11 planning tools for AI agents (Cursor, Claude Desktop, Cline). Ask your AI: "How many H100s do I need for Llama 70B at 32K context?" — get a transparent, formula-based answer with confidence intervals.
record_measured) that stores a real measurement so future estimates self-calibrate. Every tool declares its access level in the Access column below, and in the MCP behaviour hints on the wire.| Tool | What it does | Access |
|---|---|---|
| estimate_capacity | VRAM/KV/throughput/latency for one config. Returns feasibility, performance ranges, cost, confidence. | read-only |
| compare_gpus | Ranked GPU comparison for one workload. Sort by lowest cost, highest throughput, or best value. | read-only |
| recommend_topology | TP/CP topology recommendation. Returns feasible GPU + topology designs for a context length. | read-only |
| estimate_api_vs_self_host | Break-even analysis. Compares monthly API costs vs self-hosted GPU infrastructure. | read-only |
| list_models | Discover supported model IDs. Filter by family, category, MoE. | read-only |
| list_gpus | Discover supported GPU IDs. Filter by vendor, category, min VRAM. | read-only |
| get_mlperf_benchmarks | Curated MLPerf Inference v4.1 audited configs. Cross-validate estimates against reference systems. | read-only |
| find_config_for_slo | Inverse planner. Given model + context + traffic + SLO ceilings (TTFT, $/M, tok/s), searches every GPU × topology and returns feasible configs ranked by cost, throughput, or value. | read-only |
| plan_deployment | One-call decision brief: memory + KV headroom, performance, rig cost, blended build-vs-buy with break-even, concrete risks, and next steps. | read-only |
| fetch_model_spec | Fetches the model's real config.json from HuggingFace and diffs it against tokcalc's catalog (layers, KV heads, head_dim). Catches catalog drift before it skews KV math. 24h cache. | read-only |
| record_measured | Calibration loop. Record real tok/s / TTFT from your serving run; future estimates for that model+GPU+quantization are annotated with your measured/predicted ratio. | writes state |
~/.cursor/mcp.json (global) or .cursor/mcp.json (project)~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows){
"mcpServers": {
"tokcalc": {
"command": "npx",
"args": ["-y", "@tokcalc/mcp-server"]
}
}
}Restart your editor. The 11 tools will appear in your MCP tool list within seconds. No API key needed for stdio (local install).
# Install globally (or use npx)
npm install -g @tokcalc/mcp-server
# Run the HTTP server (defaults to port 3000)
tokcalc-mcp-http
# Or with a custom port:
PORT=8080 tokcalc-mcp-httpMCP_API_KEY, UPSTASH_REDIS_REST_URL, and UPSTASH_REDIS_REST_TOKENenv vars before deploying to your own server.