Cotab
English | 日本語
This VS Code extension is an AI-powered multi-line autocomplete plugin designed with maximum privacy and security in mind. It runs entirely on a local LLM and is developed to work on consumer-grade PCs.
Cotab focuses on autocomplete that takes the editing intent into account. In addition to the context of the entire file, it considers external symbols, errors, and prior edits, using AI to generate multi-line code suggestions presented as autocomplete.
Setup can be completed with a single click and is ready to use immediately. Also, you can switch the model to use with a single click, and Qwen3-Coder-30B-A3B can run on VRAM 4GB or 8GB environments.
Getting started | Questions / ideas / feedback (English / 日本語OK)
Autocomplete
The programming languages supported depend on the AI model. Despite its compact size, the default model Qwen3-4B-Instruct-2507 supports many languages.
Qwen3-Coder-30B-A3B supports even more languages and, like magic, suggests exactly the code you're about to type.
A dedicated mode that adds code comments. It analyzes code more deeply than the normal mode, understands the intent of the algorithm, and automatically adds detailed comments.
Qwen3-4B-Instruct-2507 already delivers good comments, but this use case can tolerate a slowdown, so we recommend Qwen3-Coder-30B-A3B for the best results.
Auto Translate Mode
A translation-only mode. It can translate not only code comments but also regular text files.
Qwen3-4B-Instruct-2507 also delivers high-quality translations, but in this use case, it is recommended to use Qwen3-Coder-30B-A3B for the best results, as performance degradation is tolerable.
Feature
- Prioritizes privacy, operates completely offline using local LLM
- Provides functionality focused solely on inline suggestions
- Suggests not only inline completion from cursor position, but also multi-line edits
- Provides suggestions considering the entire content of target files, symbols from other files, and edit history
- Offers fast response optimized for llama-server
- There are also modes for Auto Comment and Auto Translate.
- Open source ensuring transparency
Getting started
Install Cotab via the VS Code marketplace
Click "Install Local Server" button or configure your api.
Note:
- It may take a while the first time because it downloads a 2.5GB model.
- The server starts automatically after installation.
- Installation Supported Platforms Windows/MacOS/Ubuntu
Start typing!
| Command |
Keybinding |
| Accept All |
Tab |
| Accept First line |
Shift + Tab |
| Reject |
Esc |
Note:
- By rejecting, you can change the next completion candidates.
- Italic display overlay means the AI is still outputting the result and the result has not been determined yet. In most cases, it is the same as the final result, but in the case of italic display, there may be problems with the merge result.
Important Notes
Requests generally involve prompts exceeding 10,000 tokens.
Optimized for llama-server; we strongly recommend using llama-server
Be especially careful when using pay-per-use API servers, as token consumption can be rapid
When using a local server, we strongly recommend single-user usage
A local server is optimized for single-user.
Concurrent use by multiple users will significantly penalize inference and severely degrade response speed
Tips for use
Comment First
The default model (Qwen3-4B-Instruct-2507) is compact yet highly capable, though it is not specifically designed for code completion. Unlike many recent cloud services, it may not immediately suggest the exact code you want to write. In such cases, writing a comment that describes the code you want first will help the model generate more precise code suggestions based on your description.
Edit Prompt
While model quality matters, completion accuracy varies greatly depending on the prompt content. By customizing the prompt, you may be able to further improve accuracy.
Also, you can create your own custom mode.
To edit the prompt, open it from the menu. The default prompt is commented out. Uncomment it, edit and save, and changes will be immediately reflected in completions.
Recommend: GeForce RTX 3000 series or later GPU (or equivalent) for optimal performance.
Cotab is optimized for llama-server and Qwen3-4B-Instruct-2507 and is designed for high-speed operation. In practical environments with source code over 1,000 lines containing hundreds of external reference symbols, the prompt exceeds 15,000 tokens. Even in such situations, it understands the entire context and shows completions from the second request onward in about 0.5 seconds on a GeForce RTX 4070.
AI processing shows significant performance improvements with the GeForce RTX 3000 series and later. For comfortable response, we recommend a GeForce RTX 3000 series or later GPU or equivalent.
Details
llama-server
You can also use OpenAI compatible APIs, but we strongly recommend using llama-server. llama-server has low overhead and operates at high speed among servers using llama.cpp as backend. See details
Prompt Optimization
llama-server has a mechanism enabled by default that caches prompts from previous requests. Prompt cache is effective up to the part that matches the previous prompt, allowing prompt processing to be skipped up to that part.
To make the most of this mechanism, the original source code in prompt remains unchanged as users type. Instead, a minimal block of modified surrounding code is appended to the bottom of the prompt.
Prompt is fully customizable, and you can switch between prepared modes with a single click.
This allows you to perform completions with the optimal prompt for each purpose.
Edit History
Remembers the user's immediate previous edits and utilizes them in suggestions. Edits are categorized into add, delete, edit, rename, and copy to improve prediction accuracy.
This makes functions created immediately before more likely to be suggested, more accurately reflecting user intent.
Symbols from Other Files
Uses symbols obtainable from VSCode's language providers and utilizes them in suggestions. These symbols allow LLM to understand class structures and improve the accuracy of member function suggestions.
Note: Symbols are input in the order of files viewed in VS Code.
Error Problems
Uses diagnostic errors as input to generate code to fix errors.
Even with a small AI model, it learns to correct errors, so the quality of proposals is further improved.
Code Summary
By summarizing the source code in advance and incorporating the results into the prompt, we enable a deeper level of understanding.
This feature is disabled by default. Because the quality of completion is guaranteed even without summary, as the entire code is input.
Progress Icon Description
| Icon |
Description |
 |
Analyzing source code |
 |
Completing current line |
 |
Completing after current line |
About Available Models
All text generation models are available, but powerful instruction-following performance is required for local code generation. (Instruction-following performance refers to the ability to strictly follow prompt rules and avoid rule violations.)
Qwen3-Coder-30B-A3B
80% of the model is allocated to code learning, providing high-quality completion. Since the actual computation is equivalent to 3B, it operates at high speed like small models. This model can adjust VRAM usage without significantly degrading performance, so Cotab provides presets that work in 4GB or 8GB environments.
Qwen3-4B-Instruct-2507
Despite its very small size of 4B, it has outstanding instruction-following performance and high performance in fields such as mathematics. In Cotab, it provides good completion that makes you forget its small size.
Ministral-3-3B-Instruct-2512
Despite its very small size of 3B, it has high performance and operates at high speed with VRAM 5GB usage. Please try it if you have VRAM limitations.
granite-4.0-micro
It is not recommended because it often generates broken completions in Cotab.
LFM2-2.6B
It is designed to operate at more than twice the speed of Qwen3-4B-Instruct-2507 with only VRAM 3GB usage.
It is not recommended because cases where code breaks have been observed in Cotab's code completion.
However, it can be used when VRAM requirements are very strict or for translation purposes.
Using Remote Servers
You can use OpenAI compatible API servers, but for performance reasons, we strongly recommend using llama-server or llama-swap.
In particular, using llama-server through llama-swap allows automatic model switching when using other chat plugins.
Most Important
When using llama-server, always specify the "-np 1" option.
In the late 2025 update of llama-server, it was changed to run with 4 parallel processes by default. With the default llama-server, since Cotab frequently repeats requests and cancellations at high speed, they are mistaken for completely different requests, causing the prompt cache to not function and resulting in significant performance degradation.
Also specify the "-b 512" option.
For common NVIDIA gaming GPUs like RTX 4070, performance hardly changes even when exceeding 512. Since llama-server's cancel requests are not accepted during batch processing, with the default 2048, it may take several seconds until cancellation is executed, causing unexpected response degradation.
Privacy and Telemetry
Cotab only communicates with the default endpoint "http://localhost:8080/v1" or the LLM API specified by the user. No other external services or servers are contacted. This ensures maximum privacy and security.
- Communication is only made with the configured API
- No telemetry or usage data is ever sent
- User code or input is never shared with third parties
- No personal information is collected or stored
- This project is open-source, and all source code is available on GitHub
With this policy, you can use Cotab with complete confidence.
Note: If you install local server, it accesses the llama.cpp github repository.
💬 Questions, ideas, and usage discussions are welcome in
GitHub Discussions
(English / 日本語OK)
🐞 If you found a bug, please open an Issue instead.
Development / Contributions
- Contributions (issues, PRs, improvement proposals) are welcome.
- Bug fixes, optimizations, and sharing benchmark results are also welcome.
How to build
Setup requirements
Please install VS Code in advance.
Windows
Run this single command to automatically download and execute the setup script. Nothing is required including Git or Node.js - all portable versions are automatically downloaded and set up in ./workspace, and the project will be cloned and VS Code will launch:
mkdir cotab
cd cotab
powershell -NoProfile -Command "$f='run-vscode.bat'; (New-Object Net.WebClient).DownloadString('https://github.com/cotab-org/cotab/raw/refs/heads/main/run-vscode.bat') -replace \"`r?`n\",\"`r`n\" | Set-Content $f -Encoding ASCII; cmd /c $f"
Press F5 in vscode to start debugging the plugin.
Ubuntu
Requires Node.js(v22).
e.g., Install Node.js v22 via package manager.
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt install -y nodejs
Cotab clone & configure.
git clone https://github.com/cotab-org/cotab.git
cd cotab
npm install
code .\
Press F5 in vscode to start debugging the plugin.
MacOS
Requires Node.js(v22).
e.g., Install Node.js v22 on macos.
# install nvm
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash
# activate nvm
\. "$HOME/.nvm/nvm.sh"
# install node.js v22
nvm install 22
node -v
Cotab clone & configure.
git clone https://github.com/cotab-org/cotab.git
cd cotab
npm install
code .\
Press F5 in vscode to start debugging the plugin.
Create Package
npx vsce package
FAQ
Why does the window flicker briefly when starting to use Cotab?
The brief window flicker occurs because Cotab calculates font size during initialization. VS Code doesn't provide a direct API to get character size, so Cotab uses a Webview to calculate the font size. This causes the brief flicker when starting to use Cotab.
What do 4B and 30B in model names mean?
They represent the number of parameters. B stands for "Billion" (10 billion). For example, 4B means 4 billion parameters, and 30B means 30 billion parameters. Generally, more parameters improve model performance, but also increase memory and computational resource requirements.
What does A3B in model names mean?
It indicates that the actual computational load is equivalent to 3B (3 billion parameters). A stands for "Active". For example, Qwen3-Coder-30B-A3B is a 30 billion parameter model, but the actual computational load during inference is optimized to be equivalent to 3B. This allows the model to maintain the high-quality performance of a 30B model while achieving inference speed comparable to a 3B model.
What do Q4 and Q2 in model names mean?
They represent quantization bit counts. Quantization is a technique that reduces file size and memory usage by lowering model precision. Q4 means 4-bit quantization, and Q2 means 2-bit quantization. Lower numbers result in smaller file sizes and less memory usage, but also reduce model quality. Generally, Q4 models are said to have the best balance between data size and quality.
How much quality degradation occurs with Qwen3-Coder-30B-A3B:Q2?
With Qwen3-Coder-30B-A3B:Q2, significant degradation such as frequent syntax errors does not occur. Given its advantages in performance and accuracy in low VRAM environments, we recommend trying it out in practice.
Can't the completion extend beyond the view or avoid overlaying existing code?
VS Code's public API does not allow displaying content that extends beyond the editor view, like Cursor does. It also doesn't allow adding arbitrary blocks between lines. GitHub Copilot uses unpublished internal features to achieve this, so regular plugins cannot implement UX equivalent to GitHub Copilot.
What is a token?
A token is the basic unit that LLMs use to process text. For example, "Hello" is approximately 1 token. For code, the number of tokens is roughly one-third of the number of characters.
License
Copyright (c) 2025-2026 cotab
Apache License 2.0
