Connect Claude to your Minecraft servers
VoxelBench hosts an MCP server: connect Claude once, then ask about your benchmarks, your monitoring, your profiles and memory reports in plain language. VoxelBench computes the answers, Claude quotes them. Nothing to install, no token to copy.
Ask about your servers in plain language: VoxelBench measures, Claude reads the results with you
VoxelBench hosts an MCP server at https://voxelbench.com/api/v1/mcp. MCP, the Model Context Protocol, is how an AI assistant such as Claude calls outside tools. Once it is connected to your account, Claude can read your benchmark reports, your servers and their monitoring, and the performance profiles and memory reports your plugin sent, then answer questions such as "what limits this server?" or "which plugin holds the most memory?" with the figures VoxelBench computed.
There is nothing to install, and no token to copy.
What you need
- A voxelbench.com account with at least one server linked to it (
/bench link, see Account Linking). Without a linked server, there is little of yours for Claude to read. - Claude on the web, the desktop app or Claude Code, or any other client that speaks MCP over Streamable HTTP.
Claude on the web or the desktop app
- In Claude, open Settings → Connectors → Add custom connector.
- Paste
https://voxelbench.com/api/v1/mcp. - Claude opens a VoxelBench page. Sign in if you are not signed in yet.
- The Authorize this connector page shows what Claude asks for, in two groups: Read and Act on your behalf. Tick what you want, then click Authorize.
That is all. Claude renews its access by itself every hour, and a connector left unused for 60 days has to be authorized again.
Before you click Authorize, check the address the page sends you back to: it must belong to the application you meant to connect.
Which permissions to tick
Start with reading only. The Read group holds:
| Permission | What Claude can then read |
|---|---|
reports:read | Your benchmark reports, private ones included |
servers:read | The servers linked to your account |
monitoring:read | Everything your servers' monitoring records: live status, metrics and events, performance profiles, memory reports, maintenance windows, alert rules and alerts |
unit-tests:read | Your single-test results |
auto-bench:read | Your auto-bench targets and their runs |
offers:read | The public hosting catalogue |
The Act on your behalf group is optional: renaming servers and sorting them into groups, changing a report's visibility, description and private notes, managing alert rules and maintenance windows, launching runs on auto-bench targets you already have, and linking a new server with the code the plugin prints. Tick only what you want Claude to do.
You can change these permissions at any time, without reconnecting: Settings → Security → API tokens, then the Permissions button on the connector's row.
Claude Code
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.
For a script or a CI job, where nobody is there to click, create an API token and pass it as a header instead:
claude mcp add --transport http voxelbench https://voxelbench.com/api/v1/mcp \
--header "Authorization: Bearer vb_your_token"
Other MCP clients
A client that supports MCP authorization finds its 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 on every request:
{
"mcpServers": {
"voxelbench": {
"type": "http",
"url": "https://voxelbench.com/api/v1/mcp",
"headers": { "Authorization": "Bearer vb_your_token" }
}
}
}
The Connect an agent page gives the address and the commands to copy, and lists every tool.
What to ask
Ask in your own words. A few questions to start with:
- "How are my servers doing?"
- "What limits my server, according to its last benchmark?"
- "Compare my last two benchmarks: did the Paper update help?"
- "What happened on my server last night around 3 am?" (needs monitoring)
- "Which plugin holds the most memory on my server?" (see Find what fills your server's memory, with Claude)
- "What ran on the tick thread during the last lag?" (see Find what lags your server, with Claude)
Clients that support MCP prompts also get ready-made walkthroughs: diagnose_report, compare_reports, server_health, investigate_incident, weekly_incidents, and benchmark_now, which runs an existing auto-bench target once you confirm. In Claude Code they show up as slash commands, such as /mcp__voxelbench__server_health.
Monitoring data (live status, metrics, events, alerts) needs a plan that includes monitoring. Benchmark reports, performance profiles and memory reports work on every plan.
How the answers stay honest
- VoxelBench decides, Claude quotes. Diagnoses, comparisons and verdicts are computed by VoxelBench and handed to the model already decided. The tools tell it to quote them, not to recompute them.
- Reads and actions are labelled. Each tool tells your client whether it only reads, so a client can run the reading tools without asking and ask you before the others.
- Text is data. Report descriptions, event messages, plugin and class names come from people and from code. The tools tell the model to treat them as data to quote, never as instructions to follow.
- Some doors stay shut, whatever the model is asked, because the API refuses them: deleting anything, unlinking a server, creating an auto-bench target, setting a notification address or a webhook, creating a token, touching billing or the account.
- The tools never see your servers' link tokens, your notification addresses, the Microsoft accounts you lent for auto-bench, nor anything private of another account.
Quotas
The connector acts through the VoxelBench API, with the same data, the same quotas and the same refusals:
- 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. Connecting and listing the tools cost nothing;
- every result tells the model how many calls are left, so it sees its budget shrink instead of hitting the limit;
- a connector takes one of your plan's API token slots, and connecting the same application again reuses its slot.
A question usually takes a few calls: one to find the right report, one to read it, sometimes one more to compare it with another.
Revoking access
Settings → Security → API tokens, then Revoke on the connector's row. Access stops at once, renewal included. Until you revoke it, the authorization stays valid, and it lapses by itself after 60 days without use.
The details are in Connect Claude to Your Account.