Call Chrome's built-in AI APIs (Gemini Nano) from Node.js or Python.

Chrome's AI APIs only run inside browser pages. chrome-ai bridges that gap: a Python HTTP server manages a prompt queue, and a bridge page (open once in Chrome) processes prompts by calling the API directly.

Status: Alpha. Not yet published to npm. APIs may change, edge cases remain. Works for prototyping via local install.

Install

npm install chrome-ai-tools

Then use the chrome-ai CLI directly:

npx chrome-ai prompt "hello"
# or install globally
npm install -g chrome-ai-tools
chrome-ai summarize "some text"

Just Chrome. No agent-browser, no extensions, no API keys.

Quick Start

Start the server:

python3 server.py
# → http://localhost:8462

Open that URL in Chrome. Keep the tab open.

Now use the CLI:

npx chrome-ai prompt "What is the capital of France?"
npx chrome-ai summarize "Some long text here..."
npx chrome-ai translate --from en --to fr "Hello"
npx chrome-ai write "Write a poem about cats"
# or pipe text in
cat some-file.txt | npx chrome-ai summarize

Or from Node.js:

import { prompt, summarize, translate, write } from 'chrome-ai-tools';

const answer = await prompt({
  system: 'You are helpful.',
  user: 'What is the capital of France?',
});
const summary = await summarize('Long text...');
const french = await translate('Hello', 'en', 'fr');
const poem = await write('Write a haiku about dogs');

Or from Python:

from client import prompt, summarize, translate, write

answer = prompt('You are helpful.', 'What is the capital of France?')
summary = summarize('Long text...')
french = translate('Hello', 'en', 'fr')
poem = write('Write a haiku about dogs')

Or from any HTTP client:

# Prompt API
curl -s -X POST http://localhost:8462/prompt \
  -H 'Content-Type: application/json' \
  -d '{"api":"prompt","system":"You are helpful.","user":"Hello!"}'

# Summarizer API
curl -s -X POST http://localhost:8462/prompt \
  -H 'Content-Type: application/json' \
  -d '{"api":"summarize","text":"Long text to summarize..."}'

# Translator API
curl -s -X POST http://localhost:8462/prompt \
  -H 'Content-Type: application/json' \
  -d '{"api":"translate","text":"Hello","sourceLanguage":"en","targetLanguage":"fr"}'

# Writer API
curl -s -X POST http://localhost:8462/prompt \
  -H 'Content-Type: application/json' \
  -d '{"api":"write","text":"Write a haiku","context":"dogs"}'

# Poll for result
curl -s http://localhost:8462/result/a1b2c3d4e5f6
# → {"status": "done", "text": "..."}

How it works

Client (Node / Python / curl / CLI)
    │
    ├─ POST /prompt {api, ...params} → {id}
    │
    ▼
Python HTTP server — manages prompt queue
    │
    ├─ GET /pending ← bridge page polls every 500ms
    │
    ▼
Bridge page (Chrome, open once)
    - dispatches to LanguageModel / Summarizer / Translator / Writer
    - POSTs result to /result/{id}
    │
    ▼
Client — poll GET /result/{id} → result

API

TypeScript

import { prompt, summarize, translate, write } from 'chrome-ai-tools';

// Prompt API
const text = await prompt({
  system?: string,  // system prompt (optional)
  user: string,     // user message (required)
});

// Summarizer API
const summary = await summarize(text: string);

// Translator API
const translated = await translate(
  text: string,
  sourceLanguage: string,
  targetLanguage: string,
);

// Writer API
const written = await write(text: string, context?: string);

Python

from client import prompt, summarize, translate, write

text = prompt(system: str, user: str, timeout: float = 120)

summary = summarize(text: str, timeout: float = 120)

translated = translate(
    text: str,
    source_language: str,
    target_language: str,
    timeout: float = 120,
)

written = write(text: str, context: str = "", timeout: float = 120)

CLI

chrome-ai prompt "What is the capital of France?"

chrome-ai summarize "Some text to summarize..."

chrome-ai translate --from en --to fr "Hello world"

chrome-ai write "Write a poem" --context "cats"

# Pipe text in
cat file.txt | chrome-ai summarize
cat file.txt | chrome-ai translate --from en --to es

Auto-starts the Python server if not already running.

Server Endpoints

Method	Path	Purpose
POST	/prompt	Submit a job {api, ...params} → {id}
GET	/pending	Bridge page polls for pending jobs
POST	/result/{id}	Bridge page submits result {status, text}
GET	/result/{id}	Client polls for result
GET	/health	{ok, port, pending}

POST /prompt params

api	Required params	Optional params
prompt	user	system
summarize	text	—
translate	text, sourceLanguage, targetLanguage	—
write	text	context

Requirements

• Recent desktop Chrome (v129+, Prompt API on by default)
• Python 3.9+

Publishing

npm run build
npm publish --otp <code>

The account has 2FA enabled, so publishing requires a one-time password from your authenticator app.

Needle Tool Routing

chrome-ai-tools can train, export, and test Needle models for tool routing — a small transformer that maps natural language queries to tool calls, running entirely in-browser via the needle-rs WASM runtime (~258 KB).

# Install Python deps, write tool schemas, generate training data, fine-tune
pip install -r scripts/requirements.txt
chrome-ai generate --tools tools.json --output data/needle
bash scripts/fine-tune.sh

# Test the exported model
chrome-ai needletest --model output/needle.safetensors

# Compare with Gemini Nano (requires chrome-ai bridge)
chrome-ai compare --needle-model output/needle.safetensors

See docs/needle.md for the full guide.

Also a Pi Skill

SKILL.md — installable via skills.sh install donpark/chrome-ai for use with Pi coding agent.