Claude Code MCP prompts again? Check permission and server scope

If Claude Code asks again before using an MCP tool in a new session, first identify which approval is being requested: permission to call a tool, approval of a project MCP server, or a restriction from another rule or policy. Current Anthropic docs put user permission rules in settings files and describe server approvals separately. Issue #26765 is a closed duplicate reporting Claude Code 2.1.47; it is one historical report, not evidence of a current general bug. Documentation checked November 3, 2026.

Is this an MCP server approval or a tool permission prompt?

Check the exact prompt and tool name. A project server declared in .mcp.json can remain pending until you trust the workspace and approve it. The CLI exposes that state through claude mcp list or claude mcp get <server>; a pending project server is shown as Pending approval. Approve only a server you recognize and intend to use.

A permission rule for an MCP tool is a different control. Anthropic’s current permission syntax uses the complete tool name such as mcp__github__search_repositories, or a server-scoped rule such as mcp__github__*. A server may connect successfully while a specific tool still prompts.

Which files should contain cross-session permission rules?

Current settings documentation lists ~/.claude/settings.json for your personal rules across projects, .claude/settings.json for shared project rules, and .claude/settings.local.json for one user in one project. The separate ~/.claude.json file stores MCP server configuration, project state, account details, and global config keys; Anthropic says Claude Code writes it for itself and you do not need to edit it. Issue #26765 reports a permission list in that file, but the current docs describe permission rules in settings JSON, so verify the supported location for your installed version before copying that older layout.

For a narrow user rule, add one known tool to ~/.claude/settings.json and preserve any existing JSON:

{
  "permissions": {
    "allow": [
      "mcp__github__search_repositories"
    ]
  }
}

Replace the example with the exact tool name shown by your MCP server. The wildcard form mcp__github__* allows all tools from that named server, so prefer the single-tool rule unless the broader access is intended. Do not paste tokens or other credentials into settings examples or issue reports.

How do I confirm the right settings loaded?

1. In the session that prompts, run /status and inspect the Setting sources line. It confirms which settings files loaded, but does not show which file supplied an individual rule.
2. Run claude doctor to see settings entries Claude Code rejected, including malformed permission rules. Check its output for a typo or invalid JSON rather than assuming a saved entry was accepted.
3. Check whether the project folder is trusted. Current docs say project permissions.allow rules wait for folder trust; a shared or local project rule may not apply before that step.
4. Review every applicable settings scope for a matching ask or deny rule and check whether organization-managed rules restrict user rules. Permission lists generally merge; a saved allow does not automatically cancel an ask or deny rule.
5. Check the exact MCP tool name and rule syntax. The server segment in an allow wildcard must name one configured server; a generic mcp__* allow rule is not accepted as an auto-approval pattern.
6. Start a fresh session and repeat one harmless tool call. Record the Claude Code version, prompt text, server scope, tool name, and settings sources. Compare only one changed setting at a time.

Can an allow rule still leave a prompt?

Yes, depending on what is prompting. Anthropic documents that MCP tools marked requiresUserInteraction still prompt even if allowed. Organization policy can also constrain rules, and a project server may still need server approval before its tools are available. Diagnose these cases separately; do not disable permission checks or broaden all MCP access just to remove a prompt.

What does issue #26765 establish?

The report says a user on Claude Code 2.1.47 saw repeated MCP tool permission prompts despite entries in ~/.claude.json. GitHub marks it closed as a duplicate. That establishes one reported symptom and version, not the cause, a confirmed regression, or current behavior. The checks above follow current Anthropic documentation and do not claim to reproduce or resolve that report on every version.

How should I report a repeatable case?

Record the version from claude --version, whether the server is local, project, or user scope, the exact tool name, the wording of the prompt, and the settings-source names shown in /status. Include only a redacted minimal rule and relevant claude doctor warnings. Remove access tokens, private endpoints, customer data, and full configuration files before posting.

This is a documentation-based diagnostic, not a runtime test of all Claude Code releases, MCP servers, or managed policies. For a browser cloud session, see why MCP allow rules may not carry over to Claude Code on the web. Related: map Claude Code MCP configuration scopes, choose Claude Code permission rule scopes, and troubleshoot MCP proxy connections.

Sources and scope

• Anthropic settings files and precedence documents settings locations, /status, claude doctor, and permission-rule merging.
• Anthropic permission configuration documents MCP tool rule syntax and cases where prompts remain.
• Anthropic MCP documentation documents server scopes, pending project approval, and MCP status commands.
• Claude Code issue #26765 is a closed duplicate with one report against version 2.1.47; it is not treated as current product guidance.