MCP App Store
by KeenEthics
Indahl Models icon

Indahl Models MCP Server for Claude

By Heuremen LLC

Overview

Indahl Models gives Claude read-only access to Heuremen LLC's measurements of the AI and open-source supply chain, each one a stated claim checked against the delivered files: - Licenses: the license the 500 most downloaded Hugging Face models and datasets state in their cards against the license file in the repository, and the 1,000 most downloaded PyPI projects' declared license against the one GitHub detects. - Sizes: the parameter count in a model's name against the parameters in its safetensors files. - Installs: whether the pip install command a project documents actually resolves. - Hosts and weights: every OpenRouter host of open-weight models with its stated quantization and uptime, and whether weight files behind a model name stay the same over time. - Benchmarks: MLCommons AILuminate safety results against their reports, open-model benchmark claims rerun with the published harness, and our own Faults and Guards bench. - Lifecycles: model and API shutdown dates announced by Anthropic, OpenAI, Google and Cloudflare, and every date that moved. - Research hygiene: arXiv version drift for indexed studies and link rot in arXiv code and data links. Every answer carries the share with its 95% interval, the outcome (ABOVE, BELOW, CONSISTENT or CANNOT_TELL), the counts, and examples read in full. No account needed. All data and code are public at github.com/Heuremen-LLC/indahl.ai.

How to connect

  1. Claude: Open the Indahl Models listing in the Claude connectors directory and choose Connect. Claude listing Indahl Models exposes 15 tools here.

Tools

ailuminate

Claude
AILUMINATE: the MLCommons AILuminate Safety results against their own reports, with every read kept.

all_tests_side_by_side

Claude
The headline test of every live page in one table: what was compared, how many failed, the share with its 95% interval, and the outcome each page's own rule gave, ordered by outcome. Also a tally by outcome of the separate items on the pages that report many.

announced_shutdown_dates

Claude
Shutdown dates published by Anthropic, OpenAI, Google and Cloudflare for models and APIs, read on a schedule: what ends in the next 60 days with the notice given, notice and lifespan by vendor, every date that moved, was added or disappeared since the first reading, and the test of whether stated dates hold with its outcome.

arxiv_version_drift

Claude
For every arXiv study that QIA and IRIS index: whether a later arXiv version exists and whether the revision changed what we read (abstract numbers, title or wording, withdrawal), by class and by month indexed, with the studies whose abstract numbers changed and the numbers each version states. Descriptive, read every 7 days.

bench_results

Claude
Standards bench on open models, per finished round: per model and Fault family, the share of readable replies that commit the Fault with a plain prompt and with the Guard wording, with 95% intervals.

find_study

Claude
Search the List of published studies that QIA (quantum hardware) or IRIS (language models) has read and indexed. Returns title, link, classification codes, date indexed and the receipt of what was read.

hosts

Claude
HOSTS: open-weight models on OpenRouter, every host, its stated quantization and uptime, polled every 12 minutes.

huggingface_license_claims

Claude
For the 500 most downloaded Hugging Face models and 500 most downloaded datasets: the license each states in its card against the license file in its repository root. Counts by class for models and datasets, the disagreement share with a 95% interval and outcome, why repositories were not comparable, the share that state a license and carry no license file, two descriptive mismatch groups, and each mismatch read in full. Read every 30 days.

huggingface_model_sizes

Claude
For the 1,000 most downloaded Hugging Face models: the size stated in the name against the parameters in the safetensors files. Counts by class, the disagreement share with a 95% interval and outcome, why models were not comparable, each mismatch read in full, and the models where Hugging Face's total contradicts its own dtype breakdown. Read every 30 days.

list_divisions

Claude
List every division of indahl.ai with a one-line description and its page. Start here. Outcomes are ABOVE, BELOW, CONSISTENT or CANNOT_TELL, relative to what was stated. Nothing returned, refused and held are counted as their own outcomes.

model_claims_round

Claude
The live round that reruns an open model's stated benchmark results with the published harness: each claim, our result and the outcome.

package_license_claims

Claude
For the 1,000 most downloaded PyPI projects: the license each states in its metadata against the license GitHub detects in the repository it points to. Counts by class, the disagreement share with a 95% interval and outcome, why projects were not comparable, and every disagreement with both strings and the reason found on a full read. Read every 30 days.

pypi_install_commands

Claude
For the 1,000 most downloaded PyPI projects: the first pip install command in the project's description, kept only if it names the project, against pip's dry-run resolution on a fresh Python 3.12. Counts by class, the failure share with a 95% interval and outcome, why projects were not comparable, and every failure read in full. Read every 30 days.

weights

Claude
WEIGHTS: whether open-weight models keep the same weight files behind the same name, pass after pass.

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/indahl-models"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/indahl-models)