Anima ADK

A TypeScript framework for production multi-agent AI systems — schema-first, event-sourced, and provider-agnostic. Built and used in production at Anima to run clinical and operational agent workflows.

Documentation: adk.animahealth.com — runnable docs: every code cell on the site executes this shipped package in your browser.

The ADK is a library, not a platform: no infrastructure dependency beyond a database you choose (or none — the in-memory and SQLite backends need nothing). Every layer is independently useful.

• One session ledger. Everything an agent does is an append-only event history; all state derives from it. Time travel, forking, and auditing come for free.
• Agents that stop and ask. Tools can yield: the run pauses, the session persists as a row, and a human (or another system) resumes it — minutes or days later.
• The model sees what you choose. Context renderers are the only bridge between the ledger and the prompt, so full audit coexists with a small context.
• Deterministic testing. The test kit replaces only the model with scripted turns — tools really execute, state really writes — so agent tests run with no API key.
• Providers behind subpaths. OpenAI, Gemini, Claude (via Vertex), and EUrouter are optional integrations behind @animahealth/adk/openai and friends. Importing the core pulls in none of them.

Install

npm install @animahealth/adk

Node ≥ 22. zod is the one required peer (npm and pnpm install it automatically). The ADK accepts Zod 3 Classic from 3.25.76 and Zod 4 Classic from 4.6.5; an existing Zod 3 application does not need to migrate to Zod 4. Optional backends declare optional peers — install them only for what you use (for example better-sqlite3 for the SQLite session store).

First run — no API key

The real agent loop over a scripted model turn. Tools execute, state writes, the ledger accrues; only the model is played by the script.

import { getLastAssistantText, runTest, user, model, mockAgent } from '@animahealth/adk/testing'

const greeter = mockAgent('greeter')
const result = await runTest(greeter, [user('hi'), model('Hello! Ask me anything.')])

console.log(getLastAssistantText(result.events)) // Hello! Ask me anything.
console.log(result.events.map((e) => e.type))
// [ 'user', 'invocation_start', 'model_start', 'assistant', ... ]

Quick start — live

export OPENAI_API_KEY=sk-...

import { z } from 'zod'
import { adk } from '@animahealth/adk'
import { openai } from '@animahealth/adk/openai'

const app = adk()

const calculator = app.tool({
  name: 'calculate',
  description: 'Evaluate a mathematical expression',
  schema: z.object({
    expression: z.string().describe('Math expression to evaluate'),
  }),
  execute: (ctx) => {
    const sanitized = ctx.args.expression.replace(/[^\d\s+\-*/().eE%]/g, '')
    const result = Function(`"use strict"; return (${sanitized})`)()
    return { result }
  },
})

const assistant = app.agent({
  name: 'math_assistant',
  model: openai('gpt-5.6-luna'),
  context: [
    app.context.system(`You are helpful, use the calculator tool for arithmetic.`),
    app.context.history(),
  ],
  tools: [calculator],
})

const result = await app.run(assistant, 'What is 134 divided by 4?')
console.log(result.output.text) // 134 divided by 4 is 33.5.

Gemini ships behind @animahealth/adk/gemini (AI Studio or Vertex) and Claude behind @animahealth/adk/claude (Vertex AI). Provider setup lives in each subpath's typed config.

EUrouter ships behind @animahealth/adk/eurouter for hosted models such as DeepSeek, Kimi, and GLM. See the provider guide for credentials, routing, and a local example that needs no key.

Evaluation CLI

Text and voice cases can share one evaluation and one CLI entry point:

async function main() {
  process.exitCode = await app.evaluate.cli([...textCases, ...voiceCases], {
    concurrency: 4,
  })
}
void main()

Call the entry script with list, or run --case <name> --repeat 3 --output ./runs. The CLI writes JSON to stdout, diagnostics to stderr, and reports and case evidence to a fresh run directory. Use silent package scripts when parsing stdout. It exits successfully only when every selected execution passes. Normal imports still run before the CLI, so keep them quiet.

Use app.evaluate(cases, options) for the same mixed suite without command-line handling. Voice-specific hooks, metrics and room configuration belong in options.voice. See the mixed voice example.

For native voice with ADK backend tools, GPT Live voice handler documents openai.live(...) with app.handler.voice(...), typed result hooks, and transcript persistence.

For speech that has no structural signal, the judge metric documents app.evaluate.judge(...), an LLM-judge metric for text and voice cases, and when to use a state or event metric instead.

Execution completion and persistence

app.run() starts execution immediately. Awaiting it returns its outcome; a run deadline or run.abort() rejects promptly. run.settled always fulfills after ADK-owned execution and cleanup finish and their ledger events are buffered. Use it before committing or closing a session store:

const run = app.run(assistant, { session, timeout: 30_000 })
try {
  return await run
} finally {
  run.abort()
  await run.settled
  await app.sessions.commit(session)
}

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗