mcp server

rigbase-data

Measured local-LLM tok/s on home hardware (CC BY 4.0) - from getrigbase.com

getrigbase/rigbase-datacommunitylistedunreviewed

Set up

Hosted by its author. Add the URL to your agent — nothing to install.

~/Library/Application Support/Claude/claude_desktop_config.json

Settings → Connectors → Add custom connector

Name: rigbase-data
URL:  https://getrigbase.com/mcp

same format Cursor, GitHub Copilot CLI, Junie, JetBrains AI Assistant, Augment (Auggie CLI), Factory Droid, Kiro, Devin CLI +12

Values in angle brackets are placeholders. Restart Claude Desktop after editing its config.Claude Desktop docs

README

synced Nov 8, 2026

RigBase data

Measured local-LLM speeds on real home hardware.

getrigbase.com · Live download · Hugging Face · Licence: CC BY 4.0

This repository mirrors the free RigBase dataset:


Builds	150
Measured results	358
Models	51
Snapshot generated	2026-10-27 (from getrigbase.com/data/)

Every number links to its original public source (llama.cpp discussions, llama-bench databases, blogs, GitHub repos, Mia's AI Lab recipes, and similar). RigBase collects and normalises these numbers; it does not re-measure them or invent estimates for the tps column.

Files

builds.csv — one row per measured result (358 rows)
builds.json — one object per build with parts + nested results
LICENSE — CC BY 4.0
CITATION.cff — machine-readable citation

Quick start

# CSV
curl -LO https://getrigbase.com/data/builds.csv

# Or clone this repo
git clone https://github.com/getrigbase/rigbase-data.git

import pandas as pd
df = pd.read_csv("builds.csv")
# Plain decode only (drop speculative / MTP / draft runs):
plain = df[df["spec_decode"].fillna("none").eq("none")]
print(plain.groupby("model")["tps"].max().sort_values(ascending=False).head(10))

Columns (CSV)

column	meaning
`build_slug`, `build_title`, `page_url`	Build id, name, and RigBase page
`gpus`, `vram_gb`, `cpu`, `ram`, `power`	Hardware as stated by the source
`model`, `quant`, `engine`	What was run, as stated by the source
`tps`	Generation (decode) tokens/sec reported by the source
`spec_decode`	`none`, or the speculative-decoding method (MTP, DFlash, …)
`prefill_tps`, `prefill_tokens`, `prefill_metric`	Prefill speed only where the source reports one; never estimated
`est_ttft_8k_s_calculated`	Calculated (not measured): 8192 / prefill_tps when prefill_tokens ≥ 512
`ttft_s_measured`, `ttft_prompt_tokens`	TTFT as measured by the source
`source`	Original public measurement URL
`source_checked_by_rigbase`	`yes` = RigBase re-read the number at the source

Full column notes: https://getrigbase.com/data/

Caveats

Numbers come from different people, engines, quants and settings. Compare rows from the same source when you compute ratios.
Filter on spec_decode == "none" for plain decode speed.
Calculator pages on the site (VRAM, KV cache) are arithmetic, not measurements. They are not in this dataset.

Browse builds: https://getrigbase.com/
Can I run it? (VRAM + KV cache calculator): https://getrigbase.com/can-it-run/
DGX Spark recipes: https://getrigbase.com/dgx-spark/
Submit a correction or build: https://getrigbase.com/submit/

Attribution

Required by CC BY 4.0:

RigBase (getrigbase.com), CC BY 4.0. Each row keeps its original source URL.

Please also keep or cite the per-row source URL when you reuse an individual number.

Updates

The live files at https://getrigbase.com/data/ are the source of truth. This GitHub mirror is refreshed when the site dataset changes. Issues and PRs that fix a sourced number (with a link) are welcome.