You are reading the built-in manual

These are the pages AURA carries within itself: inside the application, the ? key opens the one for the screen you are on. They are published here exactly as they are, so they can be read without installing anything.

The inventory of MCP servers — the connectors that give Claude outside tools: Figma, Gmail, a database, an internal service.

Three different sources feed this screen, and each has its own write regime. That is the key to reading the page.

1. Connected · claude.ai

Read from ~/.claude/mcp-needs-auth-cache.json, the cache Claude Code keeps for servers authenticated through your account. I show their name and their identifier, nothing more: their connection is managed in claude.ai and in Claude Code, never here. Cache missing or unreadable ⇒ empty list, no error.

2. Configured · file

Read from ~/.claude.json, and there are two origins inside that same file:

• global scope — the root mcpServers block. These are the only ones I modify.
• project · <name> scope — the mcpServers block of a projects entry, that is, the servers a working folder declares. Read-only.

Each row states the transport — the type key if it is set, otherwise inferred: a command gives stdio, a url gives http — and the useful detail: the full command with its arguments, or the URL.

Adding, editing, deleting a global server

The form asks for a name, a transport, then the command and its arguments (one per line) or the URL. When editing, the name is locked — renaming means deleting and creating again.

Two guarantees are worth knowing:

• Keys the form does not show are preserved. A server carrying an env block, for instance, keeps it intact after an edit: the form starts from the existing configuration and overwrites only what it manages.
• Changing transport clears the other one. Switching to stdio removes url; switching to http removes command and args. No contradictory leftover survives.

Writing into ~/.claude.json

That file lives outside the ~/.claude folder and holds far more than your MCP servers. So I go about it with particular care, different from the rest of the application:

1. The diff is narrowed to the block at hand. The preview compares the old and the new mcpServers, not the thousands of lines of the file — otherwise the change would be invisible.
2. Only the mcpServers key is rewritten. Every other key is copied across as it is.
3. Concurrency is guarded by a fingerprint of the whole file. If ~/.claude.json changed — even somewhere other than mcpServers — between the preview and the confirmation, the write is refused. Reload and start again.
4. The complete file is backed up before the rewrite, in my own backup tree.

3. MCP settings

Those three keys live in settings.json and are edited in the last section, under the usual write contract (diff, backup, refusal if the file moved).

• enableAllProjectMcpServers — Auto-approved / To confirm. Since “to confirm” is the default behaviour, it is written by the absence of the key; only “auto-approved” sets true.
• enabledMcpjsonServers — the list of those approved by name.
• disabledMcpjsonServers — the list of those refused.

These three keys apply to the servers declared by a repository .mcp.json, which Claude Code offers to approve when the project opens.

Diagnosing

• A server absent from the whole screen is not configured.
• A server that is present but whose tools never appear in a session has not finished authenticating: the session will tell you on the first call.
• A project server you would like to change cannot be changed here: it belongs to the declaration of its working folder.