Quickstart

Install Tachyon, connect a provider, and drive a first change end to end.

1. Install

One line, no runtime to install alongside it:

bash
curl -fsSL https://tachyon.feyncode.com/install.sh | sh

macOS users can use Homebrew (brew install feyncode/tap/tachyon), and Windows has a native PowerShell installer. See Installation for every option, including direct binary downloads.

Confirm it landed:

bash
tachyon doctor

doctor checks your config, keys, MCP servers, and whether a sandbox backend is available on this machine. Run it first whenever something behaves oddly.

2. Connect a provider

Tachyon ships with no credentials. Save a key for any provider in the models.dev catalogue:

bash
tachyon login anthropic

You will be prompted for the key, which is stored in your user config directory rather than in the repository. To point at an OpenAI-compatible endpoint — a proxy, a gateway, or a local server — pass a base URL:

bash
tachyon login myprovider --base-url https://api.example.com/v1

Then pick a default model. Running tachyon model with no argument opens the catalogue browser; passing an id sets it directly:

bash
tachyon model anthropic:claude-sonnet-5

3. Your first session

Change into a repository and launch. Tachyon operates on the directory you start it from:

bash
cd ~/your-project
tachyon

You get an interactive REPL. Describe what you want in plain language rather than naming files and functions — the agent searches the repo itself:

session
› add rate limiting to the /auth endpoints, then run the tests

For a big or risky change, start in plan mode so nothing can be written until you have read the plan:

bash
tachyon --plan

4. Approving work

By default Tachyon runs in ask permission mode: read-only tools run freely, and anything that mutates your filesystem or shells out asks first. Each prompt offers allow once, allow for the session, or deny.

Cycle modes mid-session with Shift+Tab — ask, accept-edits, bypass, auto. Most people settle on accept-edits for routine work and drop back to ask when the agent is somewhere unfamiliar.

5. Undoing it

Every write is snapshotted before it happens, so a bad edit is not a git problem. Inside a session, /undo reverts the last change and /fork branches the conversation so you can try a different approach without losing the current one.

To keep the agent away from your working tree entirely, run it in an isolated git worktree on its own branch:

bash
tachyon --worktree experiment

From here, Providers & models covers switching models per session, and Permissions & sandboxing covers locking things down properly.