Quick answer

Measured 2026-08-19 on Windows 11: eleven common MCP servers hold 178-422MB of RAM each (median 188MB), start in 0.5-1.9 seconds warm, and a concurrent five-server stack totals 961.6MB — confirmed by two independent measurement paths. Runtime floors: 52MB bare Node, 15MB bare Python. Published as open CC BY 4.0 CSV.

Dataset ID
techfuelhq-mcp-server-bench
Version
1.1.0
Rows
11
Format
JSON, CSV
Published
Data updated
Page updated
Contents

Provenance

  • Aug 20, 2026Published
  • Nov 18, 2026Last substantive update

The short version

Versions tested: the eleven packages as npx -y and uvx fetched them on August 19, 2026, reporting the versions in the table below, over MCP protocol 2024-11-05 on one Windows 11 machine. The November time-server runs used mcp-server-time 2026.8.18 with MCP SDK 1.29.0 on Python 3.14.6. A newer release of any of these servers can use more or less memory.

In August 2026 I benchmarked eleven common MCP servers for what they cost to keep resident. The answer: 178-422MB of RAM each (median 188MB), ready in 0.5-1.9 seconds warm, with tool counts from 1 to 26. A realistic concurrent five-server stack totals 961.6MB, and two independent measurement paths confirmed that figure to the decimal. Tool count barely predicts weight. The runtime and framework dominate.

Download: mcp-server-bench-2026-08.csv. CC BY 4.0, per-iteration arrays included.

What does one installed time-server call cost?

November 4, 2026: mcp-server-time 2026.8.18, MCP SDK 1.29.0, Python 3.14.6 on Windows. Each run used one discarded warmup and five fresh processes. Each process listed tools and verified a conversion from UTC 12:34 to UTC 12:34. The first run started at 16:45 UTC. We repeated it at 16:57 after formatting the downloadable probe changed its file hash. Both medians are shown separately; minimum and maximum are from the repeat.

Operation or sampleFirst medianRepeat minRepeat medianRepeat max
Fresh process to ready (ms)1365.7891114.2661151.7991475.933
List tools (ms)1.4490.9311.2821.514
Convert UTC 12:34 to UTC (ms)6.1553.6685.7476.549
Process-tree RSS after listing (MiB)72.35572.21572.32072.520
Process-tree RSS after conversion (MiB)72.50072.36372.51272.621

RSS is the process tree's resident working set at two specific points, not peak or incremental allocation. This direct interpreter run does not use the August uvx wrapper and cannot establish a speedup over that cohort.

Reproduce this comparison

To recalculate the published table offline, save the Python script as reproduce.py in an empty folder. Inside that folder, create any subfolders shown below and save each input at its displayed path.

From that folder, run this command with Python 3.10 or later. It prints the rows and input hashes; it does not run the hardware or start a server.

python reproduce.py --data-root . --product mcp-server-bench

Compare the output with this product’s entry in the published JSON and input hashes.

To collect a new measurement: Use the linked probe with a separate installed interpreter. Keep interpreter, package and SDK versions, transport, startup path, call arguments and repetition count with the result.

The measured table

ServerVersionReadyRSSTools
GitHub (legacy npm)0.6.2547ms178MB26
Playwright1.63.0-alpha-2026-08-05618ms227MB24
Time (Python)1.29.0¹619ms181MB2
Sequential-thinking0.2.0666ms186MB1
Memory0.6.3667ms187MB9
Filesystem0.2.0692ms188MB14
Git (Python)1.29.0¹699ms185MB12
Fetch (Python)1.29.0¹717ms220MB1
Context74.0.2815ms214MB2
Desktop Commander0.2.471,705ms275MB26
Qdrant (Python, 3.12 pin)1.29.0¹1,881ms422MB2

¹ The four Python servers self-report 1.29.0 via serverInfo. That is a framework version string rather than each package’s own release number.

Controls on the same box: bare Node resident loop 52.1MB, bare Python loop 15.2MB. Five-server concurrent stack components: memory 189.8, filesystem 187.1, fetch 221.7, git 184.5, GitHub 178.5. Total: 961.6MB.

Method

Each iteration spawns the server fresh (npx or uvx, warm package cache after one discarded warmup), completes the MCP initialize handshake over stdio, requests tools/list, then samples RSS across the full descendant process tree. I report medians of five iterations, three for the two servers benched later the same evening, marked in the CSV. The August bench script is not published; the probe for the November time-server runs is linked from the comparison above. For the decision layer, meaning which servers earn their memory and which to skip, read the companion article.

Known limitations, named: one Windows 11 machine, warm caches, stdio transport only, readiness and listing timed rather than per-tool-call latency under load. To budget the rest of a homelab box, the Docker RAM planner sizes common self-hosted containers, and the self-hosted AI stack guide covers the hardware for running models locally. The GitHub entry is the legacy npm package; GitHub’s current first-party server ships as a Go binary and hosted remote. Both November time-server runs use the same directly installed interpreter and workload. Neither is pooled with the August CSV.

Frequently asked questions

How much RAM does an MCP server use?
In the August 19 package-launcher dataset, 178-422MB resident per server with a median of 188MB, measured across 11 common servers on Windows 11 with warm package caches. The floor is the runtime rather than the server code: a bare Node process idles at 52MB and bare Python at 15MB on the same machine.
How was the five-server stack measured?
Five servers (memory, filesystem, fetch, git, GitHub) were spawned concurrently, initialized over stdio JSON-RPC, and their full process trees summed: 961.6MB by psutil, and an independent PowerShell WorkingSet64 walk over the same trees returned 961.6MB — exact agreement between two measurement paths. All five were ready in 803ms.
Why does the dataset sum the whole process tree?
Because npx and uvx are launchers: the parent process holds almost nothing while a child holds the real footprint. Measured directly, the npx parent alone reports 10.3MB while its full tree holds 189.3MB — an 18x undercount. Every RSS figure in the CSV is the descendant-tree sum, sampled 0.4 seconds after tools/list returned.
What are the measurement conditions?
For the August CSV: fresh process per iteration, five iterations per server (three for the two late additions, marked in the CSV’s iterations column), one discarded warmup spawn so npm and uv caches are warm, MCP protocol 2024-11-05 over stdio, and medians reported. Single Windows 11 machine — Linux numbers will differ, likely with a lower Node floor.

Updates and corrections

Last updated
Methodology
See our methodology for research and review standards.
Update log
  • 2026-11-18 — Added the versions tested near the top and said where the November probe is. No measurement changed.
  • 2026-11-04 — Added two five-process runs of the directly installed time server, kept apart from the August CSV, with both receipts and the probe.
  • 2026-08-20 — v1.0.0 first publication: 11 servers, RAM summed over each server’s whole process tree, and a five-server stack checked two ways at 961.6 MB.
Corrections
Spotted an error? Email a correction with the affected line and what you found. The draft includes this page’s address. Verified changes are recorded here. Contact details.

About the author

I am Lowell K. Wood IV. I research, write and maintain everything on TechFuelHQ from St. Louis County, Missouri.