Getting started

horon agent is a headless coding agent. One Rust binary, one OpenAI-compatible endpoint, no terminal UI in the headless package. It runs a turn, calls tools, and exits — or stays up as a daemon that SDKs and editors attach to over a Unix socket.

What it is

• One model path. Requests go straight to /chat/completions on the endpoint you configure. There is no router process and no plugin host.
• Two binaries, one engine. horon-agent is headless (serve, run, repl, api-bridge, acp). horon-tui is the separate interactive terminal client, built on the SDK.
• Append-only sessions. History is an event log on disk; archive, compaction and rewind append records and never delete an original.
• ACP v1 and v2. horon acp serves both protocol versions and answers initialize with the highest version both sides support.

First turn

Install the CLI (see Install), then point the agent at an endpoint. Any OpenAI-compatible server works; this example uses a local one with no key.

# ~/.horon/config.toml
[provider]
default_provider = "local"

[providers.local]
type = "openai-compatible"
base_url = "http://127.0.0.1:11434/v1"
auth = "none"
default_model = "my-model"

Check that a profile resolves, then run a turn:

horon provider current --json   # "ready": true when a profile resolves
horon run "summarize this repo"

With no profile configured, a model call stops before the network and names the missing [providers.<name>] table. Nothing silently falls back to another path.

Where to go next

• Configuration — provider profiles, model roles, storage.
• CLI — every subcommand and what it does.
• Tools — the core tool set and how deferred loading works.
• TypeScript SDK and Python SDK — embed the engine in an application.