Skip to main content
Glama

@altronis/tokenmark-mcp

M8ven Score

An MCP server that gives Claude (and other agents) real local-LLM benchmark data + hardware-aware model recommendations from TokenMark.

Ask "what should I run on a Strix Halo for coding?" and the agent answers from measured configs (decode tok/s, quant, backend), each with a source link. It never invents numbers.

Add to Claude Code

claude mcp add tokenmark -- npx -y @altronis/tokenmark-mcp

Related MCP server: inferbench

Add to Cline

Cline CLI:

cline mcp add tokenmark --yes -- npx -y @altronis/tokenmark-mcp

Cline in VS Code: add the tokenmark entry from the JSON below to cline_mcp_settings.json. Step-by-step notes for agents are in llms-install.md.

Add to any MCP client

Run the server over stdio:

npx -y @altronis/tokenmark-mcp

Or in a client config:

{
  "mcpServers": {
    "tokenmark": {
      "command": "npx",
      "args": ["-y", "@altronis/tokenmark-mcp"]
    }
  }
}

Tools

  • tokenmark_recommend: { hardware, tasks?, prefer?, limit?, cards? } → ranked model + best-config picks (5 by default, 10 max) with measured tok/s + why. lane says which lane they came from: GPU cards (RTX 5090, RTX PRO 6000) rank on their own, never against the AI boxes. Naming a card, or cards: true, reads the card lane.

  • tokenmark_configs: { model?, hardware?, limit?, cards? } → tracked benchmark configs from one lane, fastest first (25 by default, 50 max; matched gives the full count), with the run mode (speculative: true for MTP/DFlash/draft-model runs). other_lane counts the matches in the other lane.

  • tokenmark_hardware: { platform? } → the hardware catalogue (Strix Halo, Gorgon Halo, DGX Spark, Mac Max/Ultra), plus gpu_cards with measured runs. With a platform: chip specs with a source per value, the boxes that ship it, Singapore prices. With a card ("RTX 5090"): VRAM and bandwidth with sources, and its tracked configs.

  • tokenmark_search: { term } → matching models/configs, with GPU-card matches apart under gpu_cards.

  • tokenmark_submit: { repo, source?, note? } → queues a GitHub/Hugging Face repo with benchmark numbers for human review.

  • tokenmark_submission_status: { id } → where a submission is.

  • tokenmark_submit_result: { bench_output?, pick?, model?, quant?, hardware?, backend?, decode_tps?, prefill_tps?, ctx?, concurrency?, ..., dry_run? } → sends a speed the user measured, for human review. Pass raw llama-bench or vllm bench serve output as bench_output, or the fields by hand; fields you pass win over what it read. Sending takes two calls: dry_run: true returns a one-line summary and a confirm code; show the user the summary, and once they agree, call again with the same fields and confirm set to that code. A send without the code, or with fields changed since the dry run, is refused.

Sending measured speeds

tokenmark_submit_result needs a one-time sign-in on the same machine, in a terminal:

npx @altronis/tokenmark-cli login

Approve the short code at https://tokenmark.app/cli. The sign-in can only send results for review and lasts 90 days; the server reads it on every call, so no restart is needed. Where there is no home directory, set TOKENMARK_TOKEN in the server's environment instead. A sign-in is only sent over https (or plain http to this machine), and every sign-in on the account can be revoked at https://tokenmark.app/cli.

Hardware missing from the catalogue, or a result typed in by hand, can also go through the signed-in form at https://tokenmark.app/submit.

Data is pulled live from https://tokenmark.app (override with TOKENMARK_URL). Zero runtime dependencies.

The same data in your terminal

The CLI lives in cli/ of this repo:

npx @altronis/tokenmark-cli recommend --hardware "Strix Halo" --tasks coding

The files in bin/ are the published builds, made from the TokenMark tracker that runs https://tokenmark.app.

MIT licensed. Data aggregated from public community benchmarks with attribution.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      InferBench's MCP server lets coding agents run, serve and benchmark local LLMs (text + image, llama.cpp + Stable Diffusion) on your own hardware on demand — measuring real tokens/sec and picking the optimal quant for your GPU from a 124-model catalog. Local-first, no cloud required.
      2
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables benchmarking of local LLM models (performance and quality) and sharing results to a public leaderboard via MCP tools.
      16 npm
      8
      Apache 2.0