Client connection settings

The settings an MCP client uses to connect to Atlan MCP Gateway: the endpoint, the tool mode, how the client signs in, and which tools it approves. Use this page to look up values. To connect a client, see Connect client to MCP Gateway.

Endpoint

Setting	Value
URL	`https://api.atlan.com/mcp`
Transport	Streamable HTTP. stdio and legacy HTTP+SSE are not supported.

Tool mode

The tools parameter on the URL sets which tools tools/list returns. Both modes reach the same tools with the same permissions. For why search mode is the default, see How search mode works.

Value	`tools/list` returns	Session opens with	Client approves	Status
Omitted, or `search`	Five tools: `mcp__search_tools`, `mcp__describe_tools`, `mcp__call_read_tool`, `mcp__call_write_tool`, and `mcp__call_tool`	The first search or runner call	Each runner	Default
`full`	Every tool you can reach, plus `mcp__initialise`	`mcp__initialise`	Each tool by name	Will be deprecated
Any other value	Same as search mode	Same as search mode	Same as search mode	—

Example: full mode in Claude Code and in a JSON configuration.

Claude Code
JSON configuration

claude mcp add --transport http atlan "https://api.atlan.com/mcp?tools=full"

Quote the URL so your shell keeps the ?. If you already added atlan, run claude mcp remove atlan first.

{
  "mcpServers": {
    "atlan": { "type": "http", "url": "https://api.atlan.com/mcp?tools=full" }
  }
}

Full mode will be deprecated later. Plan to move your clients to search mode.

Changes from full mode

Search mode is the default on the plain URL, so a client that connected before now sees five tools instead of every tool.

Area	Full mode	Search mode	What to update
Approvals, permission rules, and hooks	Match each tool’s name	Match the runner; a rule for a tool’s own name no longer matches	Replace them with runner approvals, or keep `?tools=full` until you do
Session	`mcp__initialise` returns the `session_id`	The first search or call opens the session. A `session_id` you pass must be one you opened.	Nothing, if the agent follows the tool descriptions
Direct calls by tool name	Listed in `tools/list`	Still work, though `tools/list` does not show them. Without a `session_id`, the first call opens a session from its `rationale`.	Nothing
`mcp__initialise`	Opens the session	Still opens a session, for a client that already calls it. Its earlier name, `registry__initialise`, does too, but is not listed.	Nothing
Client’s tool list	Every tool	Five tools	Reconnect the client, for example with `/mcp` in Claude Code

Method	How the client sends it	Acts as	Best for
OAuth sign-in	The client follows the gateway’s `401` to `https://api.atlan.com/.well-known/oauth-protected-resource/mcp` (RFC 9728) and opens a browser. No token goes in the configuration.	You	People using an IDE or desktop client
Bearer token	`Authorization: Bearer <token>` on each request. See Authentication for supported credential paths.	The token’s identity	CI jobs, services, and headless agents

Clients that support MCP OAuth 2.1 need no extra configuration for OAuth sign-in. Load a bearer token from a secret store, never from a committed file.

An Atlan API key sent as a bearer token acts as the key, not as you: it reaches what the organization shares at its top level and the workspaces the key is added to.

Tool approvals

In search mode your client approves the runner a call goes through, not the tool it runs. Each approval covers the runner’s tier.

Tool	Always allow covers	Name in Claude Code, for a server named `atlan`	Recommended
`mcp__search_tools`	Finding tools. Read only.	`mcp__atlan__mcp__search_tools`	Allow
`mcp__describe_tools`	Reading tool schemas. Read only.	`mcp__atlan__mcp__describe_tools`	Allow
`mcp__call_read_tool`	Tools that only read	`mcp__atlan__mcp__call_read_tool`	Allow
`mcp__call_write_tool`	Tools that write but are declared not destructive, and reads	`mcp__atlan__mcp__call_write_tool`	Allow
`mcp__call_tool`	Every tool you can reach through the gateway	`mcp__atlan__mcp__call_tool`	Leave to ask

A tool’s tier comes from its own labels. A tool labelled read-only that writes runs through mcp__call_read_tool, so allowing that runner allows it too.

Example: allow reads and non-destructive writes, and keep being asked before anything destructive.

Claude Code
Claude Code in CI
Codex
Other clients

In .claude/settings.json:

{
  "permissions": {
    "allow": [
      "mcp__atlan__mcp__search_tools",
      "mcp__atlan__mcp__describe_tools",
      "mcp__atlan__mcp__call_read_tool",
      "mcp__atlan__mcp__call_write_tool"
    ]
  }
}

A script or CI job must allow each search mode tool it uses by its full name, or the search or the call is denied:

claude -p "<task>" --allowedTools "mcp__atlan__mcp__search_tools,mcp__atlan__mcp__describe_tools,mcp__atlan__mcp__call_read_tool,mcp__atlan__mcp__call_write_tool"

Codex asks before every call through mcp__call_write_tool or mcp__call_tool, including a read run through mcp__call_tool, and refuses such a call in codex exec. See Connect Codex.

Allow the same four tools under the names your client gives them, and leave mcp__call_tool to ask. If your client keeps a per-agent allowlist, check arguments.name before letting a call through.

MCP Gateway

Get started

Bring in tools

Audit tool use

References

Concepts

Troubleshoot

Model Gateway

Get started

Configure

Monitor

References

Troubleshoot

Client connection settings

Endpoint

Tool mode

Changes from full mode

Tool approvals

See also

​Endpoint

​Tool mode

​Changes from full mode

​Sign-in methods

​Tool approvals

​See also

Endpoint

Tool mode

Changes from full mode

Sign-in methods

Tool approvals

See also