Skip to main content
Avante.nvim’s flexible provider system allows you to integrate any AI service that offers a REST API. This guide shows you how to create custom providers.

​
Provider Interface

A provider must implement specific functions to work with Avante:

​
Required Functions

  1. parse_curl_args(prompt_opts): Generate API request parameters
  2. parse_response(ctx, data_stream, event_state, opts): Parse streaming responses
  3. is_env_set(): Check if provider is properly configured

​
Optional Functions

  • setup(): Initialize provider (authentication, tokenizer, etc.)
  • parse_api_key(): Extract API key from environment
  • on_error(result): Handle API errors
  • list_models(): List available models

​
Creating a Custom Provider

​
Method 1: Inherit from Existing Provider

The easiest way is to inherit from an OpenAI-compatible provider: 
providers = {
  my_provider = {
    __inherited_from = "openai",
    endpoint = "https://api.myprovider.com/v1",
    model = "my-model-name",
    api_key_name = "MY_PROVIDER_API_KEY",
    extra_request_body = {
      temperature = 0.7,
      max_tokens = 4096,
    },
  },
}

​
Method 2: Custom Implementation

For non-OpenAI-compatible APIs, create a custom module: 
-- ~/.config/nvim/lua/custom/my_provider.lua

local Utils = require("avante.utils")
local Providers = require("avante.providers")

---@class AvanteProviderFunctor
local M = {}

M.api_key_name = "MY_PROVIDER_API_KEY"

M.role_map = {
  user = "user",
  assistant = "assistant",
}

function M:is_disable_stream()
  return false
end

function M:parse_messages(opts)
  local messages = {}
  
  -- Convert Avante message format to your provider's format
  for _, msg in ipairs(opts.messages) do
    table.insert(messages, {
      role = self.role_map[msg.role],
      content = msg.content,
    })
  end
  
  return messages
end

function M:parse_curl_args(prompt_opts)
  local provider_conf, request_body = Providers.parse_config(self)
  
  -- Get API key
  local api_key = self.parse_api_key()
  if not api_key then
    Utils.error("API key not found")
    return nil
  end
  
  return {
    url = provider_conf.endpoint .. "/chat/completions",
    headers = {
      ["Content-Type"] = "application/json",
      ["Authorization"] = "Bearer " .. api_key,
    },
    body = vim.tbl_deep_extend("force", {
      model = provider_conf.model,
      messages = self:parse_messages(prompt_opts),
      stream = true,
    }, request_body),
  }
end

function M:parse_response(ctx, data_stream, event_state, opts)
  local ok, json = pcall(vim.json.decode, data_stream)
  if not ok then return end
  
  -- Parse your provider's response format
  if json.choices and json.choices[1] then
    local delta = json.choices[1].delta
    if delta and delta.content then
      if opts.on_chunk then
        opts.on_chunk(delta.content)
      end
    end
    
    if json.choices[1].finish_reason then
      opts.on_stop({ reason = "complete" })
    end
  end
end

return M
Then configure it: 
providers = {
  my_provider = {
    __inherited_from = nil,
    endpoint = "https://api.myprovider.com",
    model = "my-model",
    parse_curl_args = require("custom.my_provider").parse_curl_args,
    parse_response = require("custom.my_provider").parse_response,
  },
}

​
Real-World Examples

​
OpenRouter

OpenRouter provides access to multiple models: 
providers = {
  openrouter = {
    __inherited_from = "openai",
    endpoint = "https://openrouter.ai/api/v1",
    model = "anthropic/claude-3-opus",
    api_key_name = "OPENROUTER_API_KEY",
    extra_request_body = {
      -- OpenRouter-specific parameters
      route = "fallback",
    },
  },
}

​
Mistral AI

providers = {
  mistral = {
    __inherited_from = "openai",
    endpoint = "https://api.mistral.ai/v1",
    model = "mistral-large-latest",
    api_key_name = "MISTRAL_API_KEY",
    extra_request_body = {
      max_tokens = 4096,
    },
  },
}

​
Together AI

providers = {
  together = {
    __inherited_from = "openai",
    endpoint = "https://api.together.xyz/v1",
    model = "meta-llama/Llama-3-70b-chat-hf",
    api_key_name = "TOGETHER_API_KEY",
  },
}

​
Groq

providers = {
  groq = {
    __inherited_from = "openai",
    endpoint = "https://api.groq.com/openai/v1",
    model = "llama-3.1-70b-versatile",
    api_key_name = "GROQ_API_KEY",
  },
}

​
Message Format

Avante uses a standardized message format: 
{
  role = "user" | "assistant",
  content = "string" | {
    { type = "text", text = "content" },
    { type = "image", source = { ... } },
    { type = "tool_use", name = "...", input = { ... } },
    { type = "tool_result", content = "..." },
  }
}
Your provider must convert this to your API’s format.

​
Tool Support

To support function calling: 
function M:transform_tool(tool)
  -- Convert Avante tool format to your provider's format
  local input_schema_properties, required = 
    Utils.llm_tool_param_fields_to_json_schema(tool.param.fields)
  
  return {
    name = tool.name,
    description = tool.description,
    parameters = {
      type = "object",
      properties = input_schema_properties,
      required = required,
    },
  }
end

function M:parse_curl_args(prompt_opts)
  -- ... existing code ...
  
  local tools = nil
  if prompt_opts.tools then
    tools = {}
    for _, tool in ipairs(prompt_opts.tools) do
      table.insert(tools, self:transform_tool(tool))
    end
  end
  
  return {
    -- ... existing code ...
    body = {
      -- ... existing code ...
      tools = tools,
    },
  }
end

​
Authentication Methods

​
API Key from Environment

M.api_key_name = "MY_API_KEY"

function M.parse_api_key()
  return os.getenv("AVANTE_" .. M.api_key_name) or os.getenv(M.api_key_name)
end

​
API Key from Command

M.api_key_name = "cmd:pass show my-service/api-key"

​
OAuth

See the Claude provider implementation for OAuth example: 
-- Store OAuth token
local token = {
  access_token = "...",
  refresh_token = "...",
  expires_at = os.time() + 3600,
}

-- Use in headers
headers["Authorization"] = "Bearer " .. token.access_token

​
Streaming Responses

​
Server-Sent Events (SSE)

Most providers use SSE: 
function M:parse_response(ctx, data_stream, event_state, opts)
  -- SSE format: data: {json}
  local json_str = data_stream:match("^data: (.+)$")
  if not json_str or json_str == "[DONE]" then
    opts.on_stop({ reason = "complete" })
    return
  end
  
  local ok, json = pcall(vim.json.decode, json_str)
  if ok and json.choices then
    local delta = json.choices[1].delta
    if delta and delta.content then
      opts.on_chunk(delta.content)
    end
  end
end

​
Newline-Delimited JSON

function M:parse_response(ctx, data_stream, event_state, opts)
  local ok, json = pcall(vim.json.decode, data_stream)
  if not ok then return end
  
  -- Handle your JSON format
end

​
Error Handling

function M.on_error(result)
  if result.status == 429 then
    -- Rate limit handling
    return
  end
  
  local ok, body = pcall(vim.json.decode, result.body)
  if ok and body.error then
    Utils.error(body.error.message, { title = "My Provider" })
  else
    Utils.error("Request failed: " .. result.status)
  end
end

​
Testing Your Provider

1

Set API key

export AVANTE_MY_PROVIDER_API_KEY=your-key
2

Configure provider

provider = "my_provider"
3

Test in Neovim

:AvanteAsk Hello, world!
4

Debug if needed

-- Enable debug mode
debug = true

​
Advanced Features

​
Prompt Caching

M.support_prompt_caching = true

function M:parse_curl_args(prompt_opts)
  -- Mark content for caching
  local messages = self:parse_messages(prompt_opts)
  if self.support_prompt_caching then
    messages[#messages].cache_control = { type = "ephemeral" }
  end
  
  return { body = { messages = messages } }
end

​
Model Listing

function M:list_models()
  local curl = require("plenary.curl")
  local response = curl.get(provider_conf.endpoint .. "/models", {
    headers = {
      ["Authorization"] = "Bearer " .. self.parse_api_key(),
    },
  })
  
  if response.status == 200 then
    local data = vim.json.decode(response.body)
    return vim.tbl_map(function(model)
      return {
        id = model.id,
        name = model.name,
        provider_name = "my_provider",
      }
    end, data.models)
  end
  
  return {}
end

​
Troubleshooting

Error: “Failed to find provider: my_provider”Ensure the provider is defined in providers table.
Error: “missing the __inherited_from attribute or a custom parse_curl_args function”Either:
  1. Set __inherited_from = "openai"
  2. Or implement parse_curl_args function
Ensure environment variable is set:
echo $AVANTE_MY_PROVIDER_API_KEY

​
Best Practices

Inherit When Possible

Use __inherited_from for OpenAI-compatible APIs to reduce code.

Handle Errors

Implement proper error handling and user-friendly messages.

Support Streaming

Implement streaming for better UX and responsiveness.

Document Your Provider

Add comments explaining API-specific quirks and requirements.

​
Complete Example

Here’s a complete custom provider: 
-- ~/.config/nvim/lua/custom/my_provider.lua

local Utils = require("avante.utils")
local Providers = require("avante.providers")

---@class AvanteProviderFunctor
local M = {}

M.api_key_name = "MY_PROVIDER_API_KEY"
M.role_map = { user = "user", assistant = "assistant" }

function M:is_disable_stream() return false end

function M:parse_messages(opts)
  local messages = {{ role = "system", content = opts.system_prompt }}
  for _, msg in ipairs(opts.messages) do
    table.insert(messages, {
      role = self.role_map[msg.role],
      content = type(msg.content) == "string" and msg.content or "",
    })
  end
  return messages
end

function M:parse_curl_args(prompt_opts)
  local provider_conf, request_body = Providers.parse_config(self)
  local api_key = self.parse_api_key()
  
  if not api_key then
    Utils.error("API key not set")
    return nil
  end
  
  return {
    url = provider_conf.endpoint .. "/chat/completions",
    headers = {
      ["Content-Type"] = "application/json",
      ["Authorization"] = "Bearer " .. api_key,
    },
    body = vim.tbl_deep_extend("force", {
      model = provider_conf.model,
      messages = self:parse_messages(prompt_opts),
      stream = true,
    }, request_body),
  }
end

function M:parse_response(ctx, data_stream, event_state, opts)
  if data_stream == "[DONE]" then
    opts.on_stop({ reason = "complete" })
    return
  end
  
  local ok, json = pcall(vim.json.decode, data_stream)
  if ok and json.choices and json.choices[1] then
    local delta = json.choices[1].delta
    if delta and delta.content then
      if opts.on_chunk then opts.on_chunk(delta.content) end
    end
    if json.choices[1].finish_reason then
      opts.on_stop({ reason = "complete" })
    end
  end
end

return M
Configuration: 
providers = {
  my_provider = {
    endpoint = "https://api.example.com",
    model = "my-model-v1",
    parse_curl_args = require("custom.my_provider").parse_curl_args,
    parse_response = require("custom.my_provider").parse_response,
    extra_request_body = {
      temperature = 0.7,
      max_tokens = 4096,
    },
  },
}

Build docs developers (and LLMs) love

Get started for freeTalk to us