Skip to main content
MCP Gateway gives your agents one MCP endpoint, https://api.atlan.com/mcp, for Atlan’s tools and your team’s MCP servers. Add a server once. Connected clients discover and call its tools with each caller’s permissions.
Your clients, such as Claude Code, Codex, Cursor and VS Code, agent frameworks, and your own agents, connect to one endpoint, https://api.atlan.com/mcp. Atlan MCP Gateway provides OAuth 2.1 sign-in, one catalog with namespaced tools, a workspace permission check on every call, credentials held at the gateway, session and call audit, and rate limits and timeouts. Behind it are Atlan platform tools and your registered MCP servers.Your clients, such as Claude Code, Codex, Cursor and VS Code, agent frameworks, and your own agents, connect to one endpoint, https://api.atlan.com/mcp. Atlan MCP Gateway provides OAuth 2.1 sign-in, one catalog with namespaced tools, a workspace permission check on every call, credentials held at the gateway, session and call audit, and rate limits and timeouts. Behind it are Atlan platform tools and your registered MCP servers.

​
What you can do

Connect any MCP client

Add one URL to Claude Code, Cursor, VS Code, or Codex.

Add your MCP servers

Put Linear, Notion, Snowflake, or your own server behind it.

Find the right tool

Search every server’s tools by meaning, then call one.

Control who runs what

Workspace permissions decide who sees and runs each tool.

Audit agent tool calls

See who called which tool, why, and how it ended.

​
Get started

Connect client to MCP Gateway

Add the gateway to your client, sign in, and make your first audited call.

​
How it works

  1. Your client connects and signs in. It points at https://api.atlan.com/mcp and signs in with OAuth, or sends a bearer token. The gateway accepts the current stateless MCP protocol and the earlier initialize handshake on the same URL.
  2. You get one catalog. tools/list combines Atlan’s tools with the tools of every enabled server you can see. Each name’s prefix says where the tool runs, so names never collide.
  3. Your agent opens a session. It calls mcp__initialise once, then passes the returned session_id and a short rationale on every later call.
  4. The gateway checks and forwards each call. It confirms you may run the tool, attaches the server’s credential, forwards the call, and records the outcome.

​
Key concepts

​
Common tasks

​
Availability and limits

  • Clients and servers must speak Streamable HTTP. stdio servers and legacy HTTP+SSE servers aren’t supported.
  • A server that signs in with each person’s account shows its tools to a person only after they connect their account.
  • Access is per workspace. There are no per-tool or per-agent allowlists.
  • The audit trail records the tool, rationale, and outcome, not arguments or results. Recording is best-effort, so a call can succeed without a record.
  • Tools from your servers are proxied; their prompts and resources aren’t.

​
Reference

​
Troubleshooting

See MCP Gateway errors.