wehub-resource-sync
145 lines
5.5 KiB
TypeScript
145 lines
5.5 KiB
TypeScript
/**
|
|
* Custom Model Provider Plugin Example
|
|
*
|
|
* Shows how a plugin can register a brand-new model provider so the agent can
|
|
* run inference against it, no changes to Cline itself required.
|
|
*
|
|
* This example registers an OpenAI-compatible provider pointed at OpenRouter.
|
|
* OpenRouter ships as a built-in provider in Cline already, so this example
|
|
* registers under a distinct id (`openrouter-plugin`) to avoid colliding with
|
|
* the built-in. It exists to demonstrate the pattern. Swap the `BASE_URL`,
|
|
* `API_KEY_ENV`, and `MODELS` below to add any OpenAI-compatible endpoint
|
|
* (a self-hosted LiteLLM/vLLM gateway, an internal proxy, a niche provider Cline
|
|
* does not bundle, etc.).
|
|
*
|
|
* How it works:
|
|
* - `Llms.registerProvider(collection)` adds the provider and its models to
|
|
* the gateway catalog. With `protocol: "openai-chat"` and
|
|
* `client: "openai-compatible"`, the gateway builds an OpenAI-compatible
|
|
* handler from the `baseUrl` and resolves the API key from the session
|
|
* config or the `env` var declared on the provider.
|
|
* - `api.registerProvider(...)` declares the contribution on the plugin
|
|
* (required by the `providers` capability) so hosts can advertise it.
|
|
*/
|
|
|
|
import { type AgentPlugin, Llms } from "@cline/core";
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Provider definition
|
|
//
|
|
// Change these constants to point at any OpenAI-compatible endpoint.
|
|
// ---------------------------------------------------------------------------
|
|
|
|
// Provider id exposed by this plugin. Kept distinct from the built-in
|
|
// `openrouter` so the two do not collide.
|
|
const PROVIDER_ID = "openrouter-plugin";
|
|
const PROVIDER_NAME = "OpenRouter (plugin)";
|
|
const BASE_URL = "https://openrouter.ai/api/v1";
|
|
/** Env var the gateway reads for the API key when one is not passed in the
|
|
* session config. */
|
|
const API_KEY_ENV = "OPENROUTER_API_KEY";
|
|
const DEFAULT_MODEL_ID = "anthropic/claude-sonnet-4.6";
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Model catalog
|
|
//
|
|
// Pricing is per 1M tokens and is illustrative; keep it in sync with the
|
|
// provider's own pricing page. `capabilities` tells the runtime what each
|
|
// model supports (tool calling, streaming, prompt caching, reasoning, images).
|
|
// ---------------------------------------------------------------------------
|
|
|
|
const MODELS: Record<string, Llms.ModelInfo> = {
|
|
"anthropic/claude-sonnet-4.6": {
|
|
id: "anthropic/claude-sonnet-4.6",
|
|
name: "Claude Sonnet 4.6",
|
|
contextWindow: 200_000,
|
|
maxTokens: 8_192,
|
|
capabilities: ["tools", "streaming", "prompt-cache", "reasoning", "images"],
|
|
pricing: { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
|
|
},
|
|
"openai/gpt-5": {
|
|
id: "openai/gpt-5",
|
|
name: "GPT-5",
|
|
contextWindow: 400_000,
|
|
maxTokens: 16_384,
|
|
capabilities: ["tools", "streaming", "reasoning", "images"],
|
|
pricing: { input: 1.25, output: 10 },
|
|
},
|
|
"google/gemini-2.5-pro": {
|
|
id: "google/gemini-2.5-pro",
|
|
name: "Gemini 2.5 Pro",
|
|
contextWindow: 1_048_576,
|
|
maxTokens: 65_536,
|
|
capabilities: ["tools", "streaming", "reasoning", "images"],
|
|
pricing: { input: 1.25, output: 10 },
|
|
},
|
|
};
|
|
|
|
function buildCollection(): Llms.ModelCollection {
|
|
const provider: Llms.ProviderInfo = {
|
|
id: PROVIDER_ID,
|
|
name: PROVIDER_NAME,
|
|
description:
|
|
"OpenAI-compatible OpenRouter endpoint registered by a plugin.",
|
|
// openai-chat protocol + openai-compatible client => the gateway builds
|
|
// an OpenAI-compatible handler from baseUrl and the resolved API key.
|
|
protocol: "openai-chat",
|
|
client: "openai-compatible",
|
|
baseUrl: BASE_URL,
|
|
defaultModelId: DEFAULT_MODEL_ID,
|
|
// env is the API key env var the gateway falls back to when the session
|
|
// config does not carry an explicit key.
|
|
env: [API_KEY_ENV],
|
|
// Provider capabilities use the provider-level schema. Image support is
|
|
// represented as "vision" here and as "images" on individual models.
|
|
capabilities: ["tools", "streaming", "prompt-cache", "reasoning", "vision"],
|
|
// "file" marks this as a user-added provider rather than a built-in
|
|
// ("system"), which affects trust-level prompts in some hosts.
|
|
source: "file",
|
|
};
|
|
|
|
return { provider, models: MODELS };
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Plugin
|
|
// ---------------------------------------------------------------------------
|
|
|
|
const plugin: AgentPlugin = {
|
|
name: "openrouter-provider",
|
|
manifest: {
|
|
capabilities: ["providers"],
|
|
// Declared so hosts that pre-validate contributions know which provider
|
|
// ids this plugin owns before importing it.
|
|
providerIds: [PROVIDER_ID],
|
|
},
|
|
|
|
setup(api, ctx) {
|
|
// 1. Register the provider + models with the gateway catalog. This is what
|
|
// makes inference actually work. The gateway can now resolve
|
|
// `openrouter-plugin` and its models and build a handler for them.
|
|
Llms.registerProvider(buildCollection());
|
|
|
|
// 2. Declare the contribution on the plugin. Required by the "providers"
|
|
// capability; lets hosts advertise the provider in pickers and surface
|
|
// its metadata.
|
|
api.registerProvider({
|
|
name: PROVIDER_ID,
|
|
description: `${PROVIDER_NAME}: OpenAI-compatible endpoint at ${BASE_URL}`,
|
|
metadata: {
|
|
baseUrl: BASE_URL,
|
|
apiKeyEnv: API_KEY_ENV,
|
|
defaultModelId: DEFAULT_MODEL_ID,
|
|
modelIds: Object.keys(MODELS),
|
|
},
|
|
});
|
|
|
|
ctx.logger?.log(
|
|
`[openrouter-provider] registered provider "${PROVIDER_ID}" with ${Object.keys(MODELS).length} model(s)`,
|
|
);
|
|
},
|
|
};
|
|
|
|
export { plugin, buildCollection };
|
|
export default plugin;
|