# outrig-cli
`outrig-cli` provides the `outrig` command: it runs an LLM agent on your host and connects it to
MCP servers running inside a [podman](https://podman.io/)-managed container, so tools like
filesystem and shell access stay inside the sandbox you define. The agent layer is the
[Rig](https://github.com/0xPlaygrounds/rig) crate; outrig's job is to run Rig in the host process
and give it a container-isolated MCP tool set.
To embed the container/MCP machinery in your own Rust program without the LLM dependency graph,
depend on the [`outrig`](https://crates.io/crates/outrig) library crate instead.
## Install
```sh
$ cargo install outrig-cli
```
This installs the `outrig` binary.
## At a glance
You define the sandbox in your repository: a `Dockerfile` for the container and an
`.agents/outrig/config.toml` describing which MCP servers run inside it and which LLM the agent
talks to. Then you run `outrig run` and talk to the agent over a stdin/stdout REPL:
```sh
$ outrig run
[outrig] agent: coding (model: fast / provider: openai / gpt-4o-mini)
[outrig] container outrig-20260502T103412-3f2a started
[outrig] mcp servers: fs, shell
> what's in this repo?
This repo is a Rust project named "outrig". The top level contains Cargo.toml,
src/, and doc/. src/ has lib.rs and main.rs...
> ^D
[outrig] session 20260502T103412-3f2a ended
```
## Commands
```sh
outrig init # set up global config and scaffold an image config
outrig config init # configure global settings interactively
outrig run # start an interactive agent session
outrig mcp # serve the configured MCP servers as one MCP over stdio
outrig design # generate AI-assisted design prompts
outrig build # build (or cache-hit) the session image
outrig image add # scaffold a new image config in the repo
outrig image init|build # create / build a standalone image project
outrig image inspect # read an image's OCI labels without starting it
outrig ls|logs|discard|clean # inspect and prune sessions
```
## Local models (optional)
The `local-llm` feature adds an in-process [mistral.rs](https://github.com/EricLBuehler/mistral.rs)
backend so the agent can run a model without an external API; `cuda` and `metal` features enable
GPU acceleration.
```sh
$ cargo install outrig-cli --features local-llm
```
## Documentation
Full docs render as an mdbook at <https://tgockel.github.io/outrig/>. Source lives in the
[repository](https://github.com/tgockel/outrig).
## License
Licensed under the Apache License, Version 2.0.