Autonomous eXecution & Orchestration Network
AXON is a high-performance, single-binary autonomous AI software engineering agent and runtime built in Go.
Designed to operate directly in your terminal, AXON goes beyond simple "LLM + wrapper" setups to provide a stateful runtime for planning, tool execution, session persistence, and self-correcting development loops. Think of it as a blazing-fast, lightweight, and model-agnostic alternative to tools like Claude Code, Aider, and Mentat.
Here is how AXON works in action:
$ export LLM_END_PROT="https://api.openai.com/v1"
$ export LLM_TOKEN="your-api-key"
$ axon --model gpt-4o
[session] new session 20260824_220130
[session] file: /home/user/.axon/sessions/20260824_220130.jsonl
> Find the bug in the auth validation and fix it.
🤖 Thinking...
🔧 Calling tool [read] {"path": "auth/validate.go"}
📝 Tool Result: (Read 1,420 bytes)
🤖 Thinking...
🔧 Calling tool [edit] {
"path": "auth/validate.go",
"old_text": "if token == \"\" {\n\treturn true\n}",
"new_text": "if token == \"\" {\n\treturn false\n}"
}
📝 Tool Result: Successfully edited auth/validate.go
🤖 Thinking...
🔧 Calling tool [bash] {"command": "go test ./auth/..."}
📝 Tool Result: PASS ok auth/validation [0.12s]
🤖 Done! I found the bug where empty tokens were mistakenly validated as true, corrected the check, and successfully verified it with the unit tests.
- 🏎️ Single-Binary & Blazing Fast: Built in Go with zero external runtime dependencies (no Node.js, no Python startup lags).
- 🧩 Model Agnostic: Works with any OpenAI-compatible API endpoint (OpenAI, Anthropic via proxy, Gemini, DeepSeek, Ollama, vLLM).
- 🛠️ Surgical File Edits: Uses a precise
edit(find-and-replace) tool rather than rewriting entire files, significantly reducing token consumption and preventing hallucinations. - 💾 Session Persistence & Resume: Automatically records your interaction history to structured JSONL files. Resume any previous task with
--continue. - 🛡️ Interactive Safety & Signal Handling: Gracefully handles interrupts (
Ctrl+C) and offers manual control before executing long-running tasks.
Ensure you have Go (1.21+) installed, then run:
git clone https://github.com/zane0x/AXON.git
cd AXON
go build -o axon main.goSet up your API endpoints. AXON supports any OpenAI-compatible base URL.
# Example: Using OpenAI API
export LLM_END_PROT="https://api.openai.com/v1"
export LLM_TOKEN="sk-proj-..."
export LLM_MODEL="gpt-4o" # Optional, defaults to gemini-3-flash-agent
# Example: Using DeepSeek API
export LLM_END_PROT="https://api.deepseek.com/v1"
export LLM_TOKEN="sk-..."
export LLM_MODEL="deepseek-coder"Start a new interactive session in your current working directory:
./axonAXON exposes a set of rich, structured primitives that let the AI inspect and modify files safely:
| Tool | Action | Description |
|---|---|---|
read |
Read File | Reads the full text content of a file (safe/read-only). |
write |
Write File | Writes a brand new file or completely replaces the contents of an existing one. |
edit |
Surgical Replace | Replaces the first verbatim match of old_text with new_text. Safe and token-efficient. |
bash |
Subprocess Bash | Runs terminal commands (e.g. go test, npm run build, git status). |
AXON provides advanced CLI flags to manage your sessions:
Usage: axon [options]
Options:
--continue, -c [path] resume the most recent session (or a specific JSONL file)
--model, -m <model> specify the model to use (defaults to LLM_MODEL env var or "gemini-3-flash-agent")
--list, -l list all saved sessions
--help, -h show this help
Environment variables:
LLM_TOKEN API key for the LLM provider
LLM_END_PROT Base URL of the LLM API endpoint
LLM_MODEL Default model name to use
To resume your last session and continue where you left off:
./axon --continueOr resume a specific session file:
./axon --continue ~/.axon/sessions/2026-08-24T22-00-00.jsonlAt its core, AXON is an execution and orchestration layer connecting models, tools, context, and the environment.
User Intent
│
▼
┌─────────────────┐
│ Agent │
│ │
│ Reasoning │
│ Planning │
│ Decision Making │
└────────┬────────┘
│
▼
┌───────────────────┐
│ AXON Runtime │
│ │
│ Execution │
│ Orchestration │
│ Context │
│ State │
│ Lifecycle │
└─────────┬─────────┘
│
┌────────────────┼────────────────┐
│ │ │
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌──────────┐
│ Tools │ │ MCP │ │ Subagent │
└─────────┘ └─────────┘ └──────────┘
│ │ │
└────────────────┼────────────────┘
▼
Environment
│
▼
Observation
│
└──────────────► Agent
- Runtime over Prompt: Reliable agents require explicit state machine and lifecycle hooks inside the engine, not just complex prompt engineering.
- Model Agnostic: Keep the execution loop strictly decoupled from LLM providers.
- Tool First: Treat capabilities as first-class, typed Go structures instead of simple JSON fragments.
- Observable: Every tool call, event, and state transition emits structured trace information.
- Interruptible: Allow developer intervention, pauses, and restarts at any phase of the execution loop.
🚧 Early Development: AXON is actively evolving. APIs and file structures are subject to change.
- Basic REPL & agent execution loops.
- Session persistence to JSONL.
- Multi-turn session resumption (
--continue). - Precise
editfind-and-replace tool. - Dynamic model switching (
--model/LLM_MODEL).
- Model Context Protocol (MCP) support for custom tool integration.
- Parallel tool execution & sub-agent orchestration.
- Streaming response support.
- Terminal UI (TUI) for interactive task tracing.
Contributions are highly welcome! Whether it is fixing a bug, suggesting a feature, or writing a custom tool:
- Fork the repository.
- Create your feature branch (
git checkout -b feature/amazing-feature). - Commit your changes (
git commit -m 'feat: add amazing capability'). - Push to the branch (
git push origin feature/amazing-feature). - Open a Pull Request.
AXON is open-source software licensed under the MIT License. (See LICENSE for details).