# TokenMark [Health: Active]

**Category:** 🗣️ Conversational AI  
**Repository:** https://github.com/sypherin/tokenmark-mcp  
**GitHub Stars:** 0  
**npm Downloads (last month):** 748  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/tokenmark

## Description
Measured local-LLM speeds and hardware-aware model picks for Strix Halo, DGX Spark and Mac.

## Tools
Capabilities this server exposes over MCP:

- **tokenmark_recommend**
- **lane**
- **tokenmark_configs**
- **matched**
- **other_lane**
- **tokenmark_hardware**
- **gpu_cards**
- **tokenmark_search**
- **tokenmark_submit**
- **tokenmark_submission_status**
- **tokenmark_submit_result**
- **bench_output**
- **confirm**

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "tokenmark": {
    "command": "npx",
    "args": ["-y","@altronis/tokenmark-mcp"],
    "env": {
      "TOKENMARK_URL": "",
      "TOKENMARK_TOKEN": ""
    }
  }
}
```

**Requires environment variables:** `TOKENMARK_URL`, `TOKENMARK_TOKEN` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What the TokenMark MCP server does

The TokenMark MCP server gives MCP-compatible agents access to live benchmark records from TokenMark. Its primary use is selecting a local language model and configuration for a particular machine or workload. Recommendations are based on measured results and include performance information such as decode tokens per second, quantization, backend, and an explanation for the ranking. Each recommendation includes a source link.

The catalogue covers platforms including Strix Halo, Gorgon Halo, DGX Spark, and Mac Max or Ultra systems. Hardware lookups can return chip specifications, source references for those values, the systems that ship with the chip, and Singapore pricing. This makes the server useful when hardware selection and model selection are part of the same decision.

## How it works

The TokenMark MCP server retrieves data live from `https://tokenmark.app`. It runs over stdio, so an MCP client starts the package locally with `npx`. Runtime dependencies are not required beyond the package execution environment. The service does not present estimates as measurements; recommendations and configuration results come from tracked benchmark data.

`tokenmark_recommend` accepts hardware and optional task or preference filters, then returns ranked picks. `tokenmark_configs` lists tracked configurations, with fastest results first, and identifies speculative runs such as MTP, DFlash, or draft-model configurations. `tokenmark_search` finds matching models and configurations, while `tokenmark_hardware` exposes the hardware catalogue.

## Setup and configuration

Install the server in a compatible MCP client with:

```sh
npx -y @altronis/tokenmark-mcp
```

For Claude Code, the documented command is:

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

Cline can add the same stdio server through its CLI or its VS Code MCP settings. Other clients can use a server entry with `npx` as the command and `-y @altronis/tokenmark-mcp` as its arguments.

Set `TOKENMARK_URL` to point the server at a different TokenMark endpoint. Most read operations need no sign-in. Sending measured results requires a one-time TokenMark CLI login on the same machine, or a `TOKENMARK_TOKEN` value when no home directory is available.

## Tools and capabilities

The server exposes tools for recommendations, configuration listings, hardware details, and searches. It also supports contribution workflows:

- Queue a GitHub or Hugging Face repository with `tokenmark_submit`.
- Check a submission using `tokenmark_submission_status`.
- Send measured benchmark output or manually supplied fields with `tokenmark_submit_result`.

Result submission uses a two-step confirmation process. A dry run returns a summary and confirmation code. The user must approve that summary before the same fields and code can be submitted for human review. Raw `llama-bench` or `vllm bench serve` output can be supplied, and explicit fields take precedence over parsed values.

## Limitations and notes

The catalogue is limited to the hardware and benchmark data tracked by TokenMark. A missing platform or manually entered result can instead be submitted through the signed-in TokenMark form. Submission access is for review rather than immediate publication, and sign-in credentials last 90 days. Benchmark data is aggregated from public community results with attribution, so coverage depends on available submissions.

_Full upstream README: https://allmcps.com/mcp/tokenmark/readme_