lgalabru/axel

Agent orchestration, designed for Rex.

Axel is a tool that runs and coordinates coding agents (AI assistants that write code) from your terminal, designed specifically for the Rex terminal app. It lets you create multiple specialized agents, give them tasks in "threads," and have them work together—even across different machines.

Key Features

🧵 Threads, not chat tabs

A thread is a single task (an "intent") plus all the work done about it. It survives even if the agent's process, the terminal pane, or your computer shuts down. You can run several threads at once, pause and resume them later, and pick them up from any machine. Every change is recorded as a signed event, so copies on different machines can be merged safely.

🪪 Named agents

Instead of one all-knowing assistant, you create a team of specialists. Each agent has its own name, its own cryptographic key (Ed25519), its own permissions, budget, and inbox, and loads only the skills it needs. For example, you could have terra build code and opus review it with a different model.

🖥️ The terminal, fixed by Rex

Rex fixes limitations of traditional terminals (like not keeping work alive or showing who's waiting). Axel runs as a program inside Rex, so threads stream as normal terminal output, and status appears in Rex's tabs.

🤖 Works with many coding tools

Axel can drive popular coding agents like Claude Code, Codex, OpenCode, Goose, and Cursor. It uses a standard protocol (Agent Client Protocol) to control them directly—no extra Node.js needed. You can specify the model right in the agent name, e.g., claude/opus or codex/gpt-5.5. Other tools can run in their own panes.

🌐 Your other machines

You can add a remote computer (like a Mac mini on your tailnet) as a "host." Agents run there, but their keys never leave your home machine—the host gets a short-lived permission per thread.

📜 One AXEL.md for every harness

A single file (AXEL.md) holds the project's context and the skills agents use. It's referenced by CLAUDE.md and AGENTS.md. Skills can be local folders or links from skills.sh or GitHub, pinned in AXEL.lock.

🛡️ Autopilot, with a guard

Agents work without asking for permission by default, but a guard refuses irreversible commands, and a budget limits what each agent can spend. You can set stricter modes for agents if you want more control.

Installation

curl -fsSL https://axel.md/install | sh

or with Homebrew:

brew install scarce/taps/axel

You'll need Rex and at least one coding agent harness (like Claude Code, Codex, etc.). The first run checks your machine and helps you create your first agent. Updates are released on every commit that passes CI.

How It Works (Briefly)

  • You create agents with axel agent new, specifying a harness and optionally a model.
  • You start a thread with axel thread new, giving it a task and naming which agents to involve.
  • Agents work autonomously, can message each other (e.g., @opus please review the diff), and the thread persists across sessions.
  • Axel uses hooks and the protocol to get status—it never scrapes the screen.

Development

Built in Rust. To build and test:

just install    # release build, linked into ~/.local/bin
cargo test --workspace

Design docs are in docs/DESIGN.md, wire formats in docs/CONTRACTS.md, and the website in web/.

License

Apache-2.0


All information above is taken directly from the project's README.

Related

  • Project
  • Project
  • Project
  • Project