Open source · MIT · on npm
Anouk is a small, friendly framework for shipping browser extensions that call LLMs. Plug in OpenAI, Anthropic, Together.xyz, Ollama, or Hugging Face. Get caching, rate-limit queueing, and a settings panel for free.
# scaffold a new AI extension
$ npm install -g anouk
$ anouk init my-ai-extension
$ cd my-ai-extension && npm run build
# load unpacked in chrome://extensions and you're live
5
provider profiles documented in the README
1
command to scaffold a runnable extension
MIT
license — fork, ship, sell on the store
What is Anouk?
Anouk is two things in one npm package. First, a runtime library — AIService, configManager, and settingsPanel — that you import into any Manifest V3 extension to add AI features. Second, a CLI (anouk init) that scaffolds a working extension with the runtime pre-wired.
The bet: for a meaningful slice of AI features, the model call belongs in the extension, not on a backend you operate. The user supplies their own provider URL, key, and model in the settings panel; the extension calls the provider directly over fetch. No server on the critical path.
Lower latency
No extra hops through a backend you host.
Stronger privacy
Content never crosses a server the developer controls.
Zero operating cost
Static site + open source; the user pays their own provider.
What you get
Anouk speaks the OpenAI chat-completions shape, so OpenAI, Anthropic, Together.xyz, Ollama, and Hugging Face Inference all plug in with a URL, an API key, and a model name. No SDK juggling, no per-provider adapters in your extension code.
AIService caches responses by a request id and cache key you control, so repeated calls (re-opening the same email, re-running the same summary) skip the round trip and the bill.
Provider quotas are unforgiving. Anouk queues calls per-provider so a clicky user does not blow your token budget or earn a 429 in the middle of a sidebar render.
Pre-built settings panel that lets users set provider, key, model, and system prompt at runtime. Values persist in chrome.storage.local. You ship one bundle; the user chooses what to run against.
anouk init my-extension drops a working Manifest V3 project on disk with the AI service, config manager, and settings panel pre-wired. Load it unpacked in Chrome and iterate.
Already have an extension? npm install anouk and import { AIService }. Starting from zero? anouk init gives you a runnable extension that talks to a model on the first build.
Three lines to an AI feature
One AIService.call() handles config, caching, queueing, and the provider request. You get back a plain string.
import { AIService } from 'anouk';
// user set provider + key in the settings panel
const summary = await AIService.call(
'Summarize this page in 3 bullets',
document.body.innerText,
`summary_${location.href}`
);
renderSidebar(summary);1. reads providerUrl / apiKey / model from config
2. builds cache key `${requestId}_${cacheKey}`
3. returns cached answer if present // no spend
4. else enqueues per-provider (rate-limit safe)
5. POSTs OpenAI chat-completions shape via fetch
6. caches + returns the textNo SDK, no backend, no per-provider glue in your extension code.
Bring your own provider
Set providerUrl to any endpoint that speaks the OpenAI chat-completions shape.
How a request flows
01 · extension.js
Pass an instruction, a piece of content, and a stable request id. Anouk handles everything from there.
02 · cache check
If the same prompt has already been answered for this id, you get the cached response. No round trip, no spend.
03 · queue + dispatch
The request is queued per provider so simultaneous user actions do not stampede the API. Anouk speaks the OpenAI chat-completions shape.
04 · response
The model output is parsed, written to cache, and returned to your extension code as a plain string.
Honest comparisons
Plasmo and WXT are great at building extensions in general. Anouk is opinionated about the AI part. Pick the right tool — sometimes that's not us.
Plasmo is the popular open-source extension framework with React-first DX, file-based routing, and multi-browser builds. Anouk is narrower: it focuses on the AI service, caching, and settings panel rather than the build system.
WXT is a newer Vite-powered extension framework with first-class TypeScript and multi-browser targets. Anouk is not a build tool; it is the AI runtime you would call from inside a WXT-built extension.
FAQ
The CLI scaffolds Chrome-style Manifest V3 extensions with host_permissions, storage, and unlimitedStorage wired up. The current template is built and tested against Chromium-based browsers; the README walks through loading via chrome://extensions. Firefox and Safari WebExtensions are not currently demonstrated by the repo.
Anything that speaks the OpenAI chat-completions request shape. The README documents OpenAI, Anthropic, Together.xyz, Ollama (localhost:11434), and Hugging Face Inference endpoints, and lets you point providerUrl at anything else compatible.
No. AIService accepts config via code, but the bundled settings panel lets the end user supply providerUrl, apiKey, model, and a system prompt at runtime. You can ship without baking a key into the extension.
anouk init produces a Manifest V3 project with src/aiService.js, configManager.js, settingsPanel.js, and a starter extension entry point. anouk generate extension|service|config drops the same kind of file into an existing project.
Both. You can npm install anouk and import { AIService } into your existing extension to use only the request, cache, and queue layer. Or you can use anouk init as a complete scaffold.
It is published to npm as anouk, MIT licensed, and developed in the open on GitHub under skelf-research/anouk. Install instructions and CLI commands are linked from the docs site.
Documentation lives at docs.skelfresearch.com/anouk (path-based, not a subdomain). The README is a quickstart; the docs site is where examples, recipes, and the full configuration surface live.
Explore the docs
The full map of the site — dig into use cases, the architecture, the glossary, comparisons, and the blog.
The provider adapter, per-request cache, rate-limit queue, runtime settings panel, and CLI — one page.
ArchitectureConfig, cache check, per-provider queue, and a direct fetch to the model. The whole request path.
Get startedFrom npm install to an unpacked extension talking to a model, in six steps.
Use casesWriting assistant, page summarizer, local Ollama extension, and an inbox copilot — with code.
CompareHow Anouk sits next to Plasmo and WXT. It is the AI runtime, not the build framework.
ReferenceAIService, configManager, cache key, providerUrl, and the rest of the vocabulary defined.
FAQProviders, config, caching, browser support, and where the framework starts and stops.
BlogNotes on distribution, scaffolding, and why client-side AI belongs in the extension.
AboutWhy Skelf Research built a framework for the AI half of a browser extension.
Install anouk, scaffold an extension, point it at a provider, and spend your weekend on the part of your product that is actually new.