Local Copilot with Cline: AI agent in VS Code (Ollama)

#Why a local Copilot

Privacy

No code snippet leaves your computer. NDA, trade secrets, client code: everything stays local.

Offline

On a plane, deep in a lab without Internet, on vacation. The model is on your drive.

Cost

Free after buying the GPU. A developer who uses Copilot for two years covers the cost of a mid-range graphics card.

Model selection

You can swap Qwen 3.5 for Devstral, Qwen3-Coder, or GLM 4.7 Flash depending on the language and your VRAM.

#Why this guide switched to Cline

This guide originally relied on Continue.dev. Continue was acquired by Cursor in June 2026, its repository is now archived, and the standalone product is no longer maintained, so we've updated it for Cline, an equivalent, active open-source assistant that is also 100% local through Ollama.

#Models to use

⭐ Qwen 3.5 9B

THE 8 GB choice in 2026. Multilingual (including FR), 256k context, vision, Apache 2.0. Good for chat and for Python, TypeScript, and Rust agents. ~6.6 GB in Q4.

Devstral 24B (Mistral AI)

Specialist in code agents, Apache 2.0. THE 16 GB choice for Cline's multi-file edits. ~14 GB in Q4.

Qwen3-Coder 30B-A3B

Code MoE (3B active, so fast), 256k context. For 24 GB: high-end chat + agent that rivals Copilot. ~19 GB in Q4.

Qwen 3.8 27B

The “Copilot-like” model of 2026 (released on 14/08/2026), with 262k context, vision, and Apache 2.0. For 24 GB—switch to “low” reasoning so it does not overthink.

#1. Install Cline

1. 01
   Install Ollama

   If you haven't already, see the guide Install Ollama on Windows/macOS/Linux. This is the engine that runs the model in the background.
2. 02
   Download the model

   Get a coder model suited to your VRAM with the command below.
3. 03
   Install the Cline extension

   In VS Code: Extensions (Ctrl+Shift+X) → search for Cline → Install. In JetBrains, use the plugin marketplace.
4. 04
   Open the panel

   A Cline icon appears in the left sidebar. Click it to open the chat and agent panel.

# Modèle principal (chat + agent)
ollama pull qwen3.5:9b

# Plus de VRAM ? Devstral, le spécialiste du mode agent (16 Go)
ollama pull devstral:24b

#2. Configure the Ollama provider

Cline needs to know where to find the model. Open its settings (the gear icon in the panel), select the Ollama provider, enter the daemon’s local address, and choose the model from the detected list.

Provider     : Ollama
Base URL     : http://localhost:11434
Modèle       : qwen3.5:9b

#3. Chat mode

Chat mode is the direct equivalent of a local Copilot Chat panel. Ask a question, paste code, request an explanation, or ask for a fix.

Open chat

Click the Cline icon in the sidebar. The conversation appears in the panel.

Attach code

Select a block in the editor, then send it to the chat context so the model can reason about it.

Mention files

Cline can reference files in your project to answer with the right context, without requiring you to copy and paste everything.

Keep the history local

Your conversations stay on your machine. Nothing is synced to the cloud.

#4. Agent mode

This is where Cline goes beyond a simple chat. In agent mode, it can read the directory tree, open files, suggest multi-file changes, and run commands—each action requires your approval.

Plan, then execution

Describe a task in French (“add a /health endpoint and its test”). Cline proposes a plan, then applies the changes step by step.

Diff validated

Each change is presented as a diff that you accept or reject. You remain in control at all times.

Commands in boxes

The agent may suggest running a command (tests, build). Nothing runs without your explicit approval.

- Ajoute du logging pour tracer les entrées/sorties
- Convertis cette fonction en async/await
- Extrais ce bloc dans un fichier util.ts et mets à jour les imports
- Écris des tests unitaires pour les cas limites

#5. Autocomplete: Tabby or Twinny

Important point: Cline is a chat assistant plus agent; it does not provide inline “grayed-out text” autocomplete while you type. To regain a Copilot-like experience, add a dedicated extension that also connects to Ollama.

Tabby

Self-hosted autocompletion extension, ideal if several developers share the same GPU. Line-by-line completion appears in gray; press Tab to accept.

Twinny

Lightweight VS Code extension, autocompletion + chat, connecting directly to Ollama. A good solo complement to Cline for high-volume typing.

The pair that works

Cline for chat and agent use, Tabby or Twinny for inline completion. Both point to the same Ollama, and your code remains local on both sides.

#Performance and advanced settings

High num_ctx

In Ollama, set num_ctx to 16384 so the agent can send a lot of context (files, diffs). Via a Modelfile.

Flash Attention

On RTX 30/40/50 cards, enable FA in Ollama (OLLAMA_FLASH_ATTENTION=1 variable). About +20% speed.

Quantized KV cache

OLLAMA_KV_CACHE_TYPE=q8_0 halves the context's VRAM usage. Crucial when the agent works with large files.

One model per role

A dedicated FIM model (qwen2.5-coder:7b-base, ~4.7 GB) for autocompletion (Tabby/Twinny), a 24B/30B model (Devstral, Qwen3-Coder) for chat and the Cline agent. Ollama loads the one requested.

# Modelfile : devstral-16k
FROM devstral:24b
PARAMETER num_ctx 16384

ollama create devstral-16k -f Modelfile
# puis sélectionnez devstral-16k comme modèle dans Cline

#Frequently asked questions

#Conclusion

In about twenty minutes, you have a 100% local code copilot: Cline for chat and agent mode, a recent code model (Qwen 3.5 9B, or Devstral 24B for the agent) running through Ollama as the engine, and Tabby or Twinny (with qwen2.5-coder:7b-base) for inline autocompletion. Your code never leaves your machine, you work offline, and you remain free to switch models based on your languages. If you want to save hours of configuration and start from a proven foundation, the paid “Local Code Copilot” guide provides a complete Ollama + Cline + Aider configuration pack, ready to use.