XFMS — Xpansion Framework Model Source

Pick the right LLM for your task — without the Twitter vibes.

State what you're using the model for. XFMS aggregates evidence from eight independent benchmark sources, normalizes it onto a common scale, lets your intent decide which dimensions matter, and returns a ranked shortlist with plain-English rationale for every pick.

XFMS is one module of the Xpansion Framework — a unified architecture for governing AI-assisted work.

What this repository is

A thin Python client and command-line tool for calling the hosted XFMS API at xfms.xpansion.dev. About 250 lines of code. It turns a one-liner into a ranked LLM shortlist.

What this repository is not: the recommender engine, the score catalog, or the ingestion pipeline. Those run on the hosted service. The methodology behind every pick is published in full at docs/methodology.md — every claim there maps to code that runs at request time, you just don't run it locally.

What you say:

"Fixing bugs in our Python codebase."

What you get:

Top picks:
   1. 0.842  GPT-5.5                 (openai/gpt-5.5)         via OpenAI
   2. 0.811  Claude Opus 4.7         (anthropic/claude-opus-4.7) via Anthropic
   3. 0.798  Gemini 3.1 Pro Preview  (google/gemini-3.1-pro-preview) via Google

Inferred quality weights from your purpose:
  • structured_output_reliability  42.0%  ← BigCodeBench, Aider Polyglot
  • instruction_following          28.0%  ← LiveBench, Tau-Bench
  • factuality                     20.0%  ← MMLU, GPQA
  • coherence                      10.0%  ← LongBench

─── Explanation ───
Picked GPT-5.5: strong on structured output and instruction following —
the two dimensions that dominate code-edit work. Beats Claude on Aider
Polyglot and matches it on LiveBench reasoning, at roughly 60% of the
per-token cost.

Want to see how the picks actually behave on your kind of query? Add --ab:

─── A/B probe ───
Ran 5 test queries against the top picks.
  • GPT-4o-mini  avg_latency=5579 ms  total_cost=$0.00156  successes=5
  • GPT-5.5      avg_latency=8190 ms  total_cost=$0.07640  successes=5
  • GPT-5.4      avg_latency=8783 ms  total_cost=$0.03493  successes=5

Commentary:
  Across 5 real test queries, GPT-4o-mini was both cheapest ($0.0016 total)
  and fastest (5579 ms avg). Clear winner — 98% cheaper and 36% faster
  than the slowest pick.

What XFMS does for you

Beyond ranking, XFMS gives you these levers to honor what you actually meant:

• --primary <branch> — sacrosanct user preference. When you say "cheapest model, period", the engine switches to lexicographic ranking: cost wins, other dimensions only break ties. No more weighted-blend surprises.
• --ab — runs the top 3 picks against 5 real test queries (expanding to 10 or 15 if results trade off) and surfaces commentary on who won what. Grounds the recommendation in actual model behavior, not just benchmarks.
• --strict-priorities — when you name two co-equal drivers ("cheap but high quality too"), the engine refuses to silently blend; it asks you which way to break the tie.
• Latent-requirement suggestions — engine surfaces capabilities you didn't ask for but probably need (streaming for real-time chat, vision for OCR), so you don't get burned by what you didn't know.
• Deterministic by design — every internal model call is content- cached; same input always returns the same answer. The "I got different picks for the same question" failure mode is gone.

Install — add one URL to your AI client

XFMS is a hosted Model Context Protocol (MCP) server. There is no package to install on your machine. You point your AI assistant — Claude Code, Cursor, Continue, Cline, or any MCP-speaking host — at the URL below and the tools appear inside your chat:

https://xfms.xpansion.dev/mcp/

That's it. The three discovery tools — rank, pick, discover — are free and work with no key. Your AI assistant does the small internal thinking work; we pay for nothing on your behalf, and you pay for nothing either.

The two live-probe tools — compare and benchmark — actually run test queries against the real candidate models on OpenRouter. That inference cost rides with you, so they require your own OpenRouter key in an X-OpenRouter-Key header. You probably already have one — you're an OpenRouter model picker's target audience. If not, grab one at openrouter.ai/keys.

The key travels encrypted to our server, is never logged, never persisted — used once per request and dropped. Same security posture as every other API key in your MCP config.

Concrete install snippets for each AI client are in the next section.

Use it inside Claude Code, Cursor, or any MCP client

XFMS speaks Model Context Protocol (MCP) — the standard your AI assistant uses to call external tools. Once connected, you can ask the assistant "which model should I use for OCR on shipping manifests?" and it calls XFMS for you. No leaving the chat. No copy-pasting between windows.

Hosted install — one line, no install required

The XFMS engine hosts the MCP server itself at https://xfms.xpansion.dev/mcp/. Two install shapes depending on which tools you want.

The free three — rank, pick, discover

Just point your client at the URL. No key, no signup.

Claude Code:

claude mcp add xfms --transport http https://xfms.xpansion.dev/mcp/

Cursor (~/.cursor/mcp.json) — or paste through Settings → MCP:

{
  "mcpServers": {
    "xfms": {
      "url": "https://xfms.xpansion.dev/mcp/"
    }
  }
}

All five tools — adds compare and benchmark

These two run real test queries against the actual candidate models on OpenRouter, so they require your OpenRouter key in an X-OpenRouter-Key header. Same install, one extra line:

Claude Code:

claude mcp add xfms --transport http https://xfms.xpansion.dev/mcp/ \
  --header "X-OpenRouter-Key: sk-or-v1-your-key-here"

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "xfms": {
      "url": "https://xfms.xpansion.dev/mcp/",
      "headers": {
        "X-OpenRouter-Key": "sk-or-v1-your-key-here"
      }
    }
  }
}

Continue / Cline / any other MCP host — same URL + headers pattern; check your host's docs for the JSON config shape.

Don't have an OpenRouter key yet? Grab one at openrouter.ai/keys. Restart your client, then ask it:

"Use XFMS to pick a model for summarizing long legal contracts."

Five tools are available to the assistant: rank (a ranked shortlist), pick (the single best pick), discover (which quality dimensions matter for your purpose, without ranking), compare (live A/B between models you've already named), and benchmark (live A/B against the engine's top 3 picks). compare and benchmark require the X-OpenRouter-Key header above; the other three don't.

Override the system's inference

If you know which quality dimension matters most for your task, say so — your preference always wins over the LLM's inference:

xfms rank "code refactor" --leaf-priorities "structured_output_reliability=1.0,factuality=0.5"

xfms.rank(
    "code refactor",
    leaf_priorities={"structured_output_reliability": 1.0, "factuality": 0.5},
)

How XFMS picks — the four principles

Methodology in full at docs/methodology.md. The short version:

1. No provider self-reports. Every score comes from a third-party evaluator running the same protocol across every model.
2. No single-source dependence. Eight independent benchmark sources contribute today; no single leaderboard determines a pick.
3. User intent beats LLM inference. The system infers weights from your purpose, but your stated leaf_priorities always override the inference.
4. Honest gaps over invented signal. Missing data is recorded as missing — no interpolation, no synthetic scores. Coverage gaps surface on every pick.

Part of the Xpansion Framework

XFMS doesn't stand alone — it's the model-selection layer of the Xpansion Framework.

The Xpansion thesis

Humans communicate with intent compressed by contextual experience. AI simply predicts patterns in language. Xpansion is the execution layer that bridges them.

Every sentence a human types carries lifetimes of context that the speaker assumes the other side will decompress — what counts as "good enough," which constraints are non-negotiable, what failures last month taught them, what their house style demands. AI doesn't share that context. It predicts patterns in language, filling in the gaps with whatever's plausible to its training data. The result reads as plausible but isn't intent-honoring: sessions lose context, security holes ship silently, contracts break without warning, and there's no way to verify that what was built actually matches what was asked for. They don't know what they don't know, and neither does AI.