Log inSubmit server
TokCalc MCP Server logo

TokCalc MCP Server

Plans LLM serving capacity by estimating VRAM, latency, throughput, topology, and deployment cost from model and workload inputs.

InstallRepositoryWebsite
Full README Compare alternatives Embed widget Claim this listing

Install TokCalc MCP Server

Pick your client, then add it in one click or paste the config.

STDIONode.js
~/Library/Application Support/Claude/claude_desktop_config.json
claude_desktop_config.json
{
  "mcpServers": {
    "tokcalc-mcp-server": {
      "command": "npx",
      "args": [
        "-y",
        "@tokcalc/mcp-server"
      ]
    }
  }
}

Settings → Developer → Edit Config, merge this into mcpServers, then fully quit and reopen Claude Desktop.

Automated check passed— started and listed 11 tools correctly (13d ago).

Overview

TokCalc MCP Server exposes tokcalc’s LLM serving capacity-planning engine for agent-driven deployment analysis. It evaluates model fit, KV-cache memory, throughput, latency, multi-GPU topology, and cost using model, accelerator, quantization, and workload parameters. Reach for it when choosing hardware, comparing self-hosting with APIs, or assessing long-context workloads before deployment. The same calculation engine also powers the tokcalc website and CLI.

Use cases

  • Estimate whether a model and KV cache fit on selected GPUs
  • Compare self-hosted inference costs with API pricing
  • Plan GPU topology for long-context workloads
  • Evaluate capacity before renting cloud accelerators
  • Gate serving-plan changes in CI

Key features

  • VRAM and KV-cache sizing
  • Throughput and latency estimation
  • Multi-GPU topology recommendations
  • Long-context concurrency planning
  • Self-host versus API break-even analysis
  • Live GPU price comparison

Tools11

What TokCalc MCP Server lets your agent do. Expand a tool for its parameters.

Verified live~3.2k tokens of context

Security & health

What it can reach and how reliably it runs — from our own checks, not the maintainer’s claims.

ActiveLast checked 1d ago

Recent health check succeeded.

9/9 checks healthy over the last 37d

What it can do

From the 11 tools this server returned when we ran it; 11 declare their own safety hints.

Handles credentials · 1Makes network requests · 1Read-only tools · 11
Network at startup3 connections
Files written at startupNone
Install scriptsNone
npm provenanceSigned build
Package age36 days
Maintainers1

Observed 2026-10-30 in an isolated sandbox. A description of reach, not a security audit.

Supply-chain signal

No high-severity advisories surfaced by our automated scan.

Critical 0High 0Medium 0Low 0

Scanned 13d ago via OSV.dev · @tokcalc/mcp-server (npm)

Adoption & maintenance

Factual signals from GitHub, npm, and our automated checks — not a rating.

GitHub stars1Stargazers on the source repository.
npm downloads1.4kPackage downloads in the last 30 days.
Last commit19d agoMost recent push to the default branch.
Availability100%Our rolling endpoint + install checks that succeeded.
Install checkPassedOur sandbox started it and listed its tools.
Tools exposed11Callable tools this server registers over MCP.
Directory activity1 viewsConfig copies, upvotes, and views on AllMCPs.

How TokCalc MCP Server works

What TokCalc MCP Server does

TokCalc MCP Server brings tokcalc’s LLM serving capacity planner into MCP-based agent workflows. It turns a selected model, accelerator, quantization format, context length, concurrency, and latency or traffic assumptions into a deployment estimate.

The calculations cover more than decode speed. Results can include model and KV-cache fit, per-stream and aggregate throughput, prefill speed, time to first token, inter-token latency, maximum concurrency at long context lengths, recommended multi-GPU topology, and operating cost. Cost outputs can extend from GPU hourly pricing to cost per million output tokens, cost per request, and monthly estimates.

The planner also supports build-versus-buy analysis. Self-hosted configurations can be compared with API pricing across providers, including a break-even assessment based on request volume.

How it works

The MCP package uses the shared tokcalc calculation engine used by the web application and command-line interface. Its estimates are based on a catalog of models, accelerators, quantization formats, API prices, and cloud GPU prices. The catalog includes 44 models, 31 accelerators, and 16 quantization formats.

Long-context planning accounts for KV memory per request and the increasing prefill cost of larger contexts. Capacity views cover contexts from 4K through 1M tokens and can recommend single-GPU, tensor-parallel, or context-parallel arrangements. The calculator also models configurable effects such as continuous batching, speculative decoding, prompt caching, and hidden reasoning tokens.

GPU pricing can be retrieved from Azure, AWS, GCP, and Vast.ai. Provider requests are handled independently, so a failed provider lookup does not prevent other pricing sources from returning results.

Setup and configuration

The package is published as @tokcalc/mcp-server. Its command-line interface can create a plan from a .tokcalc.json file or from command-line arguments. A basic configuration can specify fields such as:

  • model
  • gpu
  • quantization
  • numGpus
  • contextTokens
  • concurrency

For example, the package supports commands such as npx @tokcalc/mcp-server plan --config .tokcalc.json and direct planning with model, GPU, quantization, and output-format flags. JSON output is available for automation. The CLI returns exit code 0 for a fitting plan, 2 when the plan is infeasible, and 1 for a usage error.

GCP live pricing uses the optional GCP_API_KEY environment variable. Other listed pricing providers use public endpoints described by the project, and provider results are cached for either 24 hours or, for Vast.ai spot pricing, 5 minutes.

Tools and capabilities

  • Estimate VRAM and KV-cache requirements.
  • Calculate decode, aggregate, and prefill throughput.
  • Split latency into first-token and inter-token components.
  • Recommend tensor-parallel and context-parallel GPU layouts.
  • Compare self-hosted economics with API pricing and break-even volume.
  • Calibrate estimates against vLLM, SGLang, or TensorRT-LLM benchmark JSON.
  • Retrieve and compare live GPU prices across supported providers.

Limitations and notes

The outputs are planning estimates rather than deployment measurements. Benchmark calibration is available for observed runtime data, but accuracy depends on the supplied benchmark and workload assumptions. Catalog records and static cloud prices are curated data, while live provider pricing depends on successful upstream requests and any required provider credentials. Pricing and performance should therefore be reviewed against the target serving stack before committing to hardware.

Read the full README View source on GitHub

Reviews

No reviews yet — be the first to share how it worked for you.

Frequently Asked Questions about TokCalc MCP Server

TokCalc MCP Server is an MCP server that exposes tokcalc’s LLM serving capacity-planning calculations for model, GPU, workload, VRAM, KV cache, throughput, latency, topology, and cost estimates. It uses the shared engine behind the website and CLI, helping agents evaluate deployment plans before selecting or renting GPUs.

Run it through the published `@tokcalc/mcp-server` package with `npx`. The package supports commands such as `npx @tokcalc/mcp-server plan --config .tokcalc.json` or direct model, GPU, quantization, and output-format arguments.

It calculates model fit, VRAM and KV-cache needs, throughput, prefill speed, first-token latency, inter-token latency, GPU topology, concurrency at long context lengths, and serving costs. It can also compare self-hosting with API pricing.

Yes, it can fetch live pricing from Azure, AWS, GCP, and Vast.ai. GCP uses `GCP_API_KEY`; if one provider fails, the other provider results can still be used.

Yes, the package’s planning command returns status codes suitable for automation: `0` when the plan fits, `2` when it is infeasible, and `1` for a usage error. A GitHub Action can also comment plan results on pull requests that change a serving configuration.

Related MCP servers

All in Developer Tools All alternatives
Add a badge to your READMEShow it’s listed — live SVG, Markdown or HTML.

AllMCPs Directory Badge

Full Badge Customizer

Showcase your server listing on GitHub or your project documentation. Embed this dynamic SVG badge to highlight official listing status and live engagement.

Badge Style:
Live Dynamic SVG PreviewTokCalc MCP Server AllMCPs Directory Badge
Markdown (GitHub README)
[![AllMCPs](https://allmcps.com/api/badge/tokcalc-mcp-server?style=directory)](https://allmcps.com/mcp/tokcalc-mcp-server)
HTML Embed
<a href="https://allmcps.com/mcp/tokcalc-mcp-server"><img src="https://allmcps.com/api/badge/tokcalc-mcp-server?style=directory" alt="TokCalc MCP Server on AllMCPs" /></a>