# TokCalc MCP Server **Category:** ๐Ÿ’ป Developer Tools **Repository:** https://github.com/stevecrates489-commits/tokcalc.git **Views:** 0 **Installs:** 0 **Upvotes:** 0 **Directory Page:** https://allmcps.com/mcp/tokcalc-mcp-server ## Description LLM capacity planning tools to estimate VRAM, compute, latency, and GPU topology for agents. ## Claude Desktop Quick Installation Heuristic fallback โ€” verify the package name and runner against the repository README before running it. Uses `npx` (confidence: low): ```json "mcpServers": { "tokcalc-mcp-server": { "command": "npx", "args": ["-y","tokcalc-mcp-server"] } } ``` ## Documentation & README
# tokcalc ### The open-source LLM serving capacity planner **Plan your LLM deployment before you rent the GPUs.** [![Live demo](https://img.shields.io/badge/live-demo-10b981?style=flat-square)](https://tokcalc.vercel.app) [![License: Apache 2.0](https://img.shields.io/badge/license-Apache_2.0-blue?style=flat-square)](LICENSE) [![PRs welcome](https://img.shields.io/badge/PRs-welcome-10b981?style=flat-square)](CONTRIBUTING.md) [![Made with Next.js](https://img.shields.io/badge/made_with-Next.js_16-black?style=flat-square)](https://nextjs.org) [![MCP server](https://img.shields.io/badge/MCP-@tokcalc/mcp--server-10b981?style=flat-square)](https://www.npmjs.com/package/@tokcalc/mcp-server)
---
![tokcalc demo](https://raw.githubusercontent.com/stevecrates489-commits/tokcalc/HEAD/public/tokcalc-demo.gif) *Switch model โ†’ multi-GPU โ†’ long-context capacity planner โ†’ Build vs Buy โ†’ Reference catalog โ†’ Live GPU pricing*
--- tokcalc turns your LLM traffic, context length, latency SLOs, cache behavior, and model choice into a **defensible serving topology and cost plan** โ€” with transparent formulas and cited benchmarks. It's the engineering-grade pre-deployment decision layer for LLM inference. Not another static "tokens per second" calculator. ## Try it now **[tokcalc.vercel.app](https://tokcalc.vercel.app)** โ€” no signup, no tracking, no paywall. Pick a model, a GPU, and a workload. Get an instant capacity plan: - Generation speed (tok/s) - Time-to-first-token (TTFT) + inter-token latency (ITL) - VRAM budget with KV-cache sizing - Multi-GPU topology recommendation (Single GPU โ†’ TPร—2/4/8 โ†’ Context Parallel) - Monthly cost + break-even vs API pricing - **Live GPU pricing** across Azure + AWS + GCP + Vast.ai marketplace (fetched in parallel) - **Shareable URL** โ€” your config encoded in the URL hash, send to colleagues - **Copy as Markdown** โ€” paste the full result into GitHub issues / Slack / docs ## Why tokcalc? The market has dozens of "tokens per second" calculators and self-host-vs-API break-even tools (induwara.lk, gigagpu, kickllm, cloudparity, curlscape, profitable.ai). None of them are unified capacity planners โ€” and none show their math. ### What tokcalc answers that competitors can't > *"Can I serve Qwen 2.5 72B at 128K context on 2ร— H100 with 20 concurrent users?"* > *"How many H200s do I need for 1,000 req/min with P95 TTFT < 2s?"* > *"Does FP8 or AWQ save more money once quality, KV cache, and engine support are included?"* > *"At what daily volume does an H100 beat GPT-4o pricing?"* > *"What's the cheapest H100 right now across Azure / AWS / GCP / Vast.ai?"* > *"What happens to cost and latency if an agent makes 8 model calls, has 3 tool calls, and its context grows by 5K tokens each turn?"* > *"Would prefix caching, continuous batching, or PD disaggregation save more for this workload?"* ## Features ### Calculator tab | Feature | What it computes | |---|---| | **Model fit / VRAM** | Will the model + KV cache fit in the GPU's memory? | | **Throughput** | Decode tok/s (per-stream) + aggregate (batched) + prefill tok/s | | **Latency split** | Time-to-first-token (= prefill) + inter-token latency (= decode) | | **Continuous batching** | User-tunable 1.0โ€“4ร— multiplier (cited 1.5โ€“4ร— SOSP range) | | **Reasoning tokens** | Hidden reasoning budget added to billed output (o1 / R1 / Claude thinking) | | **Prompt caching** | Self-hosted vLLM APC + Anthropic 5m/1h TTL + OpenAI 50%-off cached tokens | | **Speculative decoding** | User-tunable 1.2โ€“4ร— boost factor | | **Multi-GPU TP** | 1ร— โ†’ 8ร— tensor parallel with NVLink efficiency factor | | **Long-context capacity** | KV memory + max concurrency + prefill time at 4K โ†’ 1M context | | **Topology recommendation** | Single GPU โ†’ TPร—2 โ†’ TPร—4 โ†’ TPร—8 โ†’ TPร—8 + Context Parallel (RingAttention) | | **Cost economics** | GPU $/hr โ†’ $/M output tokens โ†’ $/request โ†’ monthly cost | | **Observed benchmark calibration** | Paste vLLM/SGLang/TRT-LLM JSON โ†’ see formula accuracy verdict (validated / underestimated / overestimated) | ### Build vs Buy tab Independent calculator (separate state) that compares: - **Self-host**: model + GPU + quant + utilization + batch โ†’ $/M tokens + monthly cost - **API**: 13 providers (OpenAI / Anthropic / Gemini / Groq / DeepSeek / Mistral / Together) - **Verdict**: Self-host cheaper / API cheaper / Not enough volume โ€” with break-even reqs/day ### Reference tab 6 sub-tables โ€” fully transparent, every record source-linked where available: - **Models** (35 entries: Llama 4 Scout/Maverick, Qwen 3 family, DeepSeek V3/R1, Pixtral, BGE-M3, ...) - **GPUs** (30 entries: H100/H200/B200/B300, AMD MI300X/MI325X, Intel Gaudi 3, TPU v5p/Trillium, Groq LPU, Cerebras WSE-3, Apple M2/M3/M4 Ultra, ...) - **Quantization** (16 formats: FP16/BF16, GGUF Q2_Kโ†’Q8_0, GPTQ, AWQ, EXL2, FP8, NVFP4) - **API pricing** (13 models with input/cached/output + retired/current status) - **Cloud GPU pricing** (all GPUs with $/hr > 0 + typical providers โ€” static curated) - **๐Ÿ”ด Live pricing** (real-time fetch from Azure + AWS + GCP + Vast.ai in parallel โ€” LIVE badges + timestamps) ### The long-context capacity planner (the differentiator) This is the formula the Perplexity research brief called "the most important tokcalc should visibly expose": $$ \text{KV bytes/request} = 2 \cdot L \cdot T \cdot H_{\text{kv}} \cdot D_h \cdot B $$ For dense attention, prefill cost grows **superlinearly** with context length: $$ \text{prefill FLOPs} = \underbrace{2 \cdot N \cdot T}_{\text{linear}} + \underbrace{T^2 \cdot H_{\text{kv}} \cdot D_h \cdot L}_{\text{attention}} $$ tokcalc shows you: - Max concurrent users at 4K / 8K / 16K / 32K / 64K / 128K / 256K / 512K / 1M context - KV memory per request at each context length - Prefill time (with superlinear attention correction beyond 32K) - Required topology (Single GPU โ†’ TPร—2/4/8 โ†’ TPร—8 + Context Parallel) - RingAttention citation when CP is needed ### Live GPU pricing (NEW) The ๐Ÿ”ด Live pricing sub-tab in Reference fetches real-time GPU prices from 4 providers in parallel via `Promise.allSettled`: | Provider | API | Auth | Cache TTL | |---|---|---|---| | **Azure** | Retail Prices API | None (public) | 24h | | **AWS** | EC2 bulk pricing file | None (public) | 24h | | **GCP** | Cloud Billing Catalog API | `GCP_API_KEY` env var | 24h | | **Vast.ai** | Marketplace bundles API | None (public) | 5m (spot prices change rapidly) | Each provider shows a LIVE badge with timestamp + cache state. The comparison table shows the cheapest price per GPU across all 4 providers (highlighted in emerald) + per-provider breakdown. If any provider fails (e.g., GCP_API_KEY not set), the others still work โ€” graceful degradation per provider. ## The math, transparently Every number above comes from a formula you can inspect. No black boxes.
Decode tokens/sec (memory-bandwidth bound) $$ \text{decode tok/sec} \approx \frac{\text{HBM BW} \cdot \eta_{\text{mem}} \cdot \text{quant\_eff}}{\text{model size}} $$ Where: - `HBM BW` = GPU memory bandwidth (e.g., 3350 GB/s for H100 SXM) - `ฮท_mem = 0.65` = typical real-world memory utilization (35% overhead) - `quant_eff` = dequantization efficiency multiplier (1.0 for FP16, 1.5 for FP8 on H100, 0.85 for INT4) - `model size = active_params ร— bytes_per_param` (uses ACTIVE params for MoE, not total) Refs: PagedAttention paper ([arxiv.org/abs/2309.06180](https://arxiv.org/abs/2309.06180))
Prefill tokens/sec (compute bound) $$ \text{prefill tok/sec} \approx \frac{\text{GPU FLOPS} \cdot \eta_{\text{compute}}}{2 \cdot \text{active params}} $$ Where: - `GPU FLOPS` = dense FP16/BF16 TFLOPS (sparse values not used) - `ฮท_compute = 0.50` = typical compute utilization - Factor of 2 = one multiply + one add per parameter per token For long context (>32K), the superlinear attention correction above applies.
Continuous batching multiplier (workload-specific) $$ \text{aggregate tok/sec} = \text{decode tok/sec} \cdot \text{batch size} \cdot \text{continuous batching multiplier} $$ **Critical caveat**: There is no universal continuous batching multiplier. vLLM reported 14โ€“24ร— vs HF Transformers (extreme), 2.2โ€“2.5ร— vs TGI. SOSP paper finds 2โ€“4ร— typical vs FasterTransformer/Orca. tokcalc defaults to a conservative **1.5ร—** and lets you tune. Refs: - [vLLM blog (2023-06-20)](https://vllm.ai/blog/2023-06-20-vllm) - [PagedAttention paper](https://arxiv.org/abs/2309.06180) - [Anyscale continuous batching study](https://www.anyscale.com/blog/continuous-batching-llm-inference)
Prompt caching economics (Anthropic / OpenAI) For shared prefix of length $T_p$, suffix of length $T_u$, output $O$, cache hit rate $h$: $$ \text{API input cost} = N \cdot \left[ (1-h) \cdot T_p \cdot P_{\text{write}} + h \cdot T_p \cdot P_{\text{read}} + T_u \cdot P_{\text{input}} \right] $$ Anthropic multipliers (verified 2025-2026): - 5-minute cache write: 1.25ร— base input - 1-hour cache write: 2.0ร— base input - Cache read: 0.1ร— base input (90% savings) OpenAI: cached input discounted 50%, no separate write fee. Refs: - [Anthropic prompt caching docs](https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching) - [OpenAI prompt caching guide](https://platform.openai.com/docs/guides/prompt-caching)
Multi-GPU topology recommendation $$ \text{total needed} = \text{model weights} + (\text{KV per request} \cdot \text{batch size}) $$ Walk the smallest topology that fits: - **Single GPU**: total โ‰ค VRAM ร— 1 - **Tensor Parallel ร—2/4/8**: total โ‰ค VRAM ร— N (weights + KV split evenly) - **TPร—8 + Context Parallel**: total > VRAM ร— 8 โ€” use RingAttention to shard KV across nodes Refs: [RingAttention paper](https://arxiv.org/abs/2310.01889)
Self-host vs API break-even $$ \text{self-host $/M tokens} = \frac{\text{GPU $/hr}}{3600 \cdot \text{effective tok/s} \cdot \text{utilization}} \cdot 10^6 $$ $$ \text{break-even req/day} = \frac{\text{monthly self-host cost}}{30 \cdot \text{API cost per request}} $$ The decisive term is **effective utilization** โ€” not peak throughput. A GPU running at 10% utilization pays 10ร— more per token than the theoretical minimum.
## Comparison with adjacent tools | Capability | induwara / techfuelhq / pcmasterstudio | gigagpu / kickllm / cloudparity | HF Open LLM Leaderboard / MLPerf | **tokcalc** | |---|:---:|:---:|:---:|:---:| | Model ร— GPU ร— quant tok/s | โœ“ | some | some | **โœ“** | | Model-fit / VRAM | basic | rare | rare | **โœ“** | | KV-cache by context + concurrency | โ€” | โ€” | implicit | **โœ“** | | Prefill vs decode split (TTFT/ITL) | โ€” | โ€” | engine-specific | **โœ“** | | Continuous batching / paged attention | โ€” | โ€” | docs only | **โœ“** | | Long-context (128Kโ€“1M) planning | โ€” | โ€” | โ€” | **โœ“** | | Topology recommendation (TP/CP) | โ€” | โ€” | partial | **โœ“** | | Prompt-cache economics | โ€” | partial API only | โ€” | **โœ“** | | Reasoning tokens (o1/R1/Claude thinking) | โ€” | โ€” | โ€” | **โœ“** | | API vs self-host break-even | some | โœ“ | โ€” | **โœ“** | | **Live cloud GPU pricing (4 providers)** | โ€” | โ€” | โ€” | **โœ“** | | **Observed benchmark calibration** | โ€” | โ€” | โ€” | **โœ“** | | **MCP server for AI agents** | โ€” | โ€” | โ€” | **โœ“** | | **Public hosted MCP endpoint + self-serve keys** | โ€” | โ€” | โ€” | **โœ“** | | Transparent formulas / open source | mixed | usually no | mixed | **โœ“** | | Cited benchmark evidence per config | rare | rare | โœ“ (not planning) | **โœ“ (in progress)** | | Shareable URL per config | โ€” | โ€” | โ€” | **โœ“** | | Copy result as Markdown | โ€” | โ€” | โ€” | **โœ“** | ## Roadmap ### Shipped - โœ… 35 models, 30 GPUs, 16 quantization formats - โœ… Continuous batching, reasoning tokens, prompt caching - โœ… TTFT/ITL split, long-context superlinear attention - โœ… Long-context capacity planner + topology recommendation - โœ… Build-vs-Buy calculator (13 API providers with retired/current status) - โœ… Reference catalog (6 sub-tables including live pricing) - โœ… **Live GPU pricing** (Azure + AWS + GCP + Vast.ai parallel fetch with LIVE badges) - โœ… **Observed benchmark calibration** (paste vLLM/SGLang/TRT-LLM JSON โ†’ verdict) - โœ… **MCP server v0.2.0** โ€” 7 read-only tools, stdio + Streamable HTTP transports - โœ… **Public hosted MCP endpoint** at `tokcalc.vercel.app/api/mcp` with bearer auth + Upstash Redis rate limiting - โœ… **Self-serve API key generation** at `tokcalc.vercel.app/mcp` (email โ†’ instant key) - โœ… **4 SEO landing pages**: `/compare/h100-vs-h200`, `/compare/gguf-q4-k-m-vs-q5-k-m`, `/self-host-vs-openai-api`, `/mcp` (install docs) - โœ… Share URL + localStorage persistence - โœ… Copy result as Markdown (for GitHub issues / Slack / docs) - โœ… Dark mode toggle - โœ… Plain-English glossary (28 terms with hover tooltips) - โœ… OG image + Twitter card + social metadata - โœ… Plausible Analytics (privacy-friendly) + Sentry error monitoring ### Next 30 days - โณ GitHub Action (`tokcalc/plan` PR comment) - โณ i18n: Chinese, Japanese, Korean - โณ MCP server v0.2.1: Fix empty `inputSchema` in `tools/list` response (zod-to-json-schema serialization issue) - โณ MCP server v0.3.0: OAuth 2.1 with PKCE for multi-user auth ### Next 90 days - โณ Workload-trace / SLO capacity planner (prompt/output/arrival distributions, p50/p95 TTFT/ITL) - โณ P/D disaggregation planner (separate prefill + decode pools) - โณ Cache-aware economics (prefix-sharing distribution, multi-turn/agent traces) - โณ Engine-aware presets (vLLM / SGLang / TensorRT-LLM / llama.cpp) โ€” UI integration - โณ Versioned price + benchmark provenance system ### Long-term - ๐Ÿ”ฎ Agentic workflow calculator (multi-turn + tool calls + growing context) - ๐Ÿ”ฎ Multi-LoRA capacity planner (Punica / S-LoRA economics) - ๐Ÿ”ฎ VLM image-token accounting (per-model patch/tile tokenization) - ๐Ÿ”ฎ Embedding model mode (vectors/sec, separate workload) - ๐Ÿ”ฎ Training/fine-tuning estimator (LoRA / QLoRA / full-SFT FLOPs) - ๐Ÿ”ฎ Energy / carbon per million tokens (region-specific grid intensity) ## Open core model tokcalc is **open core** โ€” the calculator and catalog are open source; the cloud/data/team features are paid. | Asset | License | Notes | |---|---|---| | Source code | Apache 2.0 | This repo. Free to use, modify, distribute | | Model/GPU/quant catalog | CC0 1.0 | Public domain data. Anyone can use, no attribution required | | Benchmark provenance data | CC-BY-SA 4.0 | Anyone can use, but must attribute + share-alike | | Documentation | CC-BY 4.0 | Attribution required if copied | | "tokcalc" name + logo | Trademark | Even without formal registration, common-law rights apply | | Cloud SaaS layer | Proprietary | Real-time pricing API, benchmark DB, team workspaces (coming soon) | ### Why this structure - **Trust**: Open-source formulas build credibility vs opaque competitors - **Community**: Contributors can submit models, GPUs, quants, benchmarks - **Defensibility**: Trademark + cloud features + URL-share viral loop protect against forks - **Revenue**: Cloud tier funds ongoing development + pricing/benchmark data maintenance ## Contributing We welcome contributions! See [CONTRIBUTING.md](https://github.com/stevecrates489-commits/tokcalc/blob/HEAD/CONTRIBUTING.md) for: - How to add a model (with HuggingFace config.json as source) - How to add a GPU (with critical guardrails for B200/B300 null FP16 fields) - How to add a quantization format (with measured file sizes as source) - How to submit a benchmark (3-tier confidence model) - How to improve a formula (cite the source, no magic numbers) ### Most-needed contributions - ๐ŸŸข New models (Qwen 3, Mistral Large 4, Llama 4 variants as they release) - ๐ŸŸข New GPUs (B300, AMD MI400, Apple M5 Ultra when shipping) - ๐ŸŸข New quantization formats (mxFP8, MXFP4, BitNet 2) - ๐ŸŸข Real benchmark data (run vLLM benchmarks and submit with provenance) - ๐ŸŸข Translations (especially Chinese, Japanese, Korean) ## MCP server โ€” use tokcalc from AI agents tokcalc ships an [MCP (Model Context Protocol)](https://modelcontextprotocol.io) server that lets AI agents (Cursor, Claude Desktop, Cline) call tokcalc during design reviews. **npm package**: [`@tokcalc/mcp-server`](https://www.npmjs.com/package/@tokcalc/mcp-server) (latest: v0.2.0) **Public endpoint**: `https://tokcalc.vercel.app/api/mcp` (Streamable HTTP, bearer auth, rate-limited) **Install docs**: `https://tokcalc.vercel.app/mcp` ### 7 read-only tools | Tool | What it does | |---|---| | `estimate_capacity` | VRAM/KV/throughput/latency/cost for one config | | `compare_gpus` | Ranked GPU comparison for one workload | | `recommend_topology` | TP/CP topology recommendation | | `estimate_api_vs_self_host` | Break-even analysis | | `list_models` | Discover supported model IDs | | `list_gpus` | Discover supported GPU IDs | | `get_mlperf_benchmarks` | Curated MLPerf Inference v4.1 audited reference configs | All tools are **read-only** โ€” no side effects, no cloud credentials, no deployments. ### Option A: Public hosted endpoint (recommended โ€” no install) Point Cursor or Claude Desktop at the public endpoint. Get an instant API key at [`/mcp`](https://tokcalc.vercel.app/mcp): ```json { "mcpServers": { "tokcalc": { "url": "https://tokcalc.vercel.app/api/mcp", "headers": { "Authorization": "Bearer " } } } } ``` Features: - โœ… **Bearer API key auth** (constant-time comparison via `crypto.timingSafeEqual`) - โœ… **KV-backed rate limiting** via Upstash Redis (30 req/min per IP, 120 req/min per key) - โœ… **Self-serve key generation** โ€” enter email at `/mcp`, get instant key (5 per IP per day limit) - โœ… Keys expire after 90 days of inactivity - โœ… SHA-256 key hashing (never stored raw) ### Option B: Self-hosted stdio (for local Cursor / Claude Desktop) ```json { "mcpServers": { "tokcalc": { "command": "npx", "args": ["-y", "@tokcalc/mcp-server"] } } } ``` No API key needed for stdio (local install). Cursor spawns the process via `npx`. ### Option C: Self-hosted HTTP (for air-gapped / custom infra) ```bash npm install -g @tokcalc/mcp-server tokcalc-mcp-http # defaults to port 3000 ``` Set `MCP_API_KEY`, `UPSTASH_REDIS_REST_URL`, `UPSTASH_REDIS_REST_TOKEN` env vars for auth + rate limiting. ### Example agent prompt > "I need to serve Llama 3.3 70B at 32K context for 50 concurrent users. What GPU topology do you recommend, and how much will it cost per month?" The agent calls `list_models` โ†’ `list_gpus` โ†’ `recommend_topology` โ†’ `estimate_capacity` โ†’ returns a structured plan with VRAM, throughput, latency, cost, and confidence. ## Tech stack - **Framework**: Next.js 16 with App Router (Turbopack) - **Language**: TypeScript 5 - **Styling**: Tailwind CSS 4 + shadcn/ui (New York) - **Charts**: Recharts - **State**: React hooks (useState + useEffect + useMemo) - **Theme**: next-themes (dark mode default) - **MCP SDK**: `@modelcontextprotocol/sdk` v1.30.1 (Streamable HTTP transport, stateless mode) - **Auth**: `crypto.timingSafeEqual` (constant-time bearer key validation) - **Rate limiting**: `@upstash/ratelimit` + `@upstash/redis` (sliding window, 30/min per IP, 120/min per key) - **Analytics**: Plausible (privacy-friendly, no cookies) - **Error monitoring**: Sentry - **Database**: None for core; Upstash Redis for API key storage + rate limiting ## Project structure ``` tokcalc/ โ”œโ”€โ”€ src/ โ”‚ โ”œโ”€โ”€ app/ โ”‚ โ”‚ โ”œโ”€โ”€ api/ โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ mcp/route.ts # Public MCP endpoint (Streamable HTTP) โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ keys/route.ts # Self-serve API key generation โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ pricing/ # Live GPU pricing proxies โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ aws/route.ts โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ azure/route.ts โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ gcp/route.ts โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ vast-ai/route.ts โ”‚ โ”‚ โ”œโ”€โ”€ compare/ # SEO landing pages โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ h100-vs-h200/page.tsx โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ gguf-q4-k-m-vs-q5-k-m/page.tsx โ”‚ โ”‚ โ”œโ”€โ”€ self-host-vs-openai-api/page.tsx โ”‚ โ”‚ โ”œโ”€โ”€ mcp/page.tsx # MCP install docs + self-serve key form โ”‚ โ”‚ โ”œโ”€โ”€ page.tsx # Main calculator โ”‚ โ”‚ โ””โ”€โ”€ layout.tsx โ”‚ โ”œโ”€โ”€ lib/ โ”‚ โ”‚ โ”œโ”€โ”€ mcp/ โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ core.ts # UNIFIED tool core โ€” all 11 tools, both surfaces โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ server.ts # thin shim โ†’ core (createMcpServer factory) โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ mcp-version.ts # MCP_SERVER_VERSION โ€” single source of truth โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ mcp-store.ts # per-install store (spec cache + calibration) โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ auth.ts # validateApiKey (Redis + env var) โ”‚ โ”‚ โ”œโ”€โ”€ token-calc.ts # Core calculation engine โ”‚ โ”‚ โ”œโ”€โ”€ mlperf-curated.ts # 9 MLPerf v4.1 reference configs โ”‚ โ”‚ โ”œโ”€โ”€ price-schema.ts # PriceRecord type + helpers โ”‚ โ”‚ โ”œโ”€โ”€ benchmark-schema.ts # BenchmarkRecord + CalibrationResult โ”‚ โ”‚ โ”œโ”€โ”€ engine-presets.ts # 5 engine presets (vLLM/SGLang/TRT-LLM/llama.cpp/generic) โ”‚ โ”‚ โ””โ”€โ”€ url-state.ts # URL hash + localStorage persistence โ”‚ โ””โ”€โ”€ components/ โ”‚ โ”œโ”€โ”€ benchmark-import.tsx # Paste JSON โ†’ calibrate against estimate โ”‚ โ””โ”€โ”€ confidence-badge.tsx # ๐ŸŸข๐ŸŸกโšช confidence dots โ”œโ”€โ”€ mini-services/ โ”‚ โ””โ”€โ”€ mcp-server/ # Standalone npm package (for stdio + self-hosted HTTP) โ”‚ โ”œโ”€โ”€ server.ts # thin shim โ†’ src/lib/mcp/core.ts (no duplicated handlers) โ”‚ โ”œโ”€โ”€ auth.ts # Bearer auth middleware (Node http) โ”‚ โ”œโ”€โ”€ rate-limit.ts # Upstash rate limiter (Node http) โ”‚ โ”œโ”€โ”€ http.ts # Standalone HTTP entry (tokcalc-mcp-http) โ”‚ โ”œโ”€โ”€ index.ts # Stdio entry (tokcalc-mcp-server) โ”‚ โ””โ”€โ”€ package.json # @tokcalc/mcp-server npm package โ”œโ”€โ”€ public/ โ”‚ โ”œโ”€โ”€ sitemap.xml # 5 URLs (/, /mcp, 3 compare pages) โ”‚ โ”œโ”€โ”€ og.png # OG image โ”‚ โ””โ”€โ”€ manifest.json # PWA manifest โ”œโ”€โ”€ server.json # MCP registry entry (packages + remotes) โ”œโ”€โ”€ package.json # Web app deps โ””โ”€โ”€ next.config.ts ``` ## Acknowledgments tokcalc builds on the work of: - **vLLM team** โ€” PagedAttention, continuous batching ([arxiv.org/abs/2309.06180](https://arxiv.org/abs/2309.06180)) - **llama.cpp / ggml-org** โ€” GGUF format and quantization variants - **MLPerf / MLCommons** โ€” standardized inference benchmark methodology - **Stanford HELM** โ€” efficiency-aware model evaluation framework - **Anthropic / OpenAI / Google** โ€” published prompt-caching pricing rules - **Model Context Protocol team** โ€” MCP Streamable HTTP transport spec (2025-03-26) - **NVIDIA / AMD / Intel / Google / Groq / Cerebras** โ€” published hardware specs - **Upstash** โ€” Redis + rate limiting infrastructure Every formula has a citation. Every model/GPU/quant entry has a source URL where available. If you spot an unsourced claim, please open an issue. ## Star history [![Star History Chart](https://api.star-history.com/svg?repos=stevecrates489-commits/tokcalc&type=Date)](https://star-history.com/#stevecrates489-commits/tokcalc&Date) ---
**[Live demo](https://tokcalc.vercel.app)** ยท **[MCP install docs](https://tokcalc.vercel.app/mcp)** ยท **[Public MCP endpoint](https://tokcalc.vercel.app/api/mcp)** ยท **[npm package](https://www.npmjs.com/package/@tokcalc/mcp-server)** ยท **[Contributing](https://github.com/stevecrates489-commits/tokcalc/blob/HEAD/CONTRIBUTING.md)** ยท **[License](https://github.com/stevecrates489-commits/tokcalc/blob/HEAD/LICENSE)** ยท **[Code of Conduct](https://github.com/stevecrates489-commits/tokcalc/blob/HEAD/CODE_OF_CONDUCT.md)** Made with care by the tokcalc community. Apache 2.0 licensed.