Custom LLM Providers

Create custom provider integrations by implementing the ApiHandler interface. This guide shows you how to add support for any LLM API.

Provider Interface

All providers must implement the ApiHandler interface:

export interface ApiHandler {
  createMessage(
    systemPrompt: string,
    messages: ClineStorageMessage[],
    tools?: ClineTool[],
    useResponseApi?: boolean
  ): ApiStream
  
  getModel(): ApiHandlerModel
  getApiStreamUsage?(): Promise<ApiStreamUsageChunk | undefined>
  abort?(): void
}

Basic Template

Here’s a minimal custom provider:

import { ApiHandler, CommonApiHandlerOptions } from "../index"
import { ApiStream } from "../transform/stream"
import { withRetry } from "../retry"
import { ClineStorageMessage } from "@/shared/messages/content"
import type { ModelInfo } from "@shared/api"

interface CustomProviderOptions extends CommonApiHandlerOptions {
  apiKey: string
  baseUrl?: string
  modelId: string
}

export class CustomProviderHandler implements ApiHandler {
  private options: CustomProviderOptions
  private client: any // Your API client

  constructor(options: CustomProviderOptions) {
    this.options = options
  }

  private ensureClient() {
    if (!this.client) {
      if (!this.options.apiKey) {
        throw new Error("API key is required")
      }
      // Initialize your client
      this.client = createYourClient({
        apiKey: this.options.apiKey,
        baseURL: this.options.baseUrl
      })
    }
    return this.client
  }

  @withRetry()
  async *createMessage(
    systemPrompt: string,
    messages: ClineStorageMessage[],
    tools?: any[]
  ): ApiStream {
    const client = this.ensureClient()
    
    // Convert messages to your API format
    const apiMessages = this.convertMessages(messages)
    
    // Call your API
    const stream = await client.chat.create({
      model: this.options.modelId,
      messages: [
        { role: "system", content: systemPrompt },
        ...apiMessages
      ],
      stream: true,
      tools: tools
    })

    // Process stream and yield chunks
    for await (const chunk of stream) {
      if (chunk.content) {
        yield {
          type: "text",
          text: chunk.content
        }
      }
      
      if (chunk.usage) {
        yield {
          type: "usage",
          inputTokens: chunk.usage.input_tokens || 0,
          outputTokens: chunk.usage.output_tokens || 0
        }
      }
    }
  }

  getModel(): { id: string; info: ModelInfo } {
    return {
      id: this.options.modelId,
      info: {
        maxTokens: 8192,
        contextWindow: 128000,
        supportsPromptCache: false,
        inputPrice: 0,
        outputPrice: 0
      }
    }
  }

  private convertMessages(messages: ClineStorageMessage[]) {
    // Convert from HAI Build format to your API format
    return messages.map(msg => ({
      role: msg.role,
      content: msg.content
    }))
  }
}

Stream Chunk Types

Your provider should yield these chunk types:

Text Chunks

yield {
  type: "text",
  text: "Response text here"
}

Tool Call Chunks

yield {
  type: "tool_calls",
  tool_call: {
    function: {
      id: "call_123",
      name: "function_name",
      arguments: '{"param": "value"}'
    }
  }
}

Reasoning Chunks

yield {
  type: "reasoning",
  reasoning: "Model's thinking process",
  signature?: "optional_signature"
}

Usage Chunks

yield {
  type: "usage",
  inputTokens: 100,
  outputTokens: 50,
  cacheReadTokens?: 20,
  cacheWriteTokens?: 0,
  totalCost?: 0.0015
}

Message Conversion

Convert HAI Build messages to your API format:

private convertMessages(messages: ClineStorageMessage[]) {
  return messages.map(msg => {
    switch (msg.type) {
      case "say":
        return {
          role: msg.say === "user" ? "user" : "assistant",
          content: msg.text || ""
        }
      
      case "ask":
        return {
          role: "user",
          content: msg.ask === "tool" 
            ? `Tool: ${msg.text}`
            : msg.text || ""
        }
      
      default:
        return {
          role: "user",
          content: JSON.stringify(msg)
        }
    }
  })
}

Tool Calling Support

If your API supports tools:

async *createMessage(
  systemPrompt: string,
  messages: ClineStorageMessage[],
  tools?: ClineTool[]
): ApiStream {
  const client = this.ensureClient()
  
  // Convert tools to your API format
  const apiTools = tools?.map(tool => ({
    type: "function",
    function: {
      name: tool.name,
      description: tool.description,
      parameters: tool.input_schema
    }
  }))

  const stream = await client.chat.create({
    model: this.options.modelId,
    messages: [...],
    tools: apiTools
  })

  // Process tool calls in stream
  for await (const chunk of stream) {
    if (chunk.tool_calls) {
      for (const toolCall of chunk.tool_calls) {
        yield {
          type: "tool_calls",
          tool_call: {
            function: {
              id: toolCall.id,
              name: toolCall.function.name,
              arguments: toolCall.function.arguments
            }
          }
        }
      }
    }
  }
}

Error Handling

Use the @withRetry() decorator for automatic retries:

import { withRetry, RetriableError } from "../retry"

@withRetry({
  maxRetries: 3,
  baseDelay: 1000,
  maxDelay: 10000
})
async *createMessage(...): ApiStream {
  try {
    // Your API call
  } catch (error) {
    // Retry on rate limits
    if (error.status === 429) {
      const retryAfter = error.headers?.['retry-after']
      throw new RetriableError(
        error.message,
        retryAfter ? parseInt(retryAfter) : undefined
      )
    }
    // Don't retry on auth errors
    if (error.status === 401) {
      throw error
    }
    // Retry on server errors
    if (error.status >= 500) {
      throw new RetriableError(error.message)
    }
    throw error
  }
}

Prompt Caching

If your API supports prompt caching:

async *createMessage(
  systemPrompt: string,
  messages: ClineStorageMessage[],
  tools?: ClineTool[]
): ApiStream {
  const stream = await client.chat.create({
    model: this.options.modelId,
    messages: [
      {
        role: "system",
        content: systemPrompt,
        cache_control: { type: "ephemeral" } // Cache system prompt
      },
      ...apiMessages
    ]
  })

  for await (const chunk of stream) {
    if (chunk.usage) {
      yield {
        type: "usage",
        inputTokens: chunk.usage.input_tokens || 0,
        outputTokens: chunk.usage.output_tokens || 0,
        cacheReadTokens: chunk.usage.cache_read_input_tokens || 0,
        cacheWriteTokens: chunk.usage.cache_creation_input_tokens || 0
      }
    }
  }
}

Reasoning Models

For models with reasoning capabilities:

interface CustomProviderOptions extends CommonApiHandlerOptions {
  apiKey: string
  modelId: string
  thinkingBudgetTokens?: number
  reasoningEffort?: string
}

async *createMessage(
  systemPrompt: string,
  messages: ClineStorageMessage[],
  tools?: ClineTool[]
): ApiStream {
  const stream = await client.chat.create({
    model: this.options.modelId,
    messages: [...],
    thinking: this.options.thinkingBudgetTokens
      ? {
          enabled: true,
          budget_tokens: this.options.thinkingBudgetTokens
        }
      : undefined
  })

  for await (const chunk of stream) {
    // Handle reasoning content
    if (chunk.thinking) {
      yield {
        type: "reasoning",
        reasoning: chunk.thinking.text,
        signature: chunk.thinking.signature
      }
    }
    
    if (chunk.content) {
      yield {
        type: "text",
        text: chunk.content
      }
    }
  }
}

Register Your Provider

Add your provider to the system:

// In src/core/api/index.ts

import { CustomProviderHandler } from "./providers/custom-provider"

function createHandlerForProvider(
  apiProvider: string | undefined,
  options: Omit<ApiConfiguration, "apiProvider">,
  mode: Mode
): ApiHandler {
  switch (apiProvider) {
    // ... existing providers ...
    
    case "custom-provider":
      return new CustomProviderHandler({
        onRetryAttempt: options.onRetryAttempt,
        apiKey: options.customProviderApiKey,
        baseUrl: options.customProviderBaseUrl,
        modelId: mode === "plan" 
          ? options.planModeCustomProviderModelId 
          : options.actModeCustomProviderModelId
      })
    
    default:
      // ...
  }
}

Real-World Example: Ollama Provider

Here’s the actual Ollama provider implementation:

import { Ollama } from "ollama"
import { ApiHandler, CommonApiHandlerOptions } from "../index"
import { withRetry } from "../retry"
import { convertToOllamaMessages } from "../transform/ollama-format"
import { ToolCallProcessor } from "../transform/tool-call-processor"

interface OllamaHandlerOptions extends CommonApiHandlerOptions {
  ollamaBaseUrl?: string
  ollamaApiKey?: string
  ollamaModelId?: string
  ollamaApiOptionsCtxNum?: string
  requestTimeoutMs?: number
}

export class OllamaHandler implements ApiHandler {
  private options: OllamaHandlerOptions
  private client: Ollama | undefined

  constructor(options: OllamaHandlerOptions) {
    this.options = options
  }

  private ensureClient(): Ollama {
    if (!this.client) {
      const clientOptions = {
        host: this.options.ollamaBaseUrl,
        headers: this.options.ollamaApiKey
          ? { Authorization: `Bearer ${this.options.ollamaApiKey}` }
          : undefined
      }
      this.client = new Ollama(clientOptions)
    }
    return this.client
  }

  @withRetry({ retryAllErrors: true })
  async *createMessage(
    systemPrompt: string,
    messages: ClineStorageMessage[],
    tools?: any[]
  ): ApiStream {
    const client = this.ensureClient()
    const ollamaMessages = [
      { role: "system", content: systemPrompt },
      ...convertToOllamaMessages(messages)
    ]

    const stream = await client.chat({
      model: this.options.ollamaModelId || "",
      messages: ollamaMessages,
      stream: true,
      options: {
        num_ctx: Number(this.options.ollamaApiOptionsCtxNum || 32768)
      },
      tools: tools as any
    })

    const toolCallProcessor = new ToolCallProcessor()

    for await (const chunk of stream) {
      const delta = chunk.message

      if (delta?.tool_calls) {
        yield* toolCallProcessor.processToolCallDeltas(
          delta.tool_calls.map((tc, inx) => ({
            index: inx,
            id: `ollama-tool-${inx}`,
            function: {
              name: tc.function.name,
              arguments: typeof tc.function.arguments === "string"
                ? tc.function.arguments
                : JSON.stringify(tc.function.arguments)
            },
            type: "function"
          }))
        )
      }

      if (typeof delta.content === "string") {
        yield {
          type: "text",
          text: delta.content
        }
      }

      if (chunk.eval_count || chunk.prompt_eval_count) {
        yield {
          type: "usage",
          inputTokens: chunk.prompt_eval_count || 0,
          outputTokens: chunk.eval_count || 0
        }
      }
    }
  }

  getModel() {
    return {
      id: this.options.ollamaModelId || "",
      info: {
        maxTokens: 8192,
        contextWindow: Number(this.options.ollamaApiOptionsCtxNum || 32768),
        supportsPromptCache: false,
        inputPrice: 0,
        outputPrice: 0
      }
    }
  }

  abort(): void {
    this.client?.abort()
  }
}

Testing Your Provider

Create tests for your provider:

import { CustomProviderHandler } from "../custom-provider"

describe("CustomProviderHandler", () => {
  it("should stream text responses", async () => {
    const handler = new CustomProviderHandler({
      apiKey: "test-key",
      modelId: "test-model"
    })

    const chunks = []
    for await (const chunk of handler.createMessage(
      "You are a helpful assistant",
      [{ type: "say", say: "user", text: "Hello" }]
    )) {
      chunks.push(chunk)
    }

    expect(chunks.some(c => c.type === "text")).toBe(true)
    expect(chunks.some(c => c.type === "usage")).toBe(true)
  })

  it("should handle tool calls", async () => {
    const handler = new CustomProviderHandler({
      apiKey: "test-key",
      modelId: "test-model"
    })

    const tools = [{
      name: "test_tool",
      description: "Test tool",
      input_schema: {
        type: "object",
        properties: {}
      }
    }]

    const chunks = []
    for await (const chunk of handler.createMessage(
      "System prompt",
      [{ type: "say", say: "user", text: "Use the tool" }],
      tools
    )) {
      chunks.push(chunk)
    }

    expect(chunks.some(c => c.type === "tool_calls")).toBe(true)
  })
})

Best Practices

Client Initialization

Lazy-initialize clients in ensureClient()
Validate required options (API key, model ID)
Add proxy support if needed
Include proper headers (User-Agent, etc.)

Error Handling

Use @withRetry() for network/rate limit errors
Throw immediately on auth errors (401, 403)
Include error context (status code, message)
Log errors for debugging

Streaming

Yield chunks as soon as available
Don’t buffer entire response
Handle partial JSON in tool calls
Report usage at end of stream

Message Conversion

Preserve message order
Handle all message types
Convert tool results correctly
Sanitize invalid content

Performance

Reuse HTTP clients
Enable keep-alive
Support request cancellation
Add timeouts for long requests

Reference Implementations

Study these providers for examples:

Simple: ollama.ts - Basic streaming with tools
Advanced: anthropic.ts - Caching, thinking, retries
Complex: openai.ts - Azure, reasoning models, multiple formats
Gateway: openrouter.ts - Cost tracking, error handling

Next Steps

Provider Overview

Understand the provider system

Anthropic Provider

Reference implementation

OpenAI Provider

Complex provider example

OpenRouter Provider

Gateway pattern

Extension API

CLI Reference

LLM Providers

Provider Interface

Basic Template

Stream Chunk Types

Text Chunks

Tool Call Chunks

Reasoning Chunks

Usage Chunks

Message Conversion

Tool Calling Support

Error Handling

Prompt Caching

Reasoning Models

Register Your Provider

Real-World Example: Ollama Provider

Testing Your Provider

Best Practices

Reference Implementations

Next Steps

Provider Overview

Anthropic Provider

OpenAI Provider

OpenRouter Provider

Build docs developers (and LLMs) love

Extension API

CLI Reference

LLM Providers

Documentation Index

​Provider Interface

​Basic Template

​Stream Chunk Types

​Text Chunks

​Tool Call Chunks

​Reasoning Chunks

​Usage Chunks

​Message Conversion

​Tool Calling Support

​Error Handling

​Prompt Caching

​Reasoning Models

​Register Your Provider

​Real-World Example: Ollama Provider

​Testing Your Provider

​Best Practices

​Reference Implementations

​Next Steps

Provider Overview

Anthropic Provider

OpenAI Provider

OpenRouter Provider

Build docs developers (and LLMs) love

Provider Interface

Basic Template

Stream Chunk Types

Text Chunks

Tool Call Chunks

Reasoning Chunks

Usage Chunks

Message Conversion

Tool Calling Support

Error Handling

Prompt Caching

Reasoning Models

Register Your Provider

Real-World Example: Ollama Provider

Testing Your Provider

Best Practices

Reference Implementations

Next Steps