> ## Documentation Index
> Fetch the complete documentation index at: https://docs.fermata.run/llms.txt
> Use this file to discover all available pages before exploring further.

# Fermata and Claude Code

> You already use the CLI. What stays the same when Fermata runs it, what Fermata adds on top, and what it costs.

Fermata does not replace Claude Code. It runs it. Every agent you see in the app is the same `claude` binary you run in your terminal, launched as a real subprocess in your repository, reading the same configuration it reads today.

That is the whole design. If you already have a Claude Code setup you trust, you keep it, and Fermata builds the factory around it.

## What stays the same

**The binary.** Fermata checks a path you configured first, then your `PATH`, then the usual install locations: `~/.local/bin`, `/usr/local/bin`, `/opt/homebrew/bin`, and a few more. Nothing is vendored or bundled.

**Your login and your billing.** Fermata does not hold an account of its own. The CLI authenticates as it does today. Anthropic meters what the agents use.

**Your project instructions.** Agents run with the repository as their working directory, so the `CLAUDE.md` files the CLI would read there are the ones it reads here. Fermata can add standing per-project context on top of them; see [project instructions](/configuration/project-instructions).

**Your slash commands and skills.** Typing `/` in a prompt field opens a palette of the commands and skills from your project directory, your user directory, and any plugins enabled for that project. The command is forwarded to the CLI as plain text, so it behaves exactly as it does in the terminal.

**Your MCP servers, subagents, and settings.** The CLI reads its own configuration directory the way it always does, and Fermata passes your environment through to each subprocess rather than rebuilding it. If you keep more than one Claude configuration directory, you can say which one a project's sessions use, globally or per project, in Settings.

**The conversation.** A session is a live CLI process, not a transcript viewer. Send follow-ups, interrupt, attach files, redirect. It behaves like the CLI session it is, because that is what it is.

## What Fermata adds

**The piece.** One spec and one pull request, carried through four phases with documents and a paper trail: a guided interview, a plan you approve, a team of coding agents, and a diff you read. Start with [your first piece](/start/first-piece).

**Its own worktree per piece.** Agents commit inside a git worktree and branch belonging to the piece, never in your checkout. Nothing lands in your working copy until you merge.

**The Loop board.** Six columns from Backlog to Done, one card per piece, with **Play Loop** and **Pause Loop** over the whole board. Stack up work and let it run. See [the board](/loop/board).

**A leash you set per piece.** How far a piece advances without you, and what its agents may do without asking, are two separate settings with two different defaults. Both live on [the two layers](/control/two-layers).

**The Canvas.** Every running agent as a node you can watch and open. See [the Canvas and standalone sessions](/surfaces/canvas-and-sessions).

**Local analytics.** Fermata parses the Claude transcripts already on your disk, including nested subagents, and reports tokens, model mix, tool usage, and how much ran without you, per project; nothing is uploaded to produce them. See [cost and usage](/reference/cost-and-usage), [analytics](/reference/analytics), and [Home](/surfaces/home) for the plan-limit tiles.

**Your phone.** A paired phone approves gates and tool calls, runs the spec interview, chats with sessions, and lets you merge the pull request without going back to the Mac. See [the phone companion](/mobile/companion).

## What it costs

Fermata itself is free; running agents uses your own Claude, and Anthropic meters that usage. Everything a piece runs bills, planning included, not just the agents, and the figure on its "Usage" card is a list-price estimate. See [cost and usage](/reference/cost-and-usage) for the full breakdown and [privacy and data](/reference/privacy) for what stays on your Mac.

## Working offline

Nothing runs. Every phase is a Claude Code process, and the CLI needs its provider: Anthropic, Amazon Bedrock, or Google Vertex. Without a network you can still read specs, strategies, diffs, and transcripts already on disk.

## Providers

Claude via Anthropic, Amazon Bedrock, and Google Vertex all work, detected from your environment, and the four phases behave the same whichever you run. See [models and providers](/configuration/models).

## What it is not

Fermata is not an IDE and not a rewrite of the CLI. When you need to hand-edit a file, use your editor. It opens the pull request and stops; you review and you merge.