Everything (test server)
Everything is Anthropic's reference MCP server: a deliberate test rig that implements every part of the protocol at once. Its 13 tools do nothing useful on purpose. They echo strings, add two numbers, return a tiny image, and trigger a long-running operation so client authors can watch progress notifications, resource links, and annotations behave.
How we started it
npx -y @modelcontextprotocol/server-everythingUse it, but not the way you use other servers. Everything exists so you can prove your MCP client handles the awkward parts of the protocol before a real server exposes the bug for you. As a daily connection it's pure waste: ~1,913 tokens of tool schemas that resolve to echo and get-sum, plus a get-env tool that will happily read back whatever environment you handed the process. Connect it in a scratch profile, learn what you needed, disconnect.
Use it if
Skip it if
The bill for your context window
| Tool | Est. tokens | Described? | Required params |
|---|---|---|---|
| gzip-file-as-resource | ~285 | yes (247 chars) | 0 |
| get-structured-content | ~233 | yes (81 chars) | 1 |
| simulate-research-query | ~212 | yes (270 chars) | 1 |
| get-annotated-message | ~169 | yes (75 chars) | 1 |
| trigger-long-running-operation | ~147 | yes (60 chars) | 0 |
| get-resource-reference | ~139 | yes (60 chars) | 0 |
| get-resource-links | ~130 | yes (76 chars) | 0 |
| get-sum | ~114 | yes (30 chars) | 2 |
When we booted it, tools/list came back at 7,653 bytes, about 1,913 tokens at 4 chars per token, for 13 tools that produce no real output. That is exactly the median context tax of the batch this was measured in, which is a good joke on the rest of the ecosystem: a server built to demonstrate nothing costs as much as a typical working one. The weight is spread thin rather than concentrated, though gzip-file-as-resource, get-structured-content, and simulate-research-query carry the largest schemas between them. Nothing here is worth trimming, because nothing here should be loaded in a session doing real work.
Setup reality
There is nothing to configure. We started it as an unprivileged user with a clean environment and no credentials, and it came straight up over stdio with no keys, no OAuth dance, and no env vars requested. The harness scraped zero environment hints from stderr, which said only that it was starting the default stdio server. Boot took 7.5 seconds, and most of that is npx fetching the package on a cold cache rather than the server itself. It also ships HTTP and SSE modes for client authors who need to test those transports.
Questions people ask
Is the everything MCP server useful in production?
No. Every tool returns fixture data by design. Its value is testing an MCP client against a server that exercises the whole protocol, so keep it in a development profile and take it out before you ship.
How much context does the everything server use?
Its tools/list payload measured 7,653 bytes, roughly 1,913 tokens using a 4 chars per token estimate, when we measured it on 2026-08-19. That covers 13 tools, 7 resources, and 4 prompts.
Does the everything server need any API keys?
None. It booted with a clean environment and requested nothing. That is part of why it is the right first server to point a new client at when you are still unsure whether your own wiring works.
Why does it expose a get-env tool?
So client authors can confirm what environment their server processes inherit. It is a diagnostic, and also a reminder that any MCP server you spawn can read the variables you gave it. Treat that as a habit, not a quirk of this one.
Other dev tools servers we started
| Server | Vendor | Boots | Context tax |
|---|---|---|---|
| CircleCI | CircleCI | ✓ | ~12,932 tokens |
| Git | Anthropic (reference) | ✓ | ~1,477 tokens |
| GitHub (archived npm server) | Anthropic (archived) | ✓ | ~3,964 tokens |
| GitLab (archived) | Anthropic (archived) | ✗ | ~0 tokens |
How this page is made: the server is spawned as an unprivileged user with a clean environment and no credentials, then asked for its tools, resources and prompts over stdio. Token figures are estimates at four characters per token, not a tokenizer count. One run, one machine. Corrections: contact the desk.