A place for your
next good idea.

Download the desktop app for macOS, Windows, or Linux. Choose the build that matches your machine, run the installer, and open the app.

Current release artifacts retain the Axle filename. The download page links to those builds.

1. 1

   Connect your model

   Open Settings → Models. Add a provider API key and choose an available model.

2. 2

   Add a project

   Choose a local folder with code you want to work on. You can also start a conversation without a project.

3. 3

   Describe an outcome

   Ask a clear question or describe a change. Include any constraints that matter to you.

4. 4

   Read, review, continue

   Follow the response, inspect tool output, and review file changes. Keep the conversation going with a follow-up.

Explain how this project is structured.
Find the main entry points and tell me how to run the tests.
Do not change any files yet.

Projects connect conversations to folders on your machine. Your existing language, framework, package manager, and Git history remain part of the work.

The runtime saves conversation history under .axle/sessions. Keep private credentials out of prompts and source files.

Manage desktop API keys in Settings → Models. Select a provider, add the key from your account, and choose a model you have access to. Hosted model usage is billed by that provider.

A provider that is not on the list

Add any OpenAI-compatible endpoint with its own base URL and API key. Axle matches the host against the providers it knows, sends the request headers that endpoint expects, checks that it answers, and then lists the models your key can reach so you can choose which ones to keep.

Prefer a local model?

Run Ollama on your machine and choose a model you have installed. The runtime uses http://localhost:11434/v1 by default. A localhost endpoint does not require a provider key.

Plan mode lets you investigate a project without making changes. Build mode allows the tools needed to carry out a task. Choose the mode that fits the work in front of you.

When an approval is requested, read the action and choose whether to allow it, allow it for the session, decline it, or cancel the turn. Your choice applies to the action shown.

axle --interaction-mode plan --cwd ./my-project \
  -p "Explain the authentication flow and suggest improvements"

Open Settings → Remote access to configure access to your computer from another device. The desktop app supports connections over your local network or Tailscale.

Keep the host computer running and follow the connection instructions shown in the app. On Windows, project environments can also use WSL when it is installed and configured.

Model Context Protocol lets a server expose tools to the agent. You can connect a remote HTTP endpoint or start a local stdio server. The server determines which tools are available.

Use the installation and authentication instructions supplied by the MCP server. Third-party servers may need a separate account, subscription, or access token.

The runtime has a shortcut for GitHub's hosted MCP server. Set GITHUB_PERSONAL_ACCESS_TOKEN or GITHUB_TOKEN in your environment, then add the --github-mcp flag.

axle --github-mcp --cwd ./my-project \
  -p "Find the open issues relevant to this project"

Grant the token only the repository permissions needed for your task. The available tools and actions follow your GitHub access and the server's capabilities.

The command-line runtime currently ships as axle. The examples below assume the binary is on your path and the provider's API key is configured in your environment.

axle --provider openai --cwd ./my-project \
  -p "Run the tests and explain any failures"

Agent profiles

Specialized runtime agents can be configured in .axle/agents.toml. Profiles support instructions, provider and model overrides, and an allowed tool list.

[[agents]]
name = "code-reviewer"
description = "Read-only code review."
instructions = "Review the requested files. Explain findings with file and line references."
allowed_tools = ["read", "grep", "find", "ls"]