Modelcostsaver

What is ModelCostSaver?

ModelCostSaver is an MCP server that predicts the cost of an LLM call before you make it and selects the cheapest model that still meets the task requirements. It works offline from your editor, requires no API keys, and has no telemetry. It is intended for AI coding agents and IDEs that need cost-aware model routing.

How to use ModelCostSaver?

Run it directly with npx -y @workswarm/modelcostsaver — no install or keys needed. Add it to your editor (Cursor, Claude Code, VS Code, etc.) via the install --client <name> command or by pasting the stdio MCP config block. Once connected, the server exposes seven tools: estimate_cost, predict_cost, select_optimal_model, compare_models, list_models/get_pricing, optimize_request, and record_usage.

Key features of ModelCostSaver

• Predicts LLM call cost before making the call
• Selects cheapest model meeting task requirements
• Compares models side-by-side with cost table
• Works offline with no API keys needed
• No telemetry, two dependencies only
• Provides versioned, dated pricing catalog

Use cases of ModelCostSaver

• Forecasting the cost of a prompt across candidate models
• Selecting the optimal model for a given task and budget
• Comparing multiple models' costs for a fixed token shape
• Optimizing an existing model choice for lower cost
• Recording local usage records for tracking

FAQ from ModelCostSaver

Does ModelCostSaver need API keys?

No. The core does pricing math, not provider calls, so there is nothing to leak. No API keys are required.

Is ModelCostSaver offline by default?

Yes. Core tools return correct answers with no network access. An opt-in catalog refresh (MODELCOSTSAVER_REFRESH=on) can fetch a static JSON, but falls back to the bundled catalog on any failure.

Does ModelCostSaver collect telemetry?

No. The default is silent and local. The record_usage tool only writes when MODELCOSTSAVER_LEDGER=on is set, and only to a JSONL file under your own config directory.

How does ModelCostSaver predict token counts?

Exact token counts if you supply them; otherwise a heuristic estimate (~4 characters per token, tunable via MODELCOSTSAVER_CHARS_PER_TOKEN). Output tokens come from an explicit value, the task class cap, the model cap, or a conservative default.

How fresh is the pricing data?

The bundled catalog has a catalogVersion and asOf date on every entry. The seed is re-checked against each provider's public pricing page at release. The asOf date tells you when the data was last confirmed.