oneagent (ma)

What it is

A model, available as one more command in your pipeline.

oneagent is a one-shot command-line agent for people who already live in a shell. Instead of a session you attach to, you run it like any other Unix command: ma 'explain this project'. The task comes from its arguments, and extra context can come from stdin (git diff | ma 'summarize these changes'). Both are sent to an OpenAI-compatible Chat Completions endpoint, and the model then runs shell commands with whatever host CLIs are already installed — git, cargo, grep, curl. oneagent takes the directory you started it in as its workspace, writes the final answer to stdout and the process log to stderr, so it composes with the rest of your pipeline. Common write operations and shell network commands are refused by default; --write and --net opt in, and --skills loads named SKILL.md directories into the prompt. No conversation state is kept between runs.

The agent itself is one binary with no runtime dependency on Rust and no daemon to start. It is not a replacement for a resident coding agent such as Claude Code or aider; it is what you reach for when you want a model available as a single command inside an existing script, a Makefile, or a CI step.

How it differs

Not a session. A command.

Resident coding agents sit in your terminal and hold a conversation. oneagent runs once. That single difference decides which one belongs in a pipeline.

Install

One binary. Pick whichever path you already have.

tar -xzf ma-<version>-aarch64-apple-darwin.tar.gz
mv ma-<version>-aarch64-apple-darwin/ma ~/.local/bin/

cargo install oneagent

Then configure a model

Any endpoint that speaks the OpenAI-compatible Chat Completions API and supports tool calling will do — hosted or local.

export BASE_URL='https://your-endpoint.example/v1'
export MODEL='your-model'
export API_KEY='your-key'

BASE_URL is the API root; /chat/completions is appended. Settings can also live in a config file.

Usage

Point it at a directory and ask.

oneagent uses the directory you started it in as its workspace. -v shows model text and command execution live; the final answer goes to stdout and the process log to stderr.

cd /path/to/your/project
ma -v 'explain the structure of this project'

# Take supplementary context from stdin — it never lands on disk
git diff | ma 'summarize these changes'

# Allow writes inside the current directory and its descendants
ma -v --write 'fix the config file'

# Also allow the shell to use the network
ma -v --write --net 'upgrade the project dependencies'

# Load skills by name
ma --skills code-review,explain 'review and explain these changes'

Writes and network are opt-in

By default common write operations and shell network commands are refused; model API requests are always allowed. The built-in Guard is a best-effort command check, not a security sandbox — use a container or virtual machine when you need real isolation.

When the context gets too long

Piped input exists only in the conversation and never lands on disk. When the context overflows, the program recovers in order of increasing cost: compact old turns first, then truncate the newest results, and only then truncate the original input. That last step always warns on stderr, because the omitted part can never be read back.

Full usage reference →

FAQ

Questions people actually ask.