Provider registry

Every provider in aiplug is registered in data/registry.json and resolves to a single Transport class. The registry is the source of truth — adding a new provider is a one-file change.

What’s in the registry

data/registry.json carries the runtime-only fields:

{
  "anthropic": {
    "module": "anthropic",
    "class": "AnthropicTransport",
    "defaultBaseURL": "https://api.anthropic.com",
    "auth": "x-api-key",
    "authHeader": "x-api-key"
  },
  "openai": {
    "module": "openai",
    "class": "OpenAITransport",
    "defaultBaseURL": "https://api.openai.com/v1",
    "auth": "bearer"
  },
  "ollama": {
    "module": "ollama",
    "class": "OllamaTransport",
    "defaultBaseURL": "http://localhost:11434/v1",
    "auth": "none"
  }
}

A separate data/registry.meta.json carries cosmetic fields: displayName, notes, category, popularModels, envVar. The runtime never reads the meta file; it’s for the CLI, the TUI picker, and the introspection surface.

Discovery

AIPlug looks up providers via getEntry(slug). The entry returns the transport module path, the class name, the default base URL, and the auth scheme. The client uses this to:

  1. Validate the user-supplied config (defaultBaseURL, auth) before instantiating the transport.
  2. Dynamic-import the transport module on first use.
  3. Construct the transport with the merged config.

Introspection is sync and safe before any transport is loaded:

const ai = new AIPlug({ transport: 'anthropic', apiKey: '...', model: '...' });

// Sync — no transport instance needed
const caps = ai.capabilities();              // → { name, version, capabilities, auth, defaultBaseURL, … }
const desc = ai.describeProvider('bedrock');  // → { slug, displayName, defaultBaseURL, auth, popularModels, envVar, … }
const fields = ai.configSchema('bedrock');   // → { fields: [{ key, label, secret, required, default, help, envVar }] }
const list = ai.providers();                  // → ProviderDescriptor[] (every registered provider)

These sync APIs are the integration points memoryblock uses to render its provider picker, the auth form, and the model dropdown.

Adding a new provider

  1. Create src/providers/<slug>/index.ts and src/providers/<slug>/capabilities.ts (if it has its own metadata).
  2. Implement the transport by extending OpenAITransport (for OpenAI-shaped servers) or Transport (for custom wire formats).
  3. Add the entry to data/registry.json (module, class, defaultBaseURL, auth).
  4. Add the cosmetic entry to data/registry.meta.json (displayName, notes, popularModels, envVar).
  5. Run npm run build:registry to regenerate the thin per-provider files and the auto-generated OpenAI-compatible shims.

That’s the entire workflow. No core change required.

The OpenAICompatible shortcut

Any server that speaks the OpenAI Chat Completions wire format — Together, Groq, OpenRouter, Fireworks, llama.cpp server, vLLM, LM Studio, and many more — gets aiplug support for free by extending OpenAICompatibleTransport. You only need to set defaultBaseURL and auth. The class inherits all chat / stream / embeddings / images / audio / models logic from the OpenAI transport.

import { OpenAICompatibleTransport } from '../openai-compatible/index.js';
import { METADATA as OPENAI_METADATA } from '../openai/capabilities.js';

export class MyProviderTransport extends OpenAICompatibleTransport {
  override capabilities() {
    return { ...OPENAI_METADATA, name: 'my-provider', defaultBaseURL: 'https://api.example.com/v1', auth: 'bearer' };
  }
}

Add the entry to data/registry.json. Done.

Notable providers shipped today

aiplug ships with first-class support for every major commercial and open-source model API. Highlights:

  • Anthropic — Claude 3.5/3.7/4 family, including reasoning and prompt caching.
  • OpenAI — GPT-4o, GPT-4-turbo, o1, o3, with full tool calling and JSON mode.
  • AWS Bedrock — Claude, Llama, Mistral, Cohere, AI21, Stability, including Converse and ConverseStream wire formats.
  • Google Gemini — Gemini Pro, Flash, with native multimodal content.
  • MiniMax, DeepSeek, Moonshot, GLM, Zhipu, Yi, Qwen — Chinese model families with OpenAI-shaped APIs.
  • Ollama, vLLM, llama.cpp, LM Studio — local model servers, no auth required.
  • OpenRouter, Together, Groq, Fireworks, Anyscale, DeepInfra — inference aggregators.
  • Plus 100+ more — every server that speaks OpenAI Chat Completions or a custom Transport.

Each provider carries its own capabilities() and defaultBaseURL. The full list is in data/registry.meta.json.

What providers can advertise

Each transport’s capabilities() returns a TransportMetadata object:

{
  name: 'openai',
  version: '0.1.0',
  capabilities: ['chat', 'streaming', 'tools', 'vision', 'embeddings', 'images', 'audio-tts', 'audio-stt', 'json-mode', 'function-calling'],
  defaultBaseURL: 'https://api.openai.com/v1',
  auth: 'bearer',
}

The capabilities array is what tells the host which methods to expose. A transport advertising only ['chat', 'streaming'] will refuse to handle embeddings() or images() calls. This is the basis for aiplug’s “no implicit fallback” guarantee — calling a method the transport doesn’t support throws an explicit AIPlugError rather than silently falling through to a different transport.

Next steps

Previous Embedding guide Next CLI reference