Agent mode, get your VS code to work similarly to Cursor in linux

[bizmate]

Hi all

I am stuck, i have been trying with smaller models and several types of configs from json to yaml and ultimately i get nowhere with both asking how to configure continue in my machine to talk to ollama and act as a fully fledged agent. It would be great if continue could provide a baseline to get an agent working in visual studio....

Page at https://docs.continue.dev/docs/ide-extensions/autocomplete/model-setup#model-recommendations seems to be dead
And all attempts i have tried yield responses like

``I'm sorry but as an AI language model, I don't have access to specific projects or their codebases. If you can provide more information about the project and what kind of files you are looking for, I'd be happy to help you find them.

and various error in the logs, if anything at all.

I found now the page at https://docs.continue.dev/guides/ollama-guide#method-3-manual-configuration
but i was wondering if it was possible to ask for a confirmation of how to get this plugin to talk to my ollama running local models so that indeed it behaves as close as possible to cursor.

My machine is not massive but still has an RTX 2070 and whilst following the docs or asking online I am still unable to get it work

[ibondarenko1]

Cursor parity from Continue + Ollama is achievable but you have to wire four separate roles correctly - that is the part the docs gloss over. Here is a working config that gets you there on a 2070 (8GB VRAM is fine if you size models right).

1. The four model roles you need filled.

Continue separates models by role. Cursor-equivalent behaviour needs all four:

Role	What it does	Recommended for RTX 2070
chat	the conversation panel	qwen2.5-coder:7b-instruct
edit	inline edit / Cmd-I	same as chat works
apply	applies suggested patches	small + fast: qwen2.5-coder:1.5b
autocomplete	tab-completion	qwen2.5-coder:1.5b-base
embed	codebase RAG	nomic-embed-text

If only chat is configured, the Cmd-I and tab paths fall back to broken / generic behaviour - which produces exactly the "I don't have access to your project" replies you're seeing.

2. Working YAML config (drop into ~/.continue/config.yaml).

name: Local Cursor-like
version: 0.1.0
schema: v1

models:
  - name: chat-qwen
    provider: ollama
    model: qwen2.5-coder:7b-instruct
    apiBase: http://localhost:11434
    roles: [chat, edit, apply]

  - name: autocomplete-qwen
    provider: ollama
    model: qwen2.5-coder:1.5b-base
    apiBase: http://localhost:11434
    roles: [autocomplete]
    defaultCompletionOptions:
      maxTokens: 256
      temperature: 0.1

  - name: embed-nomic
    provider: ollama
    model: nomic-embed-text
    apiBase: http://localhost:11434
    roles: [embed]

context:
  - provider: code
  - provider: codebase
  - provider: diff
  - provider: terminal
  - provider: open

Pull the models first:

ollama pull qwen2.5-coder:7b-instruct
ollama pull qwen2.5-coder:1.5b-base
ollama pull nomic-embed-text

3. The "I don't have access to your project" failure is almost always missing context providers.

Continue passes only what you've named in context: to the model. Without codebase and code, the model sees the literal user message and nothing else - it is correct when it says it has no access. Adding codebase triggers the embedding-based RAG over your repo (which is why embed-nomic has to be configured first - silent missing-embed-model is one of the most common dead-ends).

4. Confirm the agent mode is actually engaging with tools.

After loading a workspace, in the chat panel use @codebase: where is the user authentication handled? - if Continue replies with file paths and snippets, the embedding indexer worked. If it replies with the "I'm an AI language model..." apology, the indexer hasn't run yet or embed role isn't filled. Force re-index from the command palette: Continue: Reindex codebase.

5. If you want true agent (multi-step tool use, not just chat-with-context), enable Agent mode and add MCP tools.

mcpServers:
  - name: filesystem
    type: stdio
    command: uvx
    args: ["mcp-server-filesystem", "/path/to/your/project"]

Switch the chat panel mode dropdown from "Chat" to "Agent" - now the model can call tools (read_file, write_file, list_directory) explicitly rather than hallucinating filenames. This is the closest analogue to Cursor's agent mode in Continue today.

Reality-check on what you cannot get from local-only setup. Cursor's "Composer" multi-file edit feels different because it relies on a fast remote model. With a 7b local model you get correct single-file edits and decent autocomplete; multi-file refactors land slower and benefit a lot from temperature 0 + small maxTokens. Don't expect feature parity at GPU equivalence - just expect it to actually work.

Recipe: pull all 3 ollama models -> drop the YAML above into ~/.continue/config.yaml -> add context: providers explicitly -> force-reindex once -> verify with @codebase query -> switch to Agent mode + filesystem MCP for tool use.