XF Model Source
An MCP server that turns a plain-language job description into ranked AI model recommendations, backed by evidence and optional side-by-side tests.
At a Glance
Discovery and ranking are free; first three discoveries need no sign-up and no key. Optional side-by-side tests run on your own OpenRouter account.
Engagement
Available On
Listed Nov 2026
About XF Model Source
XF Model Source (XFMS) is an MCP connector from Xpansion, built by Visionairy, that helps people choose which AI model to use for a specific job. You describe your project in plain words, and it works out the requirements, ranks models against them, and can test the top picks on your own tasks.
It runs inside MCP-capable assistants such as Claude Code, Cursor, Codex and Windsurf.
What It Is
XFMS is a model-selection service delivered as an MCP server. Instead of recommending a popular model, it begins with discovery: the assistant asks a few plain questions, and XFMS converts the answers into the job's real requirements, such as valid JSON output, field accuracy, or cost limits. Models are then scored against independent test results, prices and features. The site says this data is refreshed weekly and covers more than 450 models and more than 200 differentiating attributes.
How the workflow runs
The docs list several tools: discover, rank and pick, benchmark and compare, field_results, and watch_job. Rank and pick return a shortlist or a single best model. Benchmark and compare run the shortlist, or models you name, on your own example tasks and judge each requirement, showing the maximum possible cost before anything runs. Side-by-side tests use your own OpenRouter key, passed as the X-OpenRouter-Key header.
The homepage example extracts invoice fields as JSON. It tested six models, including OpenRouter's automatic picker, on three invoices. One model was ruled out for wrapping its JSON in chat text, and another was held after breaking the JSON once. The site notes that three invoices is a small sample.
Evidence and trust handling
For each requirement, a model is shown as measured (with a confidence level), judged from related tests, or as having no evidence. Missing evidence is shown as unknown, not as good. Some unusual abilities rest only on provider claims and are labelled lower trust. Daily test results are published on a results page with failures quoted.
XF Model Watch
XF Model Watch is a subscription add-on that saves a job and re-tests it weekly as new models and price changes arrive. It emails when a cheaper or better model passes all your checks, and warns before a model is retired. The user decides whether to switch.
Setup path
Connect through an HTTP MCP server address, with documented steps for Claude, Claude Code, Codex, Cursor and Windsurf. Sign-in with an email is described in the docs as required from 18 October 2026.
Community Discussions
Be the first to start a conversation about XF Model Source
Share your experience with XF Model Source, ask questions, or help others learn from your insights.
Pricing
Free
Discovery and ranking are free; first three discoveries need no sign-up and no key. Optional side-by-side tests run on your own OpenRouter account.
- Discovery and ranking free
- First three discoveries need no sign-up and no key
- Optional live tests on your own OpenRouter account cost a few cents (external cost)
XF Model Watch Solo
Watch up to 3 AI jobs, weekly
- Up to 3 AI jobs watched weekly
- Two months free when you pay yearly
- 14-day money-back promise
XF Model Watch Team
Up to 10 jobs, re-checked when new models arrive
- Up to 10 jobs
- Re-checked when new models arrive
- Two months free when you pay yearly
XF Model Watch Agency
Up to 40 jobs, grouped by client
- Up to 40 jobs
- Grouped by client
- Two months free when you pay yearly
XF Model Watch Enterprise
Talk to Xpansion
Capabilities
Key Features
- Plain-language discovery of a job's real requirements
- Ranking of more than 450 models across more than 200 attributes, refreshed weekly
- Single best-model pick or shortlist
- Side-by-side benchmark and compare on your own tasks with a cost cap shown first
- Evidence labels: measured, judged from related tests, or unknown
- Published daily test results with quoted failures
- Field results for similar jobs
- XF Model Watch weekly re-testing with email alerts on cheaper or better models and retirements
