ModelCostSaver

Predict the cost of an LLM call before you make it, and pick the cheapest model that still does the job, offline, from your editor.

ModelCostSaver is a Model Context Protocol server. It gives any AI coding agent or IDE a free, zero-config tool that answers the three questions every agent should ask before an LLM call:

1. What will this prompt cost on each candidate model? (predict_cost, estimate_cost)
2. Which is the cheapest model that meets the task? (select_optimal_model)
3. How do my options compare side by side? (compare_models)

It is pure pricing-and-routing math over a bundled, dated catalog, so the core needs no API keys and makes no network calls.

Quick start

Run it directly with npx (no install, no keys):

npx -y @workswarm/modelcostsaver

Or write the config for your editor in one command:

npx -y @workswarm/modelcostsaver install --client cursor

Add to Cursor — one click installs it in Cursor. Or drop the block below into ~/.cursor/mcp.json, or run npx -y @workswarm/modelcostsaver install --client cursor.

Listed on the official MCP registry and editor MCP directories as io.github.sachinuppal/modelcostsaver.

The seven tools

Every cost-bearing result carries catalogVersion and asOf so you can see how fresh the prices are. Every selection carries a reasoning array, never a black-box pick.

Trust: no keys, offline, no telemetry

For a tool that sits in your editor, trust is the whole pitch. ModelCostSaver is:

• No API keys. The core does pricing math, not provider calls. Nothing to leak.
• Offline by default. The core tools return correct answers with no network access. The only outbound request is an opt-in catalog refresh (MODELCOSTSAVER_REFRESH=on), a single GET of a static JSON, zod-validated before it can replace the bundled catalog, and it always falls back to the bundle on any failure.
• No telemetry. Ever. The default is silent and local. record_usage only writes when you set MODELCOSTSAVER_LEDGER=on, and only to a JSONL file under your own config dir.
• Two dependencies. @modelcontextprotocol/sdk and zod. Nothing else. Small supply-chain surface, fast npx cold start.
• Apache-2.0. An open-source developer tool published by Workswarm as @workswarm/modelcostsaver. The shipped bundle contains no proprietary or internal-service code: no internal-framework imports and no internal identifiers, just dependency-free pricing-and-routing math.

stdout carries only JSON-RPC; all logs go to stderr.

Install per IDE

ModelCostSaver speaks stdio MCP, so the entry is the same npx command everywhere. Use install --client <name> to write it idempotently, or paste the block by hand.

Cursor

~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{ "mcpServers": { "modelcostsaver": { "command": "npx", "args": ["-y", "@workswarm/modelcostsaver"] } } }

npx -y @workswarm/modelcostsaver install --client cursor

Claude Code

claude mcp add modelcostsaver -- npx -y @workswarm/modelcostsaver

or a .mcp.json in the repo root (which install --client claude writes):

{ "mcpServers": { "modelcostsaver": { "command": "npx", "args": ["-y", "@workswarm/modelcostsaver"], "env": { "MODELCOSTSAVER_PROVIDERS": "anthropic" } } } }

Claude clients run Claude for their own inference, so the install seeds MODELCOSTSAVER_PROVIDERS=anthropic as a sensible default for target: self recommendations. Override it per call or with the env var. See Self vs code.

Claude Desktop

Add the same mcpServers block to claude_desktop_config.json.

VS Code / GitHub Copilot

.vscode/mcp.json:

{ "servers": { "modelcostsaver": { "command": "npx", "args": ["-y", "@workswarm/modelcostsaver"], "type": "stdio" } } }

npx -y @workswarm/modelcostsaver install --client vscode

Windsurf

~/.codeium/windsurf/mcp_config.json with the same mcpServers block, or:

npx -y @workswarm/modelcostsaver install --client windsurf

Cline / Zed / Antigravity

Same stdio command/args. Use the matching installer:

npx -y @workswarm/modelcostsaver install --client cline
npx -y @workswarm/modelcostsaver install --client zed
npx -y @workswarm/modelcostsaver install --client antigravity

After adding the server, restart the client and confirm the seven tools appear in the tool list.

Two axes: self vs code

ModelCostSaver advises; it does not route traffic. So every recommendation is filtered to what you can actually act on, along two independent axes.

Axis 1, availability. Recommendations are scoped to a set of allowed providers. The default is derived from the connected client (read from the MCP handshake): a Claude client defaults to anthropic because its own inference is Claude; multi-provider clients (Cursor, VS Code, Windsurf, Cline, Zed, Antigravity) and unknown clients default to all providers. The scope and its source are always echoed in reasoning, and it is overridable: a per-call providers arg, then MODELCOSTSAVER_PROVIDERS, then config, then the client default, then all.

Axis 2, target.

• target: 'self' (default): the agent's or your own next inference in this client. The Axis-1 scope applies. In Claude Code this means cross-tier Anthropic moves (Opus to Haiku), which you can act on right now.
• target: 'code': a model you will call from your own application, where you supply that provider's key. The client scope does not apply, so all in-catalog providers are eligible.

ModelCostSaver is always honest about the gap: if the globally-cheapest model is outside your actionable set, it is surfaced as cheaperIfAvailable with the reason, never silently chosen. For example, a Claude Code target: self summarize call selects claude-haiku-4-5 and notes that a cheaper non-Anthropic model exists if you pass target: code.

How it predicts

1. Tokens. Exact counts if you supply them; otherwise a heuristic estimate (~4 chars/token, tunable via MODELCOSTSAVER_CHARS_PER_TOKEN). The heuristic is approximate but common-mode across candidates, which is what relative ranking needs. Output tokens come from your explicit value, then the task class cap, then the model cap, then a conservative default.
2. Cost. (inTok / 1e6) * inputPerMillion + (outTok / 1e6) * outputPerMillion, in full-precision USD and as integer usdMicros. A prediction is never rounded to cents.
3. Selection. Resolve the target tier (from an explicit taskClass, else a transparent keyword/length classifier), filter candidates by tier (degrade up, never below the floor), capabilities, and provider scope, forecast each, drop those over budget into rejected, and pick the cheapest survivor. Every step is recorded in reasoning, and a fallbackChain is returned for retry-on-failure.

Configuration

All config is optional. Precedence: tool-call arg, then env var, then modelcostsaver.config.json (cwd, then your user config dir), then the built-in default.

Pricing data

Prices change often, so ModelCostSaver ships a versioned, dated seed and is honest about its freshness.