@logspine/sdk-ts

Drop-in observability SDK for AI agents. TypeScript / Node / Edge. Auto-instruments OpenAI and Anthropic.

Python SDK: pip install logspine — see Python docs

The Helicone replacement built for solo devs and small teams. One line of code, zero infrastructure, pricing that doesn't punish growth — start at $19/mo for 100k spans.

→ logspine.dev · Full docs · Quickstart · Data model

Install

pnpm add @logspine/sdk-ts
# or: npm install @logspine/sdk-ts
# or: yarn add @logspine/sdk-ts

Quick start

import { Logspine } from '@logspine/sdk-ts';

Logspine.init({
  apiKey: process.env.LOGSPINE_API_KEY!,
});

const trace_id = Logspine.trace({
  name: 'support-bot',
  input: { user_question: 'How do refunds work?' },
});

Logspine.span({
  trace_id,
  name: 'retrieve-docs',
  output: { hits: 4 },
});

// Before process exit:
await Logspine.shutdown();

Get your API key at logspine.dev → Settings → API Keys.

Auto-instrumentation

Wrap your existing client and every call is tracked automatically — with real latency, token counts, and cost in USD (computed from per-model pricing tables, updated quarterly).

OpenAI

import OpenAI from 'openai';
import { Logspine } from '@logspine/sdk-ts';
import { instrumentOpenAI } from '@logspine/sdk-ts/openai';

Logspine.init({ apiKey: process.env.LOGSPINE_API_KEY! });
const openai = instrumentOpenAI(new OpenAI());

// Now every call emits a trace:
const res = await openai.chat.completions.create({
  model: 'gpt-4o-mini',
  messages: [{ role: 'user', content: 'hi' }],
});

Supports:

• chat.completions.create (sync + streaming)
• embeddings.create
• Prompt caching (cached tokens billed at the cached rate automatically)

Anthropic

import Anthropic from '@anthropic-ai/sdk';
import { Logspine } from '@logspine/sdk-ts';
import { instrumentAnthropic } from '@logspine/sdk-ts/anthropic';

Logspine.init({ apiKey: process.env.LOGSPINE_API_KEY! });
const anthropic = instrumentAnthropic(new Anthropic());

Supports:

• messages.create (sync + streaming)
• Cache reads (cache_read_input_tokens billed at 10% of input rate)

API

| Method | Description | | --- | --- | | Logspine.init(opts) | Initialize. Required: apiKey. Optional: endpoint, flushIntervalMs (default 2000), maxBatch (default 100), projectId, disabled, onError. | | Logspine.trace({ name, input, output, model, prompt_tokens, completion_tokens, cost_usd, start_ts, end_ts, status, error, metadata }) | Create top-level trace. Returns trace_id. | | Logspine.span({ trace_id, name, ... }) | Child span inside a trace. | | Logspine.flush() | Force-flush the queue. | | Logspine.shutdown() | Stop timer + flush. Call before process exit. |

Spans are batched automatically: every 2 seconds, or when the queue hits 100 spans.

Full method reference, options table, and runtime support matrix: /docs/sdk/typescript.

Pricing table

@logspine/sdk-ts ships a model→price table covering:

• OpenAI: GPT-4o, GPT-4o-mini, GPT-4.1, GPT-4, GPT-3.5, o1, o1-mini, o3, o3-mini, text-embedding-3 (small/large), text-embedding-ada-002
• Anthropic: Claude 4 (opus/sonnet/haiku), Claude 3.5 (sonnet/haiku), Claude 3 (opus/sonnet/haiku), Claude 2

If a model isn't recognized (new release, fine-tuned variant), cost_usd will be 0 and a warning is not logged — the trace still goes through.

Prices are updated each quarter. Source: pricing.ts.

Edge / serverless runtimes

Works on:

• Node 18+
• Cloudflare Workers (nodejs_compat not required)
• Vercel Edge Functions
• Deno (via npm: specifier)
• Bun

The SDK uses fetch() and crypto.subtle — no native Node dependencies.

License

MIT © Adam Kallen — see LICENSE.