Code Web Chat

AI coding without agents. Guide the model to production-grade code with engineer-level control over context.

Works with VS Code, and forks like Cursor, Windsurf, and VSCodium.

✅ Send prompts anywhere

Chatbots—ChatGPT, Claude, Gemini, AI Studio, Qwen, etc.
Model providers—Gemini API, OpenRouter, local Ollama, etc.

✅ Apply responses—sophisticated markdown parser for multi-file changes integration
✅ Fully featured—code completions, commit messages, checkpoints, and more
❤️ Free and open-source—released under the GPL-3.0 license
🔒️ Total privacy—operates 100% on your local machine
👥 Community driven—join our Discord server and get involved!

Walkthrough

Introduction

👨‍⚖️ Respect to chatbots' Terms of Use

Code Web Chat helps you use your favorite coding web tools like ChatGPT's canvas. The idea to initialize chatbots is borrowed from Firefox and because there is no further automation once the prompt is sent, by using CWC you're not violating their Terms of Use. Contributors should not submit pull requests implementing further chat automations of any kind, as these will be kindly rejected.

🧐 The limitations of LLMs

Large language models (LLMs) are trained on vast datasets targeting many use cases. For code generation, a model's training involves analyzing millions of simulated problem-solving flows, such as arriving at the accepted answer from a given StackOverflow question. For the purpose of agentic coding, models are trained on an additional layer of data that simulates gathering context and planning its next steps.

Because the model is only as smart as examples it has seen in its pre-training stage, the possible coverage of real-world problems when approached at a high level is fundamentally limited.

Therefore, CWC is designed to align with LLMs' true capabilities—that is, code generation in a controlled signal-to-noise ratio environment. Controlled by you, the engineer.

🧠 Guide the model with context

Unlike coding agents that require detailed instructions to understand your intent and locate relevant files, with CWC you provide fine-grained context up front, allowing simple, even vague instructions.

[!TIP] LLMs are pattern matchers—they love examples! Include some you believe will help the model understand the goal better.

Meet the CWC's non-agentic workflow—select folders and files, enter instructions, and send message via...

new chat—to continue in the connected browser
API call—to use a model provider of choice

Constructed message looks like the following example:

Implement a subtract function.
<system>
Whenever proposing a new or updated file use the Markdown Code Block syntax. Each code block should be a diff patch. Don't use XML for files.
</system>
<files>
<file path="src/calculator.ts">
<![CDATA[
export const addNumbers = (a: number, b: number) => a + b;
]]>
</file>
</files>
Implement a subtract function.
<system>
Whenever proposing a new or updated file use the Markdown Code Block syntax. Each code block should be a diff patch. Don't use XML for files.
</system>

[!NOTE] The prompt and edit format instructions are repeated after the context for better accuracy.

Once the response is generated, CWC's built-in parser extracts code blocks with suggested edits for one-click multi-file changes integration.

Chatbot initialization

Install the open-source Connector in your browser and never copy & paste again.

Supported chatbots

AI Studio
ChatGPT
Claude
Copilot
DeepSeek
Doubao
Gemini
GitHub Copilot
Grok
HuggingChat
Kimi
Minimax
Mistral
Open WebUI
OpenRouter
Perplexity
Qwen
Together
Yuanbao
Z

[!TIP] With the browser extension you can include markdown-parsed websites in context. Go to target website, click the extension's icon in the browser's toolbar and click Enable for context.

[!IMPORTANT] The Apply response button placed under responses is not a means of automatic output extraction, it's an alias for the original copy to clipboard button. Review the content script for implementation details.

API Tools

Anything CWC can do in chatbots, it can do calling model providers directly from the editor.

[!TIP] Get started with generous free tiers from Google, Mistral or Cerebras.

🛠️ Edit Context
Modify, create or delete files based on natural language instructions.

🛠️ Code Completions
Get accurate code-at-cursor from state-of-the-art reasoning models.

🛠️ Intelligent Update
Handle the compact "truncated" edit format and malformed diffs.

🛠️ Commit Messages
Generate meaningful summaries of changes adhering to your style.

Commands

Code completions

Code Web Chat: Code Completion - Get code-at-cursor using API tool.
Code Web Chat: Code Completion using... - ...with configuration selection.
Code Web Chat: Code Completion with Instructions - ...with instructions.
Code Web Chat: Code Completion with Instructions using... - ...with instructions and configuration selection.

Checkpoints

Code Web Chat: Checkpoints - Restore the overall workspace state to the saved checkpoint.
Code Web Chat: Create New Checkpoint - Save the current state of the workspace.

Context

Code Web Chat: Save Context - Save the currently checked files as a named context for easy reuse.
Code Web Chat: Apply Context - Apply a saved context to either replace or merge with the currently checked files.

Enterprise security

Code Web Chat operates exclusively on your machine. Your code and instructions are sent directly to chatbots via editor-browser communication channel run on local Websockets. For API tools, model providers are called directly.

Community

If you have a question, or want to help others, you're always welcome in our community.

Donations

If you enjoy using Code Web Chat, buying a $3 coffee is a great way to show your support for the project.

BTC: bc1qfzajl0fc4347knr6n5hhuk52ufr4sau04su5te

ETH: 0x532eA8CA70aBfbA6bfE35e6B3b7b301b175Cf86D

XMR: 84whVjApZJtSeRb2eEbZ1pJ7yuBoGoWHGA4JuiFvdXVBXnaRYyQ3S4kTEuzgKjpxyr3nxn1XHt9yWTRqZ3XGfY35L4yDm6R

Contributing

All contributions are welcome. Feel free to submit pull requests, feature requests and bug reports.

Robert Piosik