continue-dev

Continue.Dev is an AI Code Editing tool built on top of VSCode

Setup

Continue.dev can be used in a fully offline way with local model support via ollama. It supports a yaml configuration file structure. Here is an example of how it can be configured with a combination of Ollama on the current machine and API calls to a LiteLLM proxy server:

 
name: my-configuration
version: 0.0.1
scheme: v1
models:
  - name: qwen2.5-coder 1.5b
    provider: ollama
    model: qwen2.5-coder:1.5b
    roles:
      - chat
      - edit
      - apply
      - autocomplete
  - name: Autodetect
    provider: ollama
    model: AUTODETECT
 
  - name: Autodetect
	provider: openai
	model: AUTODETECT
	apiBase: https://litellm.yoursite.example/v1/
	apiKey: sk-some-api-key

Setting num_ctx for Ollama

By default ollama uses a context window of 2048 which is way too small for useful code assistants.

You can override the size of the context window by passing num_ctx as an extra body parameter:

  - name: Qwen 3 MoE 3x10B
    provider: ollama
    model: qwen3:30b-a3b
    roles:
      - chat
      - edit
    requestOptions:
      extraBodyProperties:
        num_ctx: 32384

It is probably worth also turning on context quantization and flash attention in Ollama first.

Agent Mode and Tool Support

Within continue’s config file you can add a block overriding the detection of tool use capability which allows you to use any model as an agent:

  - name: Qwen 3 MoE 3x10B
    provider: ollama
    model: qwen3:30b-a3b
    roles:
      - chat
      - edit
      - apply
    capabilities:
      - tool_use

Currently Continue only supports agent mode for certain models that support tool use. How do they detect whether a model supports tool usage? They have a hard coded pattern matching routine in their source code

There is an issue about this open here. I proposed a new ‘role’ that would allow people to override the tool detection.