@coralbricks/ai-sdk-provider

Vercel AI SDK provider for CoralBricks — open-model inference (GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash) with up to 1M-token context and free cached input.

A thin, typed wrapper over @ai-sdk/openai-compatible that presets the gateway endpoint and model catalog, so it works with zero configuration like the first-party @ai-sdk/* providers.

Install

npm install @coralbricks/ai-sdk-provider ai

Usage

Set CORALBRICKS_API_KEY (get a key at coralbricks.ai/api-keys), then:

import { generateText } from 'ai'
import { coralbricks } from '@coralbricks/ai-sdk-provider'

const { text } = await generateText({
  model: coralbricks('glm-5.3-fp4'),
  prompt: 'Explain prompt caching in one sentence.',
})

Streaming, tools, and structured output work through the standard AI SDK functions (streamText, generateObject, …) — CoralBricks speaks the OpenAI wire format, including streaming tool calls.

Custom instance

import { createCoralBricks } from '@coralbricks/ai-sdk-provider'

const coralbricks = createCoralBricks({
  apiKey: process.env.MY_KEY,          // default: CORALBRICKS_API_KEY
  baseURL: 'https://inference.coralbricks.ai/v1', // default
  headers: { 'x-my-header': '1' },
})

Models

| Model | Model ID | Context | Input $/M | Cache Write $/M | Output $/M | | --- | --- | --- | --- | --- | --- | | GLM 5.3 | glm-5.3-fp4 | 1M | $1.12 | $1.68 | $4.40 | | GLM 5.3 Flash | glm-5.3-flash-fp4 | 1M | $0.15 | $0.23 | $0.50 | | DeepSeek V4.1 Flash | deepseek-v4.1-flash-fast-fp4 | 1M | $0.30 | $0.09 | $1.20 |

Cached input tokens are always free; prompt tokens CoralBricks has not cached yet bill once at the Cache Write rate, in place of the Input rate. Rates per coralbricks.ai/pricing. The model-id type is open (string & {}), so a newly launched model works before this package updates.

License

Apache-2.0