Connect Claude to Your Account

What the Connector Is

VoxelBench hosts an MCP server at https://voxelbench.com/api/v1/mcp: there is nothing to install. It acts for your account through the API, with the permissions you grant it: the same data, the same quotas, the same refusals. The connector decides nothing on its own; the API does.

The Connect an agent page gives the address and the commands to copy, and lists every tool.

Claude Web and Desktop

1. In Claude, open Settings → Connectors → Add custom connector, and paste https://voxelbench.com/api/v1/mcp.
2. Claude opens a VoxelBench page. Sign in if you are not signed in yet.
3. The consent screen shows what Claude asks for (see below). Choose the permissions, then click Authorize.

That is all: no token to copy anywhere. Claude renews its access on its own every hour. A connector left unused for 60 days has to be authorized again.

The Consent Screen

The page, Authorize this connector, shows:

• the application's name, either as the application declares it or with the domain that publishes its identity;
• the address it sends you back to: check that it belongs to the application you meant to connect;
• the permissions, in two groups, Read and Act on your behalf. Those the application asks for are ticked in advance. The second group is optional: tick only what you want this connector to do.

Authorize grants exactly what is ticked, and at least one permission must be. Refuse grants nothing.

A connector takes one of your plan's API token slots. When they are all used, the page says No token slot available and offers to manage your tokens or see the plans. Connecting the same application again reuses its slot.

Claude Code

Add the server, with no token:

claude mcp add --transport http voxelbench https://voxelbench.com/api/v1/mcp

Then type /mcp in Claude Code, pick voxelbench and choose Authenticate. Your browser opens the same consent page as Claude Web, and Claude Code renews its access on its own.

For scripts and CI, where nobody is there to click, give it an API token instead:

claude mcp add --transport http voxelbench https://voxelbench.com/api/v1/mcp \
  --header "Authorization: Bearer vb_your_token"

Other MCP Clients

The endpoint speaks MCP over Streamable HTTP. It needs a token on every request, even to read public reports (the API serves those without one):

• a client that supports MCP authorization finds the way by itself: it registers, opens the VoxelBench consent page and receives a token limited to what you approved;
• any other client sends an API token in an Authorization: Bearer vb_… header.

{
  "mcpServers": {
    "voxelbench": {
      "type": "http",
      "url": "https://voxelbench.com/api/v1/mcp",
      "headers": { "Authorization": "Bearer vb_your_token" }
    }
  }
}

What the Tools Can Do

Read, with the read permissions:

• reports: list, read, diagnosis, tuning advice, comparison of two reports, before/after verdict;
• single-test results;
• servers and groups, and whether a server's performance has dropped;
• monitoring: live status, overview of all your servers, metrics, events, alert rules, alerts, maintenance windows;
• performance profiles and memory reports, and their comparisons;
• auto-bench targets and their runs;
• the hosting offers catalogue.

Act, each with its own permission:

• link a new server with the 8-character code the plugin prints: Claude can only link a server whose code you gave it;
• rename a server, change its visibility, create groups and sort servers into them;
• change a report's visibility, description and private notes;
• create, edit, pause and mute alert rules, send a test notification, acknowledge alerts, start and end maintenance;
• launch, cancel and reschedule auto-bench runs on targets you already have. A launch connects a bot to your real server for about ten minutes.

Comparisons and diagnoses are computed by VoxelBench and handed over already decided: the model quotes them, it does not recompute them. Each tool also tells your client whether it only reads, so a client can run the reading tools without asking you and ask before the others.

The server also offers ready-made walkthroughs (MCP prompts): diagnose_report, compare_reports, server_health, investigate_incident, weekly_incidents, and benchmark_now, which runs an existing target once you confirm.

What They Never Do

Whatever the model is asked, these stay closed, because the API refuses them:

• delete anything (report, group, alert rule, target);
• unlink a server, or link again a server that is already linked;
• create an auto-bench target, which engages a Microsoft account you lend;
• set a notification address or a webhook, which would decide where your data goes;
• create a token, or touch billing and the account.

The tools never see your servers' link tokens, your notification addresses or the Microsoft accounts you lent, nor anything private of another account. The tool descriptions tell the model to treat the text people write into reports and events as data to quote, never as instructions to follow.

Permissions and Quotas

Change the permissions at any time: Settings → Security → API tokens. The connector's row carries the Connector badge; its Permissions button lets you tick Read and Act on your behalf permissions. The change applies on the connector's next call, with no need to reconnect.

When a tool needs a permission you did not grant, the call is refused: add the permission there. When the refusal is plan_required, it is your plan that lacks the feature (auto-bench, monitoring), not the connector.

Quotas are those of the API (see API and Tokens):

• a connector counts as one of your plan's tokens;
• each tool call is one API call, taken from your account's daily quota (10 on Free, 50 on Pro, 150 on hosting provider accounts, 500 on Enterprise), reads and writes alike; connecting and listing the tools cost nothing;
• every result tells the model how many calls are left, so it sees the budget shrink instead of hitting the limit;
• launching auto-bench runs also uses your daily run quota.

A resource you may not see answers "not found", exactly like one that does not exist: the model must not conclude that it does not exist.

Revoking Access

Go to Settings → Security → API tokens and click Revoke on the connector's row. Access stops at once, renewal included: to use the connector again, you will have to authorize it again.

Revoking on VoxelBench is what ends the access. Until you do, the authorization stays valid, and it lapses on its own after 60 days without use.