ai-powered

v0.3.2

Published

2 months ago

Unified AI client and CLI: multi-modal, multi-provider, browser-safe, fully mock-able

0High
0Medium
0Low

mytechtoday

ai openai anthropic venice xai grok ollama cli multimodal text image audio video llm typescript browser

ai-powered

Unified AI client and CLI — multi-modal, multi-provider, browser-safe, fully mock-able.

TL;DR

# Install
npm install -g ai-powered

# Run a quick text generation (uses mock provider — no API key needed)
ai-powered text --mock "Explain REST APIs in one sentence."

# Or as a library (Node.js ≥ 18, ESM only)
import { getAiClient } from "ai-powered";
const client = getAiClient({ mock: true });
const result = await client.generateText("Hello, AI!");
console.log(result.content);

Key features at a glance:

| Feature | Detail | | ----------------- | --------------------------------------------------------------------------------------------------- | | Modalities | Text · Image · Audio (transcribe + speak) · Video · Structured JSON | | Providers | OpenAI · Anthropic · xAI (Grok) · Venice.ai · Custom/Ollama · Mock | | Resilience | Per-provider circuit breakers · automatic provider fallback · configurable retry | | Security | API key masking in all logs · SHA-256 prompt hashing in audit log · git-tracked credential warnings | | Plugin system | onRequest / onResponse / onError hooks · frozen config sandboxing | | Browser | Vite ESM+UMD bundle · proxy mode (recommended) · direct mode (dev only) | | ESM only | "type": "module" throughout — CommonJS is not supported (Design Decision D1) |

Requirements

Node.js ≥ 18 (ESM native support required)
npm ≥ 9 (or pnpm / yarn equivalent)
An API key for at least one provider — or use --mock / AI_MOCK=true for zero-cost testing

⚠️ ESM only. This package sets "type": "module" in package.json. You cannot require() it. If your project uses CommonJS, use a dynamic import() wrapper or migrate to ESM.

Installation

# Global CLI install
npm install -g ai-powered

# Local library install
npm install ai-powered

# Development (from source)
git clone https://github.com/mytech-today-now/ai-powered.git
cd ai-powered
npm install
npm run build

Configuration

Config file locations

Config is loaded from multiple layers and merged in priority order (lowest → highest):

| Layer | Path | Notes | | ---------------- | ------------------------------ | ----------------------------------------- | | Schema defaults | — | Zod defaults apply first | | Global config | ~/.ai-powered/config.json | Shared across all projects | | Local config | ./.ai-powered/config.json | Per-project overrides | | Named profile | config.profiles[name] | Selected by profile key or AI_PROFILE | | Environment vars | AI_*, OPENAI_API_KEY, etc. | See table below | | CLI flags | --provider, --model, etc. | Highest precedence |

Environment variables

| Variable | Config key | Example | | ------------------- | -------------------- | ------------ | | OPENAI_API_KEY | apiKey (OpenAI) | sk-… | | ANTHROPIC_API_KEY | apiKey (Anthropic) | sk-ant-… | | XAI_API_KEY | apiKey (xAI) | xai-… | | VENICE_API_KEY | apiKey (Venice) | ven-… | | AI_CUSTOM_API_KEY | apiKey (custom) | any | | AI_PROVIDER | provider | openai | | AI_MODEL | model | gpt-4o | | AI_PROFILE | profile | production | | AI_MOCK | mock | true | | AI_BUDGET_SESSION | budgetSession | 1.00 | | LOG_LEVEL | debug | debug |

Example config file

{
  "provider": "openai",
  "model": "gpt-4o",
  "temperature": 0.7,
  "maxTokens": 2048,
  "systemPrompt": "You are a helpful assistant.",
  "stream": false,
  "fallbackProviders": ["anthropic", "mock"],
  "budgetSession": 1.0,
  "warnBudget": 0.8,
  "plugins": ["audit-log"],
  "profiles": {
    "production": {
      "temperature": 0.3,
      "maxTokens": 4096
    },
    "creative": {
      "temperature": 1.2
    }
  }
}

Custom / self-hosted providers

{
  "provider": "custom",
  "baseUrl": "http://localhost:11434/v1",
  "customProviderType": "ollama",
  "model": "llama3",
  "customHeaders": { "X-Internal-Token": "my-token" }
}

Supported customProviderType values: "openai-compatible" · "ollama" · "other"

CLI Reference

All examples use --mock to avoid real API calls. Remove --mock and set your API key for live use.

Global flags

| Flag | Description | | ------------------- | ---------------------------------------------------------------------------- | | --provider <name> | Override provider (openai, anthropic, xai, venice, custom, mock) | | --model <id> | Override model identifier | | --profile <name> | Use named profile from config | | --mock | Force mock provider | | --dry-run | Estimate cost; skip API call | | --quiet | Print raw content only (no decorators) | | --json | Print JSON envelope | | --session <id> | Attach request to a named conversation session | | --log <path> | Write structured JSONL log to file | | --debug | Enable verbose debug logging | | --no-color | Disable ANSI colors (also NO_COLOR=1) |

`text` — Generate text

# Basic
ai-powered text --mock "Explain quantum entanglement."

# Quiet (raw content only — great for piping)
ai-powered text --mock --quiet "What is 2 + 2?" > answer.txt

# JSON envelope
ai-powered text --mock --json "Summarise TCP/IP in one sentence."

# Dry-run cost estimate
ai-powered text --mock --dry-run "Write a 1000-word essay."

# Multi-turn session
ai-powered text --mock --session my-session "My name is Alice."
ai-powered text --mock --session my-session "What is my name?"

# Streaming
ai-powered text --mock --stream "Tell me a story."

# Built-in template
ai-powered text --mock --template summarize --var text="Long article…"

# Custom system prompt
ai-powered text --mock --system "Reply only in French." "Hello!"

`image` — Generate image

ai-powered image --mock --output image.png "A serene mountain lake at sunrise"

`audio transcribe` — Transcribe audio

ai-powered audio transcribe --mock --file recording.mp3

`audio speak` — Text-to-speech

ai-powered audio speak --mock --output speech.mp3 "Hello, world!"

`video` — Generate video

ai-powered video --mock "A timelapse of clouds over a city"

`structured` — Generate structured JSON

# Using an inline JSON schema file
cat > schema.json <<'EOF'
{
  "type": "object",
  "properties": {
    "name":       { "type": "string" },
    "capital":    { "type": "string" },
    "population": { "type": "number" }
  },
  "required": ["name", "capital", "population"]
}
EOF
ai-powered structured --mock --schema schema.json "Describe France."

`batch` — Batch processing (JSONL)

# Create input file
printf '{"prompt":"Speed of light?"}\n{"prompt":"Who wrote Hamlet?"}\n' > input.jsonl

# Process batch
ai-powered batch text --mock --input input.jsonl --output output.jsonl

# View results
cat output.jsonl

`batch` stdin/stdout piping

Pass - as --input or --output to use standard streams:

# Read prompts from stdin, write results to stdout
printf '{"prompt":"Speed of light?"}\n{"prompt":"Who wrote Hamlet?"}\n' \
  | ai-powered batch text --mock --input - --output -

# Pipe a JSONL file directly into the batch command
cat prompts.jsonl | ai-powered batch text --input - --output -

# Pipe a video shot list and pretty-print NDJSON results with jq
cat shots.jsonl | ai-powered batch video --input - --output - | jq .

`serve` — Start HTTP proxy server

ai-powered serve --mock --port 3001
# Exposes: GET /health, GET /config, GET /models, GET /pricing, POST /text, POST /stream, POST /image, POST /batch, and /v1/* compat routes

`session` — Manage conversation sessions

ai-powered session list
ai-powered session clear my-session

`config` — Manage configuration

ai-powered config list                  # show all config values (keys masked)
ai-powered config get provider          # get one key
ai-powered config set temperature 0.5   # set a key
ai-powered config delete model          # remove a key
ai-powered config reset                 # restore defaults
ai-powered config path                  # print config file path
ai-powered config validate              # validate current config

`list-models` — List available models

ai-powered list-models --provider openai
ai-powered list-models --provider openai --modality image

`list-templates` — List built-in templates

ai-powered list-templates

`health-check` — Check configuration and connectivity

ai-powered health-check         # validates config + API key + git credential safety

`wizard` — Interactive setup

ai-powered wizard               # guided provider/model/API key setup with live validation

File Input

The proxy server and web demo support batch file input on the Video tab. Drop (or click-to-browse) a .json, .jsonl, or .md shot-list file onto the batch drop zone. The app parses it into shots, shows a pre-flight preview, and then sends every shot to POST /batch sequentially. Results appear as live shot cards with play buttons, individual download links, a downloadable HTML results page, and a ZIP export of all video files.

JSONL format (one shot per line)

Text

{"prompt":"Explain quantum entanglement in one sentence"}
{"prompt":"What is the half-life of carbon-14?","model":"gpt-4o"}

Image

{"name":"Hero shot","prompt":"A cinematic product photo of a red sports car on wet asphalt, studio lighting"}
{"name":"Detail shot","prompt":"Close-up of the dashboard with ambient LED lighting","provider":"openai"}

Video

{"name":"Opening","prompt":"Aerial crane shot descending over a gleaming city skyline at dawn, 5 seconds","modality":"video"}
{"name":"Title card","prompt":"Logo reveal with light-ray particle effect on black background, 3 seconds","modality":"video"}

Audio / TTS

{"name":"Intro VO","prompt":"Welcome to the future of AI-powered video production."}
{"name":"Outro VO","prompt":"Thank you for watching. Subscribe for more.","model":"tts-1-hd"}

Structured

{
  "name": "Product info",
  "prompt": "Extract product name, price, and SKU from: Blue Widget XL, $29.99, SKU-4821",
  "modality": "structured"
}

JSON array format

A .json file whose root value is an array is treated identically to JSONL:

[
  {
    "name": "Shot 1",
    "prompt": "Wide establishing shot of a rooftop at sunrise",
    "modality": "video"
  },
  {
    "name": "Shot 2",
    "prompt": "Close-up of solar panels glistening with morning dew",
    "modality": "video"
  }
]

A JSON object with a shots or items key is also accepted:

{
  "shots": [
    { "name": "Opening", "prompt": "Drone flyover of downtown skyline at golden hour" },
    { "name": "Cut", "prompt": "Street-level pedestrian traffic timelapse" }
  ]
}

Markdown shot-list format

A .md file is parsed as a shot list. Headings become shot names; the following paragraph text becomes the prompt. See ai-prompts/roofing-commercial-shot-list2.md for a full production example.

## Scene 1 — Establishing Shot

Wide crane shot descending over a gleaming city skyline at dawn,
warm golden light, 5 seconds.

## Scene 2 — Product Detail

Macro close-up of the finished roofing tile surface, water beading
and rolling off, 3 seconds.

POST /batch API

The proxy exposes POST /batch which accepts an items array and streams results as NDJSON:

curl -X POST http://localhost:3001/batch \
  -H 'Content-Type: application/json' \
  -d '{
    "items": [
      { "modality": "video", "name": "Shot 1", "prompt": "Aerial dawn cityscape" },
      { "modality": "video", "name": "Shot 2", "prompt": "Street-level timelapse" }
    ]
  }'

Each response line:

{"index":0,"name":"Shot 1","modality":"video","prompt":"Aerial dawn cityscape","status":"ok","result":{…}}
{"index":1,"name":"Shot 2","modality":"video","prompt":"Street-level timelapse","status":"ok","result":{…}}

Per-item errors use "status":"error" and include an "error" field — they do not abort the remaining items.

HTML export and base64 trade-off

The ⬇ Download HTML export produces a self-contained batch-results.html file with all generated media embedded as base64 data URIs — no external server or file paths required. This means the file works completely offline and can be emailed or archived as a single artifact.

Trade-off: base64 encoding inflates binary data by ~33 %. A batch of ten 10 MB video clips will produce a roughly 130 MB HTML file. For large video batches, prefer ⬇ Download ZIP, which packages the raw .mp4 files together with a lightweight results.html that links to them by relative path.

| Export | Best for | Approximate size | | --------------- | -------------------------------------- | ----------------------- | | HTML (base64) | Sharing, archiving, offline viewing | ~1.33× total video size | | ZIP (raw files) | Large batches, local playback, editing | ~1× total video size |

Library Usage

ESM only. Import from "ai-powered" — no require().

Quick start

import { getAiClient } from "ai-powered";

// Use mock provider for testing (no API key required)
const client = getAiClient({ mock: true });

const result = await client.generateText("Hello, AI!");
console.log(result.content); // string
console.log(result.usage); // { promptTokens, completionTokens, totalTokens }
console.log(result.cost); // { totalUsd, isEstimate }
console.log(result.latencyMs); // number

Text generation

import { getAiClient } from "ai-powered";

const client = getAiClient({
  provider: "openai",
  model: "gpt-4o",
  temperature: 0.7,
  maxTokens: 1024,
  systemPrompt: "You are a helpful assistant.",
});

const result = await client.generateText("Explain WebAssembly.");

Image generation

const result = await client.generateImage("A serene mountain lake at sunrise");
console.log(result.url); // string | undefined
console.log(result.base64); // string | undefined

Audio transcription

import { readFileSync } from "node:fs";

const audio = readFileSync("recording.mp3");
const result = await client.transcribeAudio(audio, "recording.mp3");
console.log(result.transcript);
console.log(result.durationSeconds);

Speech synthesis

const result = await client.synthesizeSpeech("Hello, world!");
// result.audioData is a Buffer (Node) or Uint8Array (browser)

Video generation

const result = await client.generateVideo("A timelapse of clouds over a city");
console.log(result.url);

Structured JSON output

import { z } from "zod";

const CountrySchema = z.object({
  name: z.string(),
  capital: z.string(),
  population: z.number(),
  in_europe: z.boolean(),
});

const result = await client.generateStructured("Describe France.", CountrySchema);
console.log(result.data); // typed as { name, capital, population, in_europe }

Streaming text

const stream = client.streamText("Tell me a long story.");
for await (const chunk of stream) {
  process.stdout.write(chunk);
}

Multi-turn conversation sessions

import { getAiClient, ConversationSession } from "ai-powered";

const client = getAiClient({ mock: true });
const session = new ConversationSession("system: You are helpful.");

session.addMessage("user", "My name is Alice.");
const r1 = await client.generateText(session.buildPrompt());
session.addMessage("assistant", r1.content);

session.addMessage("user", "What is my name?");
const r2 = await client.generateText(session.buildPrompt());
console.log(r2.content); // "Alice"

Loading config manually

import { loadConfig, getAiClient } from "ai-powered";

const config = loadConfig({
  profileOverride: "production",
  flags: { temperature: 0.3 },
});

const client = getAiClient(config);

Provider fallback

const client = getAiClient({
  provider: "openai",
  fallbackProviders: ["anthropic", "mock"],
  fallback: true, // default: true
  circuitBreakerThreshold: 5, // open circuit after 5 consecutive failures
  circuitBreakerResetMs: 60_000, // probe after 60 s
});

Budget enforcement

const client = getAiClient({
  budgetSession: 0.5, // max $0.50 per session
  warnBudget: 0.8, // warn at 80% ($0.40)
});
// BudgetExceededError thrown before API call if projected cost exceeds limit

AI Agent / Tool-Calling Usage

ai-powered is designed as a first-class tool for AI agent frameworks (LangChain, AutoGPT, custom orchestrators). The library exposes all five modalities as discrete, idempotent functions that map cleanly to function-calling schemas.

OpenAI function-calling example

import OpenAI from "openai";
import { getAiClient } from "ai-powered";

const openai = new OpenAI();
const aiClient = getAiClient({ mock: true });

// Define tools backed by ai-powered
const tools: OpenAI.Chat.Completions.ChatCompletionTool[] = [
  {
    type: "function",
    function: {
      name: "generate_text",
      description: "Generate text from a prompt using the configured AI provider.",
      parameters: {
        type: "object",
        properties: {
          prompt: { type: "string", description: "The user prompt." },
          temperature: { type: "number", description: "Sampling temperature 0–2." },
          maxTokens: { type: "integer", description: "Max tokens to generate." },
        },
        required: ["prompt"],
      },
    },
  },
  {
    type: "function",
    function: {
      name: "generate_image",
      description: "Generate an image from a text description.",
      parameters: {
        type: "object",
        properties: {
          prompt: { type: "string", description: "Image description." },
        },
        required: ["prompt"],
      },
    },
  },
  {
    type: "function",
    function: {
      name: "generate_structured",
      description: "Generate structured JSON matching a Zod schema.",
      parameters: {
        type: "object",
        properties: {
          prompt: { type: "string" },
          schemaName: { type: "string", description: "Schema identifier known to the agent." },
        },
        required: ["prompt", "schemaName"],
      },
    },
  },
];

// Tool dispatch map
async function dispatchTool(name: string, args: Record<string, unknown>): Promise<string> {
  switch (name) {
    case "generate_text": {
      const result = await aiClient.generateText(args.prompt as string, {
        temperature: args.temperature as number | undefined,
        maxTokens: args.maxTokens as number | undefined,
      });
      return result.content;
    }
    case "generate_image": {
      const result = await aiClient.generateImage(args.prompt as string);
      return result.url ?? result.base64 ?? "No image data";
    }
    case "generate_structured": {
      const { z } = await import("zod");
      // Example: resolve a named schema
      const schema = z.object({ answer: z.string() });
      const result = await aiClient.generateStructured(args.prompt as string, schema);
      return JSON.stringify(result.data);
    }
    default:
      return `Unknown tool: ${name}`;
  }
}

// Agentic loop
async function agentLoop(userMessage: string) {
  const messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[] = [
    { role: "user", content: userMessage },
  ];

  while (true) {
    const response = await openai.chat.completions.create({
      model: "gpt-4o",
      messages,
      tools,
      tool_choice: "auto",
    });

    const choice = response.choices[0];
    if (!choice) break;

    messages.push(choice.message);

    if (choice.finish_reason === "stop") {
      console.log("Agent response:", choice.message.content);
      break;
    }

    if (choice.finish_reason === "tool_calls" && choice.message.tool_calls) {
      for (const toolCall of choice.message.tool_calls) {
        const args = JSON.parse(toolCall.function.arguments) as Record<string, unknown>;
        const result = await dispatchTool(toolCall.function.name, args);
        messages.push({
          role: "tool",
          tool_call_id: toolCall.id,
          content: result,
        });
      }
    }
  }
}

await agentLoop("Generate an image of a sunset and then describe it.");

Machine-readable tool schema (JSON)

Agents that use raw JSON schemas (e.g. Anthropic tool_use) can use this schema block directly:

{
  "tools": [
    {
      "name": "generate_text",
      "description": "Generate text using ai-powered. Supports all configured providers.",
      "input_schema": {
        "type": "object",
        "properties": {
          "prompt": { "type": "string" },
          "provider": {
            "type": "string",
            "enum": ["openai", "anthropic", "xai", "venice", "mock"]
          },
          "model": { "type": "string" },
          "temperature": { "type": "number" },
          "maxTokens": { "type": "integer" },
          "systemPrompt": { "type": "string" }
        },
        "required": ["prompt"]
      }
    },
    {
      "name": "generate_image",
      "description": "Generate an image from a text description.",
      "input_schema": {
        "type": "object",
        "properties": {
          "prompt": { "type": "string" },
          "provider": { "type": "string" },
          "model": { "type": "string" }
        },
        "required": ["prompt"]
      }
    },
    {
      "name": "transcribe_audio",
      "description": "Transcribe audio file (base64-encoded) to text.",
      "input_schema": {
        "type": "object",
        "properties": {
          "audioBase64": { "type": "string" },
          "filename": { "type": "string" }
        },
        "required": ["audioBase64"]
      }
    },
    {
      "name": "synthesize_speech",
      "description": "Convert text to speech. Returns base64-encoded audio.",
      "input_schema": {
        "type": "object",
        "properties": {
          "text": { "type": "string" },
          "model": { "type": "string" }
        },
        "required": ["text"]
      }
    },
    {
      "name": "generate_structured",
      "description": "Generate a structured JSON object validated against a schema.",
      "input_schema": {
        "type": "object",
        "properties": {
          "prompt": { "type": "string" },
          "jsonSchema": { "type": "object", "description": "JSON Schema for the output object." }
        },
        "required": ["prompt", "jsonSchema"]
      }
    }
  ]
}

HTTP proxy tool-calling (serve mode)

When running ai-powered serve, all five modalities are available as HTTP endpoints. Agents can call them directly:

POST http://localhost:3001/text         { "prompt": "…", "provider": "openai" }
POST http://localhost:3001/image        { "prompt": "…" }
POST http://localhost:3001/structured   { "prompt": "…" }
POST http://localhost:3001/audio/transcribe  { "audioBase64": "…" }
POST http://localhost:3001/audio/speak  { "text": "…" }
POST http://localhost:3001/video        { "prompt": "…" }
GET  http://localhost:3001/health       → { "status": "ok" }
GET  http://localhost:3001/models       → [{ "id": "…", "name": "…" }]

All endpoints accept per-request overrides (provider, model, temperature, profile).

Standard API Compatibility

ai-powered exposes industry-standard wire-format endpoints so that existing OpenAI and Anthropic client libraries work without modification — just point the baseURL at your proxy server. API keys are managed centrally in ai-powered; consumers send no credentials.

Endpoint table

| Route | Standard | Providers | Streaming | | --- | --- | --- | --- | | POST /v1/chat/completions | OpenAI Chat | openai · anthropic · xai · venice · mock | ✅ SSE (stream: true) | | POST /v1/messages | Anthropic Messages | openai · anthropic · xai · venice · mock | ✅ SSE (6-event sequence) | | GET /v1/models | OpenAI Models | all active providers | — | | POST /v1/images/generations | OpenAI Images | openai · venice · mock | — | | POST /v1/audio/transcriptions | OpenAI Audio | openai · mock | — | | POST /v1/audio/speech | OpenAI TTS | openai · mock | — | | POST /v1/video/generations | ai-powered-native | lumaai · mock | — |

⚠️ /v1/video/generations uses an ai-powered-native request/response shape. There is no external industry standard for video generation; the route exists to give proxy consumers a consistent /v1/ namespace.

Provider × modality support matrix

| Provider | text | image | audio | video | structured | | --- | :---: | :---: | :---: | :---: | :---: | | openai | ✅ | ✅ | ✅ | — | ✅ | | anthropic | ✅ | — | — | — | ✅ | | xai | ✅ | — | — | — | ✅ | | venice | ✅ | ✅ | — | — | — | | lumaai | — | — | — | ✅ | — | | mock | ✅ | ✅ | ✅ | ✅ | ✅ |

OpenAI client quick-start (FilmBuff pattern)

Point the official OpenAI SDK at your proxy. The apiKey field is required by the SDK but is not forwarded — ai-powered manages all real credentials server-side.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://localhost:3001/v1",
  apiKey:  "not-used", // ai-powered manages credentials
});

const response = await client.chat.completions.create({
  model:    "gpt-4",
  messages: [{ role: "user", content: "Summarise the plot of Metropolis." }],
});

console.log(response.choices[0]?.message.content);

Anthropic client quick-start

Override baseURL on the official Anthropic SDK so requests hit the proxy instead of api.anthropic.com. Again, credentials are not forwarded.

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL:  "http://localhost:3001/v1",
  apiKey:   "not-used", // ai-powered manages credentials
});

const message = await client.messages.create({
  model:      "claude-3-5-sonnet-20241022",
  max_tokens: 1024,
  messages:   [{ role: "user", content: "What is the golden ratio?" }],
});

console.log(message.content[0]?.type === "text" ? message.content[0].text : "");

💡 API key management: consumers never embed provider credentials. All keys live in ~/.ai-powered/config.json or environment variables on the machine running the proxy.

Browser / Web Usage

The ai-powered/web entry point ships a Vite-built ESM+UMD bundle (dist-web/) with no Node.js built-in dependencies.

Two modes

| Mode | Use case | API key exposure | | ---------- | --------------- | ----------------------------------------------------------- | | proxy | Production | Key stays on your server — browser never sees it | | direct | Dev / demo only | Key visible in DevTools — non-suppressible DOM banner shown |

Proxy mode (recommended)

Start the proxy server on your backend:

ai-powered serve --port 3001

Then in your browser app:

import { createWebClient } from "ai-powered/web";

const client = createWebClient({
  mode: "proxy",
  proxyUrl: "http://localhost:3001",
});

const result = await client.generateText("Hello from the browser!");
console.log(result.content);

Streaming in proxy mode (SSE)

const response = await fetch("http://localhost:3001/stream", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({ prompt: "Tell me a story." }),
});

const reader = response.body!.getReader();
const decoder = new TextDecoder();

while (true) {
  const { done, value } = await reader.read();
  if (done) break;
  const lines = decoder.decode(value, { stream: true }).split("\n");
  for (const line of lines) {
    if (!line.startsWith("data: ")) continue;
    const payload = line.slice(6).trim();
    if (payload === "[DONE]") return;
    const { delta } = JSON.parse(payload);
    document.getElementById("output")!.textContent += delta;
  }
}

Direct mode (development only)

const client = createWebClient({
  mode: "direct",
  provider: "openai",
  apiKey: "sk-…", // ⚠️ visible in DevTools — dev/demo only
});

const result = await client.generateText("Hello!");

Browser conversation sessions

import { BrowserConversationSession } from "ai-powered/web";

const session = new BrowserConversationSession("chat-1");
// State is persisted to sessionStorage automatically
session.addMessage("user", "Hello!");
const history = session.getMessages();

HTML quick-start

<!doctype html>
<html lang="en">
  <head>
    <meta charset="UTF-8" />
    <title>ai-powered demo</title>
  </head>
  <body>
    <textarea id="prompt">Explain WebAssembly.</textarea>
    <button id="btn">Generate</button>
    <pre id="output"></pre>

    <script type="module">
      import { createWebClient } from "https://cdn.jsdelivr.net/npm/ai-powered/dist-web/ai-powered.esm.js";

      document.getElementById("btn").addEventListener("click", async () => {
        const client = createWebClient({ mode: "proxy", proxyUrl: "http://localhost:3001" });
        const result = await client.generateText(document.getElementById("prompt").value);
        document.getElementById("output").textContent = result.content;
      });
    </script>
  </body>
</html>

Vite / bundler integration

// vite.config.ts — resolve ai-powered/web to local source in dev
import { defineConfig } from "vite";
export default defineConfig({
  resolve: {
    alias: { "ai-powered/web": "/src/ai-powered/web/index.ts" },
  },
});

Cross-Language Shell Integration

ai-powered is callable from any language that can invoke a subprocess. The CLI is the integration point. All examples set AI_MOCK=true to avoid real API calls during development.

Bash

# integrations/bash-example.sh
export AI_MOCK=true
result=$(ai-powered text --quiet "What is 2+2?")
echo "Answer: $result"

# Batch processing
printf '{"prompt":"Speed of light?"}\n{"prompt":"Who wrote Hamlet?"}\n' > input.jsonl
ai-powered batch text --input input.jsonl --output output.jsonl

Python

# integrations/python-example.py
import subprocess, os

def ask(prompt: str) -> str:
    result = subprocess.run(
        ["ai-powered", "text", "--mock", "--quiet", prompt],
        capture_output=True, text=True, check=True,
        env={**os.environ, "AI_MOCK": "true"},
    )
    return result.stdout.strip()

print(ask("Explain REST APIs in one sentence."))

PowerShell

# integrations/powershell-example.ps1
$env:AI_MOCK = "true"
$result = ai-powered text --mock --quiet "What is quantum computing?"
Write-Host "Answer: $result"

Windows Batch

REM integrations/batch-example.bat
set AI_MOCK=true
for /f "delims=" %%i in ('ai-powered text --mock --quiet "What is the speed of light?"') do set RESULT=%%i
echo Answer: %RESULT%

Go

// integrations/go-example.go
package main

import (
    "fmt"
    "os"
    "os/exec"
    "strings"
)

func ask(prompt string) (string, error) {
    cmd := exec.Command("ai-powered", "text", "--mock", "--quiet", prompt)
    cmd.Env = append(os.Environ(), "AI_MOCK=true")
    out, err := cmd.Output()
    return strings.TrimSpace(string(out)), err
}

func main() {
    answer, _ := ask("Explain TCP/IP in one sentence.")
    fmt.Println(answer)
}

Java

// integrations/java-example.java
import java.io.*;
import java.util.*;

public class AiPoweredExample {
    public static String ask(String prompt) throws Exception {
        ProcessBuilder pb = new ProcessBuilder("ai-powered", "text", "--mock", "--quiet", prompt);
        pb.environment().put("AI_MOCK", "true");
        pb.redirectErrorStream(true);
        Process process = pb.start();
        return new String(process.getInputStream().readAllBytes()).trim();
    }
    public static void main(String[] args) throws Exception {
        System.out.println(ask("Explain REST APIs in one sentence."));
    }
}

C#

// integrations/csharp-example.cs
using System.Diagnostics;

static string Ask(string prompt) {
    var psi = new ProcessStartInfo("ai-powered", $"text --mock --quiet \"{prompt}\"") {
        RedirectStandardOutput = true,
        UseShellExecute = false,
        Environment = { ["AI_MOCK"] = "true" }
    };
    using var p = Process.Start(psi)!;
    return p.StandardOutput.ReadToEnd().Trim();
}

Console.WriteLine(Ask("Explain quantum computing."));

Ruby

# integrations/ruby-example.rb
require 'open3'

def ask(prompt)
  stdout, _status = Open3.capture2(
    { 'AI_MOCK' => 'true' },
    'ai-powered', 'text', '--mock', '--quiet', prompt
  )
  stdout.strip
end

puts ask('What is machine learning?')

Rust

// integrations/rust-example.rs
use std::process::Command;

fn ask(prompt: &str) -> String {
    let output = Command::new("ai-powered")
        .args(["text", "--mock", "--quiet", prompt])
        .env("AI_MOCK", "true")
        .output()
        .expect("Failed to run ai-powered");
    String::from_utf8_lossy(&output.stdout).trim().to_string()
}

fn main() {
    println!("{}", ask("Explain REST APIs."));
}

PHP

<?php
// integrations/php-example.php
putenv('AI_MOCK=true');
$prompt = escapeshellarg('Explain REST APIs in one sentence.');
$result = trim(shell_exec("ai-powered text --mock --quiet $prompt"));
echo $result . PHP_EOL;

Perl

#!/usr/bin/env perl
# integrations/perl-example.pl
$ENV{AI_MOCK} = 'true';
my $result = `ai-powered text --mock --quiet "Explain REST APIs."`;
chomp $result;
print "$result\n";

Full integration scripts are available in the integrations/ directory.

Security Best Practices

API key protection

Never commit API keys to version control. Use ~/.ai-powered/config.json (global) or environment variables.
The health-check command warns if a config file containing an API key is tracked by git.
All log output passes through maskApiKey() which redacts key values:
- sk-… → sk-...****
- sk-ant-… → sk-ant-...****
- xai-… → xai-...****
- ven-… → ven-...****
The pre-commit hook scans staged files for any of these patterns and aborts if found.

Audit log plugin

Enable the built-in audit-log plugin for a tamper-evident, key-safe audit trail:

{ "plugins": ["audit-log"] }

Each entry in ai-powered-audit.jsonl looks like:

{
  "type": "request",
  "timestamp": "2026-03-28T12:00:00.000Z",
  "modality": "text",
  "provider": "openai",
  "model": "gpt-4o",
  "promptHash": "a3f1…",
  "apiKeyMasked": "sk-...****",
  "options": { "temperature": 0.7, "maxTokens": 1024 }
}

Raw prompts are stored as SHA-256 hashes — they can be verified but not reversed.

Browser security

Always use proxy mode in production. The API key never leaves your server.
Direct mode renders a non-suppressible DOM banner warning users the key is in DevTools.
The Vite build post-process step scans dist-web/ for leaked key prefixes and aborts if found.

Prompt injection defense

The built-in prompt-shield plugin heuristically detects common injection patterns:

{ "plugins": ["prompt-shield"] }

Set reject: true in the plugin config to block flagged requests instead of only logging:

import { createPromptShieldPlugin } from "ai-powered";
const shield = createPromptShieldPlugin({ reject: true });

Budget limits

Set budgetSession in config or via flags to cap spend per session:

ai-powered text --budget-session 0.10 --mock "Write an essay."

A BudgetExceededError is thrown before the API call if the projected cost would exceed the limit.

Architecture Overview

ai-powered supports four invocation modes, each suitable for different contexts:

┌──────────────────────────────────────────────────────────────────────────┐
│                         ai-powered architecture                          │
├─────────────────┬───────────────┬──────────────────┬────────────────────┤
│   Mode 1: CLI   │ Mode 2: Lib   │ Mode 3: Agent    │  Mode 4: Browser   │
│                 │               │  Tool-Calling    │                    │
│ ai-powered text │ getAiClient() │ HTTP POST /text  │ createWebClient()  │
│ ai-powered image│ AiClient.*    │ HTTP POST /image │ proxy or direct    │
│ ai-powered serve│ loadConfig()  │ GET /health      │ BrowserConvSession │
│ ai-powered batch│               │                  │ SSE streaming      │
└────────┬────────┴───────┬───────┴────────┬─────────┴────────────────────┘
         │                │                │
         └────────────────┼────────────────┘
                          ▼
              ┌─────────────────────┐
              │    AiClient core    │
              │  Plugin pipeline    │
              │  Budget tracking    │
              │  Circuit breakers   │
              │  Retry / fallback   │
              └─────────┬───────────┘
                        │
          ┌─────────────┼─────────────┐
          ▼             ▼             ▼
    OpenAiProvider  AnthropicProvider  VeniceProvider
    GrokProvider    CustomProvider     MockProvider

Config layers (lowest → highest precedence)

Schema defaults → Global config → Local config → Named profile → Env vars → CLI flags

Plugin pipeline execution order

Request:  plugin[0].onRequest → plugin[1].onRequest → … → provider call
Response: … → plugin[1].onResponse → plugin[0].onResponse
Error:    plugin[0].onError → plugin[1].onError → …

Plugins receive a frozen snapshot of AiConfig — mutations throw TypeError. An unhandled error in a plugin is caught, wrapped as PluginError, logged, and the plugin is bypassed for the remainder of the session.

ESM-only design (Decision D1)

This package uses "type": "module" and ships only ES Modules. This decision was made to:

Support top-level await in CLI entry points
Enable tree-shaking in Vite browser bundles
Align with the direction of the Node.js and npm ecosystems
Avoid dual-package hazards (CJS/ESM singleton state issues)

Migration path for CJS consumers: wrap the import in a dynamic import() or migrate to "type": "module".

Writing a Plugin

Plugins are the primary extension point for ai-powered. A plugin is any ESM module that exports an object conforming to the AiPlugin interface. Plugins can observe, modify, or react to every request/response/error cycle without touching core library code.

The `AiPlugin` interface

import type { AiPlugin, RequestContext, ResponseContext, AiPoweredError } from "ai-powered";

export const myPlugin: AiPlugin = {
  /** Required: unique plugin identifier. Used in logs and error messages. */
  name: "my-plugin",

  /** Optional: semver version string. */
  version: "1.0.0",

  /** Optional: human-readable description. */
  description: "A short description of what this plugin does.",

  /**
   * Called before every provider API call.
   * Receives the full request context; must return the (possibly modified) context.
   * Throw PluginError to signal a non-fatal failure — the plugin is bypassed.
   * Throw any other error to abort the request entirely.
   */
  async onRequest(ctx: RequestContext): Promise<RequestContext> {
    // ctx.config      — frozen AiConfig snapshot (mutations throw TypeError)
    // ctx.messages    — mutable message array
    // ctx.modality    — "text" | "image" | "audio" | "video" | "structured"
    return ctx;
  },

  /**
   * Called after every successful provider response.
   * Receives the response context; must return the (possibly modified) context.
   */
  async onResponse(ctx: ResponseContext): Promise<ResponseContext> {
    // ctx.result      — typed result object (TextResult, ImageResult, etc.)
    // ctx.modality    — same modality as the request
    return ctx;
  },

  /**
   * Called for every AiPoweredError (provider errors, budget errors, etc.).
   * Return void; errors thrown here are logged but do not propagate.
   */
  async onError(error: AiPoweredError): Promise<void> {
    // error.code      — machine-readable error code string
    // error.message   — human-readable message
  },
};

Key `RequestContext` fields

| Field | Type | Description | | ---------- | ------------------------ | ---------------------------------------------------- | | config | Readonly<AiConfig> | Frozen config snapshot — mutations throw TypeError | | messages | Array<{role, content}> | Mutable message list | | modality | Modality | Active modality for this request |

Key `ResponseContext` fields

| Field | Type | Description | | ---------- | -------------------------------- | ----------------- | | result | TextResult \| ImageResult \| … | Provider response | | modality | Modality | Active modality |

Full example: rate-limiter plugin

// plugins/rate-limiter.ts
import type { AiPlugin, RequestContext, AiPoweredError } from "ai-powered";

export interface RateLimiterOptions {
  /** Max requests per window. Default: 10. */
  maxRequests?: number;
  /** Window duration in milliseconds. Default: 60_000 (1 minute). */
  windowMs?: number;
}

export function createRateLimiterPlugin(opts: RateLimiterOptions = {}): AiPlugin {
  const maxRequests = opts.maxRequests ?? 10;
  const windowMs = opts.windowMs ?? 60_000;
  const timestamps: number[] = [];

  return {
    name: "rate-limiter",
    version: "1.0.0",
    description: `Limits to ${maxRequests} requests per ${windowMs / 1000}s window.`,

    async onRequest(ctx: RequestContext): Promise<RequestContext> {
      const now = Date.now();
      // Evict timestamps outside the current window
      while (timestamps.length > 0 && now - timestamps[0]! > windowMs) {
        timestamps.shift();
      }

      if (timestamps.length >= maxRequests) {
        const resetIn = windowMs - (now - timestamps[0]!);
        throw new Error(
          `Rate limit exceeded: ${maxRequests} req/${windowMs}ms. ` +
            `Reset in ${Math.ceil(resetIn / 1000)}s.`,
        );
      }

      timestamps.push(now);
      return ctx;
    },

    async onError(error: AiPoweredError): Promise<void> {
      // Optionally log rate limit errors differently
      if (error.message.includes("Rate limit exceeded")) {
        console.warn("[rate-limiter]", error.message);
      }
    },
  };
}

Registering plugins

Via config file (string identifiers — built-ins or npm packages):

{
  "plugins": ["audit-log", "rate-limiter", "./plugins/my-plugin.js", "@my-scope/ai-powered-plugin"]
}

Via the library API (plugin objects — for programmatic control):

import { getAiClient } from "ai-powered";
import { createRateLimiterPlugin } from "./plugins/rate-limiter.js";

const client = getAiClient(
  {
    mock: true,
    plugins: ["audit-log"], // string: built-in or npm package
  },
  [
    createRateLimiterPlugin({ maxRequests: 20, windowMs: 30_000 }), // object: programmatic
  ],
);

Plugin sandboxing

Frozen config: ctx.config is a deep-frozen snapshot of AiConfig. Any attempt to mutate it throws a TypeError. This ensures plugins cannot alter global configuration.
PluginError isolation: If a plugin's hook throws an error that is not intentionally re-thrown by the caller, it is caught, wrapped as a PluginError, and logged. The plugin is then bypassed for subsequent hooks in the current request cycle.
Bypass behaviour: A bypassed plugin still receives onError callbacks (from a separate try-catch) so audit-style plugins never miss error events.

Error handling in plugins

import { PluginError } from "ai-powered";

async onRequest(ctx: RequestContext): Promise<RequestContext> {
  try {
    await someExternalCall();
  } catch (err) {
    // Throw PluginError for non-fatal plugin failures:
    // the plugin is bypassed but the request continues normally.
    throw new PluginError("my-plugin", "External call failed", { cause: err });
  }
  return ctx;
}

Built-in plugins

| Plugin ID | Factory | Description | | ----------------- | -------------------------------- | ------------------------------------------------------------ | | "audit-log" | createAuditLogPlugin(opts) | JSONL audit trail with masked keys and SHA-256 prompt hashes | | "rate-limiter" | createRateLimiterPlugin(opts) | Token-bucket rate limiter per sliding window | | "prompt-shield" | createPromptShieldPlugin(opts) | Heuristic prompt injection detector |

Publishing a plugin to npm

Create an ESM package with "type": "module" in package.json.
Export your plugin factory as a named export:

// index.ts
export { createMyPlugin } from "./my-plugin.js";
export type { MyPluginOptions } from "./my-plugin.js";

Name the package with the ai-powered-plugin- prefix by convention:

{
  "name": "ai-powered-plugin-my-feature",
  "version": "1.0.0",
  "type": "module",
  "peerDependencies": {
    "ai-powered": ">=0.1.0"
  }
}

Users register it by package name:

{ "plugins": ["ai-powered-plugin-my-feature"] }

ai-powered dynamically imports plugin strings via import(pluginId). The default export or a named export called default or plugin is used if the import resolves to a module rather than an AiPlugin object directly.

Contributing

See CONTRIBUTING.md for branch naming, commit conventions, PR process, and the full plugin authoring guide.

Quick summary:

# Clone and install
git clone https://github.com/mytech-today-now/ai-powered.git
cd ai-powered
npm install

# Build
npm run build

# Run all tests (mock provider — no API key required)
AI_MOCK=true npm test

# Lint and format
npm run lint && npm run format

# Start the web dev server
npm run dev:web

# Start the proxy server (mock)
npm run serve

Branch naming: feat/<slug> · fix/<slug> · docs/<slug> · refactor/<slug> · ci/<slug> · release/v<semver>

Commit format: feat(scope): description following Conventional Commits.

Made with ❤️ and AI by the ai-powered contributors.

Published

Vulnerabilities

Links

Maintainers

Keywords

Readme

ai-powered

TL;DR

Table of Contents

Requirements

Installation

Configuration

Config file locations

Environment variables

Example config file

Custom / self-hosted providers

CLI Reference

Global flags

text — Generate text

image — Generate image

audio transcribe — Transcribe audio

audio speak — Text-to-speech

video — Generate video

structured — Generate structured JSON

batch — Batch processing (JSONL)

batch stdin/stdout piping

serve — Start HTTP proxy server

session — Manage conversation sessions

config — Manage configuration

list-models — List available models

list-templates — List built-in templates

health-check — Check configuration and connectivity

wizard — Interactive setup

File Input

JSONL format (one shot per line)

JSON array format

Markdown shot-list format

POST /batch API

HTML export and base64 trade-off

Library Usage

Quick start

Text generation

Image generation

Audio transcription

Speech synthesis

Video generation

Structured JSON output

Streaming text

Multi-turn conversation sessions

Loading config manually

Provider fallback

Budget enforcement

AI Agent / Tool-Calling Usage

OpenAI function-calling example

Machine-readable tool schema (JSON)

HTTP proxy tool-calling (serve mode)

Standard API Compatibility

Endpoint table

Provider × modality support matrix

OpenAI client quick-start (FilmBuff pattern)

Anthropic client quick-start

Browser / Web Usage

Two modes

Proxy mode (recommended)

Streaming in proxy mode (SSE)

Direct mode (development only)

Browser conversation sessions

HTML quick-start

Vite / bundler integration

Cross-Language Shell Integration

Bash

Python

PowerShell

Windows Batch

Go

Java

C#

Ruby

Rust

PHP

`text` — Generate text

`image` — Generate image

`audio transcribe` — Transcribe audio

`audio speak` — Text-to-speech

`video` — Generate video

`structured` — Generate structured JSON

`batch` — Batch processing (JSONL)

`batch` stdin/stdout piping

`serve` — Start HTTP proxy server

`session` — Manage conversation sessions

`config` — Manage configuration

`list-models` — List available models

`list-templates` — List built-in templates

`health-check` — Check configuration and connectivity

`wizard` — Interactive setup

The `AiPlugin` interface

Key `RequestContext` fields

Key `ResponseContext` fields