SavvyTokens CLI

SavvyTokens CLI connects local LLM builds to SavvyTokens pricing, estimate, snapshot, usage telemetry, and MCP tools.

Install

npm install -g @savvysignalsllc/savvytokens

You can also run it without a global install:

npx @savvysignalsllc/savvytokens prices

Login

Create a scoped API key in SavvyTokens, then point the CLI at the hosted API:

savvytokens login --token aisp_your_key --api-base https://savvytokens.io/api/v1

For the pre-Clerk owner setup, the same token can manage workspace API keys when its configured hash includes owner:write. Regular customer ingestion keys should use only usage:write and pricing:read.

Common Commands

savvytokens prices
savvytokens compare --input 1000000 --output 250000
savvytokens init --workspace demo-workspace --project demo-project
savvytokens estimate SPEC.md BLUEPRINT.md TODOS.md --json
savvytokens snapshot SPEC.md BLUEPRINT.md TODOS.md --local --json
savvytokens usage send --provider openai --model gpt-5.4-mini --input 100000 --output 25000 --feature build --environment local
savvytokens usage rankings --task coding
savvytokens usage forecast-actual --model gpt-5.4-mini --expected 12.50 --actual 10.80 --task coding
savvytokens usage recommendation-outcome --current gpt-5.4 --recommended gpt-5.4-mini --outcome accepted --task coding
savvytokens mcp start

The legacy aispend command is still published as an alias for existing local scripts.

Telemetry Boundary

savvytokens usage send sends structured usage telemetry: provider, model, token counts, request count, timestamp, environment, feature/team/cost-center labels, latency/retry status when supplied, and allowlisted metadata. Logged-in CLI and MCP commands also send small command-level telemetry events so SavvyTokens can improve forecast accuracy and model recommendations.

SavvyTokens does not collect prompts, completions, message bodies, secrets, personal identifiers, session IDs, trace IDs, IP addresses, or raw request/response bodies through the CLI telemetry flow. Set AISPEND_CLI_TELEMETRY=off or AISPEND_MCP_TELEMETRY=off to disable automatic command telemetry.