles only the calls a client sends to it: an agent that calls its provider or tools directly is not routed through Atlan Trident. Model Gateway handles model calls. Your agents call Atlan’s endpoint instead of a provider directly — provider keys stay in Atlan, every call is logged with its cost, and you can fall back across providers without changing code. MCP Gateway handles tool calls. One MCP URL gives every agent and coding tool access to Atlan’s tools and your team’s MCP servers, with each person signed in as themselves and a record of every call.

Where does it run?

Atlan hosts both gateways. A client reaches them over HTTPS with an Atlan sign-in or key: You don’t install or run either one. A call goes through a gateway because the client’s configuration names that endpoint. A client you haven’t changed keeps calling its provider or its MCP server directly, and the gateway takes no part in those calls.

Do I have to send everything through it?

No. You choose per client. A team can use MCP Gateway for tools and keep calling its model provider directly, or the other way round. What a direct call goes without is what the gateway adds: the permission check, the stored credential, the limits, and the record of the call.

I already have my own cloud or warehouse gateway

Keep it. The two can serve different calls, and one client can hold both connections. Three things are specific to Atlan Trident’s gateways:
  • Access follows your workspaces. MCP Gateway checks the caller’s workspace permission on every tool call. See MCP permissions.
  • Calls are tied to who made them. A tool call is recorded in a session with the person or agent, the tool, the reason given, and the outcome. A model call is recorded with who made it, the model, tokens, and cost. Neither record holds the arguments, the results, or the prompt.
  • Credentials stay at the gateway. Your provider keys and your MCP servers’ credentials are stored in Atlan Trident. Clients hold only an Atlan sign-in or key.
A gateway you already run can also sit behind this one. Model Gateway connects to an OpenAI-compatible endpoint or a LiteLLM proxy as a provider; see Supported model providers. MCP Gateway registers an MCP server it can reach over Streamable HTTP; see Supported features and limits.

Does Atlan Trident need its own connection to my MCP servers?

Yes, for every server you want agents to reach through MCP Gateway. An organization admin registers the server’s URL once. After that the gateway calls the server itself each time an agent runs one of its tools. The agent’s client connects only to the gateway. That has two consequences:
  • The server must be reachable from Atlan. Private network addresses are refused, so a server that only answers inside your network can’t be registered.
  • The gateway presents the server’s credential, not the caller’s Atlan token. That is either one shared credential, or each person’s own account when the server signs people in individually.
A server you don’t register stays outside the gateway. A client can still connect to it directly, next to its connection to Atlan Trident. See How agents reach your servers.

What happens when calls hit limits?

The gateway refuses the call and says how long to wait. Both gateways limit how fast an organization can call. Model Gateway also enforces spending and token limits. For the limits and what a refusal carries, see Rate limits.

Get started

MCP Gateway

ToolsOne URL for every tool your team has connected. Add it to Claude Code, Cursor, or your own agent.

Model Gateway

ModelsOne endpoint for every model your organization has enabled. Use your existing SDK.

Review model usage

UsageSee what model calls cost and who made them, or look up a single call.