Install
Official MCP Registry•npm•v0.2.3

Ask which model is best right now.

ModelShortlist gives your AI assistant independent Artificial Analysis benchmark evidence plus current OpenRouter facts—price, context, capabilities, providers, and optional ZDR—so it can recommend the best fit for the workload you actually have.

Install ModelShortlistSee the selection framework
Current
Upstream evidence
Local
Runs on your machine
BYOK
Your API credentials
Open source
MIT licensed
modelshortlist / recommend_models
Example
Workload

“Long-running coding agent. Tools required. At least 100k context. Quality matters most, but I care about value.”

Artificial Analysis
Independent benchmark + performance evidence
OpenRouter
Current catalog + operational facts
Decision structureNo universal score
Apply hard constraints first
Compare quality and performance evidence
Weigh current price and operational fit
Explain best-fit and best-value tradeoffs
Source freshness is part of the tool output, so degraded evidence can be qualified instead of silently presented as current.

Why freshness matters

Why yesterday’s model recommendation may be wrong today.

Static comparison content gets stale because the model market is not static. ModelShortlist is built to put current evidence inside the assistant that already understands your workload.

A new model launches

The competitive set can change immediately, even if your workload stays exactly the same.

Pricing moves

A price cut can turn a previously expensive model into the best-value option.

Benchmark evidence changes

Independent evaluations can change the quality picture as new results appear.

Capabilities change

Tool support and other parameters can appear, disappear, or differ across model variants.

Providers change

Endpoint availability and provider options move independently of a model’s benchmark quality.

ZDR availability changes

Privacy-sensitive workloads need current endpoint evidence, not assumptions about a model family.

See how ModelShortlist handles freshness and degraded upstreams

Evidence, with clear attribution

Benchmarks tell you how models perform. Current operational facts tell you what fits.

ModelShortlist keeps those evidence types conceptually separate, then lets the host AI reason across them for the workload you described.

Independent performance evidence

Artificial Analysis

Artificial Analysis supplies the independent benchmark and performance evidence. ModelShortlist does not create those benchmarks and only attaches them when it can confidently reconcile the exact model identity.

  • Intelligence, coding, and agentic evaluation context
  • Performance and pricing fields exposed by the upstream API
  • Conservative matching: missing evidence is preferred over a wrong match
How Artificial Analysis contributes

Current operational facts

OpenRouter

OpenRouter supplies the current catalog and deployment-relevant facts that can make a benchmark-strong model the wrong operational choice for a specific job.

  • Model availability, context, and supported parameters
  • Current input/output and tiered pricing evidence
  • Provider and ZDR endpoint facts when privacy is explicitly required
Compare OpenRouter models by workload
The judgment stays workload-specific. ModelShortlist retrieves and normalizes evidence; the AI in your MCP client weighs that evidence against the task. There is no permanent global score pretending one model is best for every job.

How it works

Current evidence in. Workload-specific reasoning out.

ModelShortlist is a read-only context layer, not a model router. It retrieves evidence; your AI assistant makes the recommendation.

01

Describe the workload

State what you are building, what matters, and any hard constraints such as tools, context, budget, creator, or ZDR.

02

Build the eligible set

ModelShortlist starts from the current OpenRouter catalog and applies explicit hard constraints rather than a hand-picked leaderboard.

03

Attach current evidence

Operational facts are combined with Artificial Analysis evidence when the exact model identity can be matched confidently.

04

Reason to a shortlist

Your host AI explains best-fit, best-value, and constraint-driven tradeoffs using the evidence returned by the MCP.

Ask naturally

Start with the decision you actually need to make.

No special prompt syntax is required. Be explicit about hard constraints; let the assistant reason about everything else.

PROMPT 01

“I need a long-running coding agent. Tool use is required, I need at least 100k context, and I care about quality-per-dollar.”

PROMPT 02

“What is the strongest current model under $10 per million output tokens for this workload?”

PROMPT 03

“I need strict structured output from long documents. Which current models are the best fit?”

PROMPT 04

“Compare these OpenRouter models and separate benchmark evidence from price, context, and provider facts.”

PROMPT 05

“I need 200k context and reliable tool use. What are my best current options?”

PROMPT 06

“Same workload, but ZDR is mandatory. Which real endpoints still satisfy every hard constraint?”

Guides

Learn the decision, not a temporary top-10 list.

The guides explain what criteria matter and how to ask for a current answer without hard-coding today’s model ranking into tomorrow’s advice.

Browse all guides

How to choose an AI model

A practical framework for hard constraints, quality evidence, economics, and current operational fit.

Read guide

Best models for coding agents

Tool use, context, coding and agentic evidence, and long-running token economics.

Read guide

Models for document extraction

Extraction reliability, structured output, context, and high-volume cost tradeoffs.

Read guide

Models under $10 / 1M output

Apply a hard price ceiling, then compare the strongest capable models that remain.

Read guide

ModelShortlist vs leaderboards

Why independent benchmarks are important evidence but not the whole deployment decision.

Read guide

ModelShortlist vs model routers

Recommendation and runtime inference routing solve different problems.

Read guide

Install

Bring two keys. Keep them local.

ModelShortlist is a local stdio MCP server. You supply Artificial Analysis and OpenRouter credentials; there is no ModelShortlist account or hosted key vault in the middle.

1. Get upstream API keys

Use your own Artificial Analysis and OpenRouter API keys. ModelShortlist does not issue or proxy either credential.

2. Add the local MCP server

Use the npm package directly in your MCP client or follow the guided install page for client-specific config.

3. Ask the real model-selection question

Describe the workload, hard constraints, and what you value. Your assistant can call ModelShortlist when it needs current evidence.

MCP config · npm
{
  "mcpServers": {
    "modelshortlist": {
      "command": "npx",
      "args": ["-y", "@agentic.artists/modelshortlist"],
      "env": {
        "ARTIFICIAL_ANALYSIS_API_KEY": "YOUR_KEY",
        "OPENROUTER_API_KEY": "YOUR_KEY"
      }
    }
  }
}
Client-specific configuratornpm package

Supported setup guidance includes Claude Desktop, Claude Code, Cursor, Hermes Desktop, and VS Code / Copilot. Windows configurations use npx.cmd where appropriate.

Local + open source

Model selection evidence without another hosted account.

The project is intentionally a local BYOK context layer. It stays out of the inference path and keeps the architecture inspectable.

View the source on GitHub

No hosted ModelShortlist backend

Your local MCP process calls upstream APIs directly.

No MCP telemetry

ModelShortlist does not add product telemetry to the local MCP runtime.

MIT licensed source

Inspect the code, package contents, matching logic, and release workflow yourself.

Read-only MCP tools

The MCP retrieves and compares evidence; it does not place model calls or mutate external systems.

FAQ

What ModelShortlist is—and what it is not.

A local evidence layer for model selection, with explicit attribution and current operational context.

Does ModelShortlist choose one “best” model for everyone?

No. The host AI weighs current evidence against your workload and constraints. Coding, extraction, long-context, latency-sensitive, budget-sensitive, and privacy-sensitive work can produce different shortlists.

What does Artificial Analysis contribute?

Artificial Analysis supplies independent benchmark and performance evidence. ModelShortlist uses that evidence when it can confidently reconcile the exact model identity; it does not create or claim ownership of those benchmarks.

What does OpenRouter contribute?

OpenRouter supplies the current model catalog plus operational facts such as context, supported parameters, pricing, providers, and ZDR endpoint information.

What happens if upstream data is stale or unavailable?

ModelShortlist tracks source freshness separately. It can use explicitly stale last-known-good evidence where appropriate, marks unavailable sources, and fails closed when a required source or hard privacy constraint cannot be verified.

Does it only recommend ZDR models?

No. The full OpenRouter catalog is considered by default. Zero Data Retention becomes a hard eligibility requirement only when you explicitly request ZDR.

Do my API keys go through a ModelShortlist server?

No. ModelShortlist is a local stdio MCP server. Your local process uses the API credentials you provide to call the upstream services directly; there is no hosted ModelShortlist backend or MCP telemetry.

Ask the current model market, not yesterday’s recommendation.

Install the local MCP, add your Artificial Analysis and OpenRouter keys, and let your assistant reason from current evidence for the workload in front of you.

Install ModelShortlist Browse guides