Oven

A coding agent that runs in your terminal. It connects to an LLM, reads and writes files in your project, runs bash commands, and helps you get things done.

Usage

# One-shot: ask a question and get an answer
oven "what does this project do?"

# Interactive mode (opens a TUI)
oven

# With a specific model and API key
OVEN_MODEL=deepseek-v4-flash OVEN_DEEPSEEK_API_KEY=sk-xxx oven

# Resume a previous session
oven --session my-session

# Work in a different directory
oven --root /path/to/project

How it works

Oven runs inside your project directory. It:

1. Takes your prompt and sends it to an LLM provider
2. The LLM can read files, write files, and run shell commands
3. Oven streams the response back in real time

Supported providers

Oven speaks the OpenAI chat completions format, so it works with any provider that supports it:

Provider	Model prefix	Env var for API key
OpenAI	gpt-*, o1-*	OPENAI_API_KEY
DeepSeek	deepseek-*	DEEPSEEK_API_KEY
Zhipu (智谱)	glm-*	OVEN_ZHIPU_API_KEY
Moonshot (Kimi)	kimi-*	MOONSHOT_API_KEY
Anthropic*	claude-*	ANTHROPIC_API_KEY

* Anthropic requires an OpenAI-to-Anthropic proxy (e.g. ANTHROPIC_BASE_URL pointing to one).

Set OVEN_MODEL to pick the model. Override the base URL with OVEN_BASE_URL (or the provider-specific vars like ANTHROPIC_BASE_URL, OPENAI_BASE_URL) if you need a proxy or self-hosted endpoint.

Configuration

Put a .oven.toml in your project root, or configure globally at $XDG_CONFIG_HOME/oven/config.toml (default ~/.config/oven/config.toml). On first run Oven creates the global config as a template if it is missing. Available options:

request_timeout_secs = 60
max_retries = 2
base_backoff_ms = 500

skills = []
tools = ["file_read", "file_write", "bash"]

[provider]
model = "deepseek-v4-flash"
base_url = "https://api.deepseek.com"
api_key = "sk-xxx"

[mcps]
# MCP server declarations (transport support coming)

If neither file exists, Oven uses sensible defaults. Sessions are stored as JSONL under $XDG_DATA_HOME/oven/sessions/ (default ~/.local/share/oven/sessions/).

skills opts into guidance modules that inject instructions into the system prompt; no skills are bundled yet, so entries are accepted and currently skipped. tools selects which capabilities the agent can actually invoke; an empty tools: list mounts the built-in default set (file_read, file_write, bash). The two are independent.

Modes

One-shot — pass a prompt as arguments. Oven replies and exits.

oven "what is love?"

Interactive — run without arguments for a TUI. Type your prompt, press Enter to send. Esc cancels the current request, Ctrl-C quits.

oven

Piped — pipe input in:

echo "summarize this" | oven

Sessions

Use --session <id> to resume an existing session; its history is loaded and continued. A new session always gets an auto-generated id (uuid v7) that the app manages internally, so you never provide an id for a new chat.

oven --session my-project "continue the previous chat"

Sessions are stored as JSONL files in the platform's data directory.

Slash commands

In both one-shot and interactive modes:

Command	What it does
/clear	Clear the screen and start a new chat (switches to a new uuid v7 session; the old file is kept)
/exit	Exit the agent

Building from source

cargo build -r
./target/release/oven

License

MIT