Free local copilot: Cline, Tabby & CodeGeeX in VS Code (2026)

#Why use a free local Copilot for VS Code?

Copilot and Cursor send your code to OpenAI or Anthropic. On an NDA-covered client project, with proprietary code, or simply as a matter of principle, that isn’t always acceptable. A local copilot solves three problems at once: zero leakage (nothing leaves the laptop), zero subscription fees, and zero network latency.

The tradeoff is straightforward: the quality of a local Qwen 3.5 9B does not match the very latest Copilot Pro model. But for everyday tasks—completing a loop, writing a test, refactoring a 30-line function—a 9B to 24B model in Q4 gets the job done. And for tasks the local model cannot handle, you can keep Copilot running alongside it.

#Prerequisites

Recent VS Code

Version 1.85+ for the inline completion API. JetBrains extensions also exist for all three tools.

Ollama installed

Serves as the inference backend for Cline. Listens by default on http://localhost:11434. For Tabby, the runtime is included; CodeGeeX can point to Ollama.

GPU with 8 GB VRAM minimum

RTX 3060 12 GB, 4060 Ti 16 GB, unified-memory Mac M2/M3 16 GB. Below that, stick to 1.5B models (qwen2.5-coder:1.5b) for completion; they remain responsive.

20 GB of disk space

A 7B coder in Q4 = 4-5 GB. Allow plenty of headroom if you're testing several models or tools.

#1. Cline (chat + agent)

Cline is the most complete option for assistance: a sidebar chat and agent mode capable of reading and modifying multiple files, Cursor-style, all locally through Ollama. Open source under the MIT license and BYO-LLM, it is the natural replacement for Continue.dev. Roo Code, a fork of Cline, is also shutting down: Cline therefore becomes the consolidation point for this ecosystem. Its limitation: it does not provide inline autocompletion — Tabby covers that lower down.

1. 01
   Install the extension

   In VS Code, open the extensions palette (Ctrl+Shift+X), search for Cline, and install it. A Cline icon appears in the left sidebar.
2. 02
   Download a model

   Qwen 3.5 9B handles chat and agent tasks respectably on 8 GB of VRAM. Move up to Devstral 24B or Qwen 3.6 35B-A3B if you have the headroom; Devstral is specifically designed for agentic workloads.
3. 03
   Configure the Ollama provider

   Open Cline's settings, choose the Ollama provider, enter the local URL, and select the model.
4. 04
   Test the chat

   Ask a question, paste code, or request a fix. The response appears in the panel without sending anything to the cloud.
5. 05
   Test the agent

   Describe a multi-file task. Cline proposes a plan, then applies the changes as diffs that you approve one by one.

ollama pull qwen3.5:9b
# Plus de VRAM ? Devstral, spécialiste agent de code, suit mieux le mode agent
ollama pull devstral:24b

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

#2. Tabby (self-hosted autocomplete)

Tabby (TabbyML) covers exactly what Cline does not: inline autocompletion. It is designed for teams—a single server hosts the model, and all developers connect to it from their IDE. The server includes its own runtime, so it does not need Ollama.

1. 01
   Launch the Tabby server

   The Docker method is the simplest. The server exposes an HTTP endpoint on :8080 and a web admin interface.
2. 02
   Create an admin account

   Open http://localhost:8080 in your browser. On first connection, create the owner account. Go to Settings → Information to retrieve an auth token.
3. 03
   Install the VS Code extension

   In the extensions palette, search for Tabby (publisher: TabbyML). Install it.
4. 04
   Connect the extension

   Open Settings (Ctrl+,), search for tabby. Set Endpoint = http://localhost:8080 and paste the token. The VS Code status bar shows Tabby: ready; completions appear grayed out as you type.

docker run -d --name tabby \
  --gpus all -p 8080:8080 \
  -v $HOME/.tabby:/data \
  registry.tabbyml.com/tabbyml/tabby \
  serve --model StarCoder-1B --device cuda

#3. CodeGeeX

CodeGeeX (Tsinghua KEG / Zhipu AI) is the “all-in-one” option: the extension already includes completion and chat with a free proprietary model hosted by them. But it can also connect to a local model, which is what interests us here.

1. 01
   Install the extension

   In VS Code, search for CodeGeeX (publisher: aminer). Install it. A CodeGeeX icon appears in the sidebar.
2. 02
   Switch to local mode

   Open the CodeGeeX settings (the gear icon in the side panel). Find the Local mode section and enable it.
3. 03
   Point to Ollama

   Enter the local model URL. CodeGeeX speaks an OpenAI-compatible endpoint—the one from Ollama on http://localhost:11434/v1 will work.
4. 04
   Choose the model

   codegeex4 (9B) is the built-in model available in Ollama. Otherwise, any recent coder will do (qwen3-coder, devstral).

ollama pull codegeex4

#Comparison table

Use-case coverage

Cline: chat + agent (no completion). Tabby: completion (no advanced chat). CodeGeeX: completion + basic chat. The Cline + Tabby combo covers everything.

Setup effort

CodeGeeX (5 min) < Cline + Ollama (10 min) < Tabby self-hosted (15–20 min).

Multi-utilisateurs

Only Tabby is designed for this natively. Cline/CodeGeeX = one workstation = one Ollama.

Privacy

Cline and Tabby remain local by default. CodeGeeX goes to the cloud until local mode is enabled—keep an eye on that.

#Troubleshooting

Cline is not responding

Check ollama ps: the model should appear there when you submit a request. If Ollama does not respond, test curl http://localhost:11434/api/tags from the terminal. Also check that the provider selected in Cline is Ollama.

No grayed-out completion

Cline doesn't provide inline autocompletion: that's normal. Install Tabby (or Twinny) for line-by-line completion, and check editor.inlineSuggest.enabled in VS Code.

Tabby displays “Connection refused”

The container isn't running, or the port isn't exposed. docker ps should list tabby. docker logs tabby shows the cause (often: the model wasn't downloaded on first launch; wait 5-10 min).

CodeGeeX returns answers in Chinese

Specify the language in the system prompt (Settings CodeGeeX → Custom system prompt → "Toujours répondre en français"), or use Qwen 3.5 9B, which responds in French when addressed in French.

VRAM saturated when everything is running

You may have two models loaded (Tabby autocompletion + Cline chat). On 8 GB, keep a lightweight FIM model for completion and a Qwen 3.5 9B for chat. ollama ps displays usage.

#FAQ

#Go further

You have a local copilot up and running. Three logical directions depending on your profile:

Learn more about Cline

The dedicated guide to local Copilot with Cline goes deeper into the configuration: provider Ollama, agent mode, autocompletion via Tabby/Twinny, and fine-tuning the settings for agentic workflows.

You came from Continue.dev

The Continue → Cline migration guide details the transition (equivalent settings, models, and local data).

Choose your hardware

Before moving up to a 14B or 32B model, the Choose Your GPU for Local AI guide helps you avoid buying too little VRAM.

If you want to save time on configuration, the paid “Local Code Copilot” guide provides a complete Ollama + Cline + Aider configuration pack, ready to paste, for both a solo workstation and a team.