CodeFlux Documentation

CodeFlux is a local agent tool. It runs as a desktop application on your computer, works directly inside your project folders, and uses a large language model to understand your request, plan the work, edit files, run commands and check the results.

The model is entirely your choice. CodeFlux can use a model deployed on your own machine or network, or the API of a hosted model vendor. It is already adapted to the major vendors in China and abroad. To use one, you only fill in the vendor's address, a model ID and your key, and select the protocol format the vendor speaks.

Key concepts

Three steps from download to your first task.

1. Install CodeFlux

   Download the package for your system from the download section and install it. See Installation for platform-specific notes.

2. Add a model

   Open the model settings, add a provider, choose its protocol, then enter the Base URL, Model ID and API key. For a local model, point the address at your local server instead. Details in Connect a model.

3. Open a project and start

   Choose a project folder, then describe what you want in plain language, for example: "Add an endpoint to the orders module that exports Excel, with date filtering, and write unit tests."

CodeFlux is available for Windows, macOS and Linux. All packages are published on the GitHub release page.

Windows

1. Download CodeFlux-win-x64.exe and run it.
2. If Windows shows "Windows protected your PC", click More info, then Run anyway.
3. Follow the installer, then start CodeFlux from the Start menu.

Prefer no installation? Download CodeFlux-windows-x64-portable.zip, extract it to any folder and run the application inside.

macOS

1. Download CodeFlux-mac-arm64.app.tar and extract it (double-click it, or run the command below).
2. Move CodeFlux.app to your Applications folder.
3. Because the app is not signed yet, remove the quarantine flag once, then open it.

tar -xf CodeFlux-mac-arm64.app.tar
xattr -cr CodeFlux.app

Only Apple Silicon (M-series) Macs are supported in this release.

Linux

Choose the package that fits your distribution.

# AppImage (any distribution)
chmod +x CodeFlux-linux-x86_64.AppImage
./CodeFlux-linux-x86_64.AppImage

# Debian / Ubuntu
sudo apt install ./CodeFlux-linux-amd64.deb

# Fedora / RHEL
sudo dnf install ./CodeFlux-linux-x86_64.rpm

Verify your download

Every release includes a SHA256SUMS.txt file. Compare the checksum of your download with the matching line in that file.

# Windows (PowerShell)
Get-FileHash .\CodeFlux-win-x64.exe -Algorithm SHA256

# macOS
shasum -a 256 CodeFlux-mac-arm64.app.tar

# Linux
sha256sum -c SHA256SUMS.txt --ignore-missing

CodeFlux does not ship with a model of its own. Instead, you connect the models you want to use. Each connection is a provider made of a protocol, an address, a key and one or more model IDs. You can add as many providers as you like and mix local and remote ones.

Configuration fields

Nameoptional

A label for your own reference, such as "DeepSeek" or "Office GPU server".

ProtocolAPI format

The API dialect the provider speaks. See Protocol formats.

Base URLAPI address

The provider's API address, as given in its documentation, usually ending in /v1. Do not append the endpoint path such as /chat/completions.

API keysecret

The credential issued by the vendor. Local runtimes usually do not check it; if the field cannot be left empty, enter any placeholder such as local.

Model IDexact name

The identifier of the model at that provider, copied exactly from the vendor's model list, for example deepseek-chat or qwen-plus.

Protocol formats

Vendors expose their models through different API formats. CodeFlux supports several, so a model works as long as you select the format its provider uses.

Remote vendor APIs

1. Get your credentials

   Create an API key in the vendor's console. Make sure the account has quota or balance, and that the model you want is enabled.

2. Add the provider

   In CodeFlux, add a provider, pick the protocol, then paste the Base URL and API key.

3. Add the model ID

   Enter the exact Model ID. A provider can hold several models, for example a fast one for simple edits and a stronger one for planning.

4. Test it

   Start a session with the new model and send a short message. If it answers, the setup is complete.

Local models

To keep everything on your own hardware, run a model with a local runtime and point CodeFlux at it. Any runtime that provides an OpenAI-compatible API works, including Ollama, LM Studio and vLLM.

Example with Ollama:

# 1. Download a model
ollama pull qwen2.5-coder:7b

# 2. Check the server is running and the model is listed
curl http://localhost:11434/v1/models

Then add a provider in CodeFlux with the following values:

Protocol

OpenAI-compatible

Base URL

http://localhost:11434/v1

API key

Any placeholder, for example ollama

Model ID

qwen2.5-coder:7b (as shown by ollama list)

A model served on another machine in your network works the same way: replace localhost with that machine's address and make sure its server listens on the network interface and its firewall allows the port.

Vendor reference

Common addresses to get you started. Model IDs change often, so always check the vendor's current model list.

Vendors not listed here, including gateways, aggregators and your company's internal model service, work the same way: use the address and protocol from their documentation.

Managing and switching

• Add several providers and models, then choose which one each session uses.
• Switch models at any time without leaving your project. Use a strong model for planning and a fast, inexpensive one for routine edits.
• To rotate a key, edit the provider and paste the new one. Existing sessions pick it up automatically.
• Remove providers you no longer use to keep the list tidy.

Projects and sessions

Open a project by choosing its folder. The agent reads and edits files inside that folder only, so it always works against your real code. Each task lives in its own session, shown as a tab. You can run several sessions in parallel and switch between them without them interfering.

Giving tasks

Describe the outcome you want, not the keystrokes. Good requests state the goal, the scope and how you will judge success.

• "Fix the login timeout bug. Start from the auth middleware and add a regression test."
• "Upgrade the dependencies in this project and make sure the build still passes."

For larger work, CodeFlux first analyzes the codebase and drafts a plan you can edit. It starts only after you confirm. It then edits files, runs commands and tests, and repairs its own errors along the way.

Commands and references

Reviewing changes

Every change is shown as a diff, so you can read exactly what was modified, line by line. You can then accept it, roll it back, or commit it to Git. Because each step can be undone, you stay in control even when the agent makes many edits.

MCP

The Model Context Protocol (MCP) lets CodeFlux use external tools and data sources, such as GitHub, databases, Figma, a browser or Docker. Each integration is an MCP server that you add once and the agent can then call when relevant.

1. Find the MCP server you want and note its launch command or its URL.
2. Add it in CodeFlux's MCP settings and supply its command, arguments and any environment variables (such as an access token).
3. Confirm the server's tools appear, then ask the agent to use them.

Most servers document a configuration like this one:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "<your-token>" }
    }
  }
}

Multi-agent mode

For complex tasks, CodeFlux can split the work between specialised roles that advance in parallel: a planner that drafts the approach, a coder that implements it, a tester that writes and runs tests, and a reviewer that checks the result. Each role can use whichever model suits it best, so you can combine a strong planning model with a fast coding model.

CodeFlux is designed so that you decide where your data goes.

Recommendations

• For sensitive code, use a local model or a model service hosted inside your organization.
• Review the data policy of any hosted vendor before sending proprietary code.
• Use dedicated, limited keys and rotate them regularly.
• Keep your project under version control so every change can be reviewed and reverted.

Model connection

Installation and startup

Still stuck? When you contact us, include your CodeFlux version, your operating system, the provider and protocol you are using, and the exact error message. Never include your API key.

• Technical support: [email protected]
• Pricing, Team and Enterprise plans: [email protected]
• Releases and source: GitHub