待翻譯：A proposed browser standard for provider-agnostic AI inference

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Notifications You must be signed in to change notification settings Fork 1 Star 5 BranchesTags Open more actions menu Latest commit History 70 Commits 70 Commits Folders and files NameName Last commit message Last commit date .cursor/skills/ship-ipa-tools .cursor/skills/ship-ipa-tools .github/workflows .github/workflows examples examples packages/ipa-tools packages/ipa-tools .gitignore .gitignore LICENSE LICENSE README.md README.md SPEC.md SPEC.md Repository files navigation A proposed browser standard for provider-agnostic AI inference. Status: Experimental Draft Spec: SPEC.md Try It Reference Implementation Inference Bridge — official Chrome extension that injects window.inference and routes to OpenAI, Anthropic, OpenRouter, Ollama, or experimental OpenAI-compatible servers (source) Install from the Chrome Web Store, or for development clone the repository and load it unpacked from chrome://extensions (Developer mode → Load unpacked → select the repo root). Example Applications Examples index — gallery of demo apps (source) Chat demo — minimal chat UI that uses the API (source) Social demo — post + replies with a Grok-like Ask AI panel (source) Translate demo — short haiku translated with ipa-tools complete (source) Nostr feed demo — a feed of Nostr notes filtered with natural language using Inference Bridge experimental tool calling (source) The specification defines the standard. Inference Bridge implements that standard and may also include experimental features that are not part of the API contract yet. Applications should target the Inference Provider API (request and getFeatures), not extension-specific namespaces. Motivation Today, every AI-powered web application has to reinvent the same infrastructure: Ask users for API keys Integrate every inference provider separately Proxy requests through their own backend Build custom permission systems The Inference Provider API (IPA) proposes a standard browser interface that allows web applications to request inference from a user-approved browser extension without ever accessing API keys. Inspired by NIP-07, IPA separates applications from providers, giving users complete control over where inference is performed. Design Principles Users own their API keys. Applications request inference, not providers. Users choose providers. Users choose models. Applications should be provider agnostic. Local and remote inference are first-class citizens. Permission is explicit. API keys never leave the browser extension. Example for await (const chunk of window.inference.request({ method: "chat", messages: [ { role: "user", content: Is this true?:\n\nNostr is dead. } ] })) { if (chunk.type === "accepted") { // permission resolved; provider call may begin } else if (chunk.type === "reasoning_delta") { // optional: model reasoning / chain-of-thought } else if (chunk.type === "delta") { // append chunk.content to the reply UI } else if (chunk.type === "done") { // final message / usage; message.reasoning when reasoning was streamed } } request is required. getFeatures reports optional capabilities such as tool calling and request options (for example reasoningEffort, temperature); implementations that omit it advertise none. If the app only needs the final message, drain to done (inline sketch — or use ipa-tools’s complete): async function complete(request) { let done; for await (const chunk of window.inference.request(request)) { if (chunk.type === "done") done = chunk; } return done; } const { model, message, usage } = await complete({ method: "chat", messages: [{ role: "user", content: "Is this true?:\n\nNostr is dead." }], }); The helper is application code, not part of window.inference. It throws InferenceError the same way iterating request does. Feature-detect optional capabilities before sending tools. Missing getFeatures means none: const features = window.inference.getFeatures?.() ?? {}; if (features.toolCalling) { const tools = [ { type: "function", function: { name: "get_weather", description: "Get the current weather for a city", parameters: { type: "object", properties: { city: { type: "string" } }, required: ["city"], }, }, }, ]; for await (const chunk of window.inference.request({ method: "chat", messages: [{ role: "user", content: "What's the weather in Austin?" }], tools, })) { if (chunk.type === "done" && chunk.message.toolCalls?.length) { // page executes the function, appends role: "tool" results, calls request again } } } Request options can be sent without feature detection — unsupported keys are ignored: // Prefer less thinking / lower temperature for translation for await (const chunk of window.inference.request({ method: "chat", messages: [{ role: "user", content: "Translate to Spanish: Hello" }], options: { reasoningEffort: "none", temperature: 0.2, }, })) { // ... } Any multi-round tool loop is application code. Implementations that do not advertise toolCalling reject tools with invalid_request. Unsupported options keys are ignored (not rejected) so apps may send them for forward compatibility. For a ready-made loop (plus types and complete), see the non-normative ipa-tools package (npm install ipa-tools). Sending tools on IPA request without a toolCalling advertisement is invalid_request. Prefer getFeatures().toolCalling before enabling tools; see ipa-tools. The extension prompts the user for permission: Allow inference? primal.net Provider [ Ollama ▼ ] Model [ Gemma 4 ▼ ] Request preview user: Is this true?: Nostr is dead. [ ] Remember for this site Allow once, or deny only this request. [Allow] [Deny] Request preview is optional extension UX for this draft, not part of the API contract. When the request includes tools, the permission UI must list the function names; a persistent chat grant does not silently cover a later tools request. A follow-up that only appends role: "tool" results may not re-prompt and may not appear in any preview, so applications should disclose what data those tools will send to the provider before Allow (in the first messages, the tool description, or the page UI). See SPEC.md — Tool calling. The user chooses the provider and model. With “Remember for this site” checked, Allow persists access for that origin together with the chosen provider and model; Deny permanently blocks it. Changing the extension’s global default does not alter existing origin grants. Text chat is required. Tool calling is optional: implementations that support it return { toolCalling: true } from getFeatures and accept tools on request. The page defines and executes function tools; the extension only relays schemas, toolCalls, and results. Optional options (for example options.reasoningEffort: "auto" | "none" | "low" | "medium" | "high", options.temperature: number in [0, 2]) lets apps prefer generation settings when the matching getFeatures().options key is true — not a permission change; user override or clamp controls are optional extension UX. See SPEC.md. Goals Standard browser API Provider agnostic Bring Your Own Key (BYOK) Local-first compatible Per-origin permissions Streaming support Zero backend required Optional capability discovery (getFeatures) Optional function tools, executed by the page Optional request options (for example reasoningEffort, temperature) Non-goals Replacing provider SDKs Billing Authentication Defining inference protocols Choosing the "best" model Potential Providers An IPA-compatible browser extension could route requests to any provider, including: OpenAI Anthropic Google Gemini xAI OpenRouter ppq.ai Routstr Ollama LM Studio Local inference servers Applications should not need to know which provider the user has selected. Local providers (Ollama, LM Studio, etc.) Local servers often reject requests that carry a chrome-extension:// Origin header (commonly HTTP 403). IPA extensions that support local inference should strip or rewrite that header on their own requests to loopback endpoints so users are not asked to set OLLAMA_ORIGINS=chrome-extension://* or similar allowlists. Widening the local server's origin allowlist remains a fallback, not the preferred path. Chrome MV3 reference: Inference Bridge does this with declarativeNetRequestWithHostAccess and dynamic rules in src/ollama-origin-bypass.js and src/loopback-origin-bypass.js that remove Origin / Referer for local Ollama and other loopback OpenAI-compatible servers. See SPEC.md Security for the normative guidance. That permission lets the extension modify request headers only for hosts already listed in host_permissions—it is not a browser-wide rewrite capability. Still treat it as privileged: a compromised or overly broad extension could alter headers on those hosts. Prefer port-scoped loopback permissions (for example http://localhost:11434/*) over http://localhost/*, keep DNR rules limited to local inference endpoints, and do not use DNR to touch remote provider traffic. This is still preferable to asking every user to set OLLAMA_ORIGINS=chrome-extension://*, which trusts every installed extension talking to Ollama. Example Use Cases A "Grok" button on every social post. AI-powered documentation. Browser-based coding tools and other page-executed function tools. Translation. Writing assistance. Local-first AI applications. Open Questions Some topics that still need community discussion: Is window.inference the right namespace? Which further capability constraints, if any, do applications need beyond tools and options? Are "auto" | "none" | "low" | "medium" | "high" the right options.reasoningEffort levels, or should the field become a provider-mapped budget/token object? Which further keys belong under options (for example maxTokens), and should clamp/override UX stay optional? Should model selection always remain under user control? Should images, embeddings, and speech use this API or separate APIs? How should extensions surface token usage? Should estimated cost remain optional UX until pricing metadata is defined? Should getFeatures grow beyond booleans (for example nested tool kinds), or stay one key per capability? Should hosted / provider-executed tools (web search, MCP) be specified, or remain implementation-specific? Should tool calls stream as their own chunk type, or stay on done.message.toolCalls only? Should structured outputs (e.g. JSON Schema / responseFormat) be part of IPA, or left to prompt engineering until providers converge? How should permission UIs present multi-message requests — e.g. emphasize the last user message and collapse system/context by default? Should applications be encouraged or required to round-trip message.reasoning on later turns for providers that benefit from it? Contributing This proposal is intentionally in an early draft stage. The goal is to collaboratively design an open browser standard for provider-agnostic inference—not a specific implementation. Contributions of all kinds are welcome, including: Design feedback API suggestions Security considerations Alternative approaches Reference implementations Browser extension prototypes Related standards or prior art If you have an idea or concern, please open an issue. License This project is licensed under the MIT License. See LICENSE for details. MIT license Activity Stars 5 stars Watchers 1 watching Forks 1 fork Report repository