@trimly/core

Tokenization, pricing, currency conversion, local storage, prompt-optimization strategies, and multi-agent adapters for Trimly.

The engine behind Trimly — token counting and cost computation for Anthropic, OpenAI, and Mistral, a local SQLite store, filler/prompt-slimming strategies in 24 EU languages, and a factory of agent adapters (Claude Code, Codex, Cursor). Used by @trimly/cli and the Trimly plugin, but usable standalone.

Install

npm install @trimly/core

Usage

Count tokens

import { countTokens } from '@trimly/core'

countTokens('anthropic', 'claude-sonnet-4-6', 'Hello, world!') // → number
countTokens('openai', 'gpt-4o', 'Hello, world!')

Compute and format cost

import { computeCost, convertCost, formatCost } from '@trimly/core'

const usage = { input_tokens: 1200, output_tokens: 350 }
const usd = computeCost('anthropic', 'claude-sonnet-4-6', usage) // → number (USD)

const eur = convertCost(usd, 'EUR') // USD → EUR (configurable rate, default 0.92)
formatCost(eur, 'EUR', 'fr-FR') // → "0,01 €"

TokenUsage also accepts cache_read_input_tokens and cache_creation_input_tokens for cache-aware pricing.

Local storage

import { createStorage, getDefaultDbPath } from '@trimly/core'

// Picks node:sqlite when available, falls back to libSQL
const storage = await createStorage(getDefaultDbPath())
await storage.recordEvent(/* TrimlyEvent */)
const events = await storage.queryEvents({ days: 30 })

Optimization strategies

import { cleanFiller, analyzePrompt } from '@trimly/core'

cleanFiller(text)     // strip filler phrases — 24 EU languages, auto-detected
analyzePrompt(text)   // detect heavy code blocks, stack traces, etc.

Agent adapters

A factory resolves the host agent (from config override or env detection) and abstracts each agent's hook payloads, output protocol, and model resolution.

import { resolveAgent, detectAgent } from '@trimly/core'

const adapter = resolveAgent('auto') // 'claude-code' | 'codex' | 'cursor'
adapter.source            // event source for storage
adapter.capabilities      // { promptOptimization, costTracking, toolTracking }
adapter.parsePayload('UserPromptSubmit', rawStdinJson)
adapter.formatOutput({ context: '…' }, 'UserPromptSubmit')

Browser build

A dependency-light entry point (no WASM, no Node APIs) is available for browser/extension use:

import { countTokens } from '@trimly/core/browser'

License

Apache 2.0 © JOYS LAB