A lightweight CLI tool for managing and running single-responsibility LLM agents via declarative TOML configuration, featuring sub-agent delegation, stdin piping, persistent memory, and MCP tool support.
Axe is a lightweight CLI tool written in Go, designed for running single-responsibility LLM agents in terminal environments. Each agent is declared as an independent TOML file covering model selection, system prompts, tool sets, sub-agent delegation, memory policies, and token budgets — all suitable for version control.
Core capabilities include: multi-LLM provider support (Anthropic, OpenAI, Ollama, AWS Bedrock); sub-agent delegation with parallel execution (up to 5 levels deep, max 50 tool call rounds); timestamped Markdown persistent memory with LLM-assisted pattern analysis and garbage collection; reusable Skill instruction sets (SKILL.md); file operation sandboxing and SSRF protection (allowed_hosts whitelist, private/reserved IPs always blocked); and MCP server connectivity via SSE or streamable-HTTP for extended tooling.
Axe follows a single-execution model, natively fitting Unix pipes (e.g., git diff | axe run code-reviewer), and provides JSON structured output with semantic exit codes (0=success, 1=runtime error, 2=config error, 3=provider/network error, 4=token budget exceeded) for integration into Git hooks, cron jobs, and CI/CD pipelines. All LLM calls use Go's standard net/http library with zero third-party SDK dependencies — only 4 direct dependencies total (cobra, toml, mcp-go-sdk, x/net). Available as precompiled binaries (Linux/macOS/Windows) and Docker images with hardened defaults (non-root user, read-only root filesystem, all capabilities dropped).
Agent discovery follows priority: --agents-dir → <cwd>/axe/agents/ → $XDG_CONFIG_HOME/axe/agents/. Config directory structure: $XDG_CONFIG_HOME/axe/ containing config.toml, agents/, skills/. Key CLI commands: axe config init, axe agents init <name>, axe agents edit <name>, axe run <agent>. User message priority: -p flag → stdin pipe → built-in default message.
Unconfirmed: no Hugging Face model/dataset page found; no academic papers found; OpenCode provider API spec not detailed; SKILL.md community format spec link not provided; v1.10.0 exact release date not shown; pkg/runner/ public API stability not versioned or documented.