Set up MCP
Every client connects to the same URL,https://api.atlan.com/mcp, and signs in with your Atlan account through OAuth. Your client must support Streamable HTTP; stdio and legacy HTTP+SSE are not supported. Pick your client:
- Claude Code
- Codex
- Cursor
- VS Code
- Other clients
-
Add the server:
This adds
atlanfor you in the current project. Add--scope userto use it in every project. -
Start Claude Code, run
/mcp, selectatlan, and complete the sign-in in your browser. -
Run
claude mcp listand confirmatlanis connected.
List my Atlan workspaces. and find the call under Tool calls in MCP Studio. For CI jobs and headless agents that can’t open a browser, send a bearer token instead, as described in Connect client to MCP Gateway.
Tool providers compared
Each tool name says where it comes from:
The tool groups in this section follow the jobs an agent does: connect to Registry, find artifacts, manage artifacts, register and monitor agents, publish and monitor skills, audit runs, use ontologies, evaluate runs, find metrics, and register repositories.
Core offerings
All MCP tools
Look up every current first-party tool, and how the gateway catalogs registered servers’ tools.
Start work session
Open the session every later tool call must name.
Search mode tools
Find, describe, and run every tool through the five tools a default connection lists.
Track usage and audit
Review recorded calls, with their rationale and outcome.
Choose where to start
- Managing servers from an application? The MCP category of the Public API checks what a new server needs before you register it, and lists, inspects, tests, and refreshes registered servers.
- A call failed? MCP Gateway errors lists errors by symptom.
Every call runs as the signed-in identity and only reaches tools it may use. Your client’s live
mcp__describe_tools result in search mode, or tools/list result in full mode, is the source of truth for callable names and input schemas, so read each tool’s description and schema before calling it, and give an agent only the tools its workflow needs. Every tool call needs a session_id and, by default, a short rationale. In search mode, the default, the first mcp__search_tools call returns the session_id; in full mode, ?tools=full, mcp__initialise does. Recorded calls include the tool, rationale, outcome, and duration, but not arguments or output. Recording is best-effort; a call can succeed even when its audit write fails. Read a write tool’s retry and partial-success rules before calling it.See also
- Serve MCP tools: Register your own MCP servers behind the same endpoint.
App tools and downstream tools
The complete first-party catalog includes Registry, agent, skill, file, eval, ontology, metric, and repository tools. Apps with no declared MCP tools still provide their documented HTTP or product interfaces. Remote servers contribute account-specific tools. Inspect its tools withmcp__describe_tools and the server’s name in search mode, or with authenticated tools/list in full mode, after a server is connected or refreshed; the REST tools listing is a saved discovery projection and can lag the callable catalog. Skill allowed-tools metadata describes dependencies and does not register a tool.