AALLM Docs

Quickstart

Find a compatible deployment and calculate a deterministic quote in under five minutes with TypeScript, Python, or the REST API.

Prerequisites

  • Node.js 20+ for the TypeScript SDK, or Python 3.11+ for the Python SDK.
  • An ALLM API key stored in ALLM_API_KEY.
  • A server-side environment. Never expose a secret key in browser code.

1. Install an SDK

Terminal · TypeScript
pnpm add @allm/sdk
Terminal · Python
pip install allm

You can skip installation and use the REST examples throughout the reference with curl.

2. Configure the client

TypeScript
import { ALLM } from "@allm/sdk";

const allm = new ALLM({ apiKey: process.env.ALLM_API_KEY });
Python
import os
from allm import ALLM

allm = ALLM(api_key=os.environ["ALLM_API_KEY"])

3. Find a compatible deployment

TypeScript
const result = await allm.deployments.list({
  model: "anthropic/claude-haiku-4.5",
  capability: "tool_calling",
  region: "global",
});

const deployment = result.data[0];
if (!deployment) throw new Error("No compatible deployment found");

console.log(deployment.id);
// dep_anthropic_claude_haiku_4_5_messages_global

Filter semantics

The capability filter returns only deployments whose assertion is supported. It excludes conditional, unknown, and unsupported.

4. Calculate a quote

TypeScript
const quote = await allm.pricing.calculate({
  deployment: "dep_openai_gpt_5_1_responses_global",
  inputTokens: 1_000_000,
  outputTokens: 100_000,
});

console.log(quote.totalUsd); // "2.250000"
console.table(quote.lineItems);
Console output
┌─────────────────────┬──────────┬────────────┐
│ dimension           │ quantity │ amountUsd  │
├─────────────────────┼──────────┼────────────┤
│ input_tokens        │ 1000000  │ 1.250000   │
│ cached_input_tokens │ 0        │ 0.000000   │
│ output_tokens       │ 100000   │ 1.000000   │
└─────────────────────┴──────────┴────────────┘

Money values are exact decimal strings. Keep them as strings or parse them with a decimal library.

5. Production checklist

  • Handle structured errors and retain x-request-id for diagnostics.
  • Use timeouts and retry only 429, transient 5xx, and network failures.
  • Store the selected deployment with x-allm-catalog-release.
  • Pin important decisions in a lockfile and review catalog drift.