TokenMark

@altronis/tokenmark-mcp

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

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.