# Start a session

Start a chat or terminal session in the harness of your choice, with your repo, your org's kit and its context.

A session is one harness working on one folder: Claude Code, Codex, Cursor or another, on your own login. EasyCode
starts it, gives it your org's kit and context, and keeps its record. This page walks through the composer you start
one from.

> **Where sessions run**
>
> Sessions run on your machine, so you start them in the [desktop app](/setup/desktop-app) (or the local app). The web
> app shows your team's sessions, and its **Open in desktop** button takes you to the desktop app.

## Before you start

- At least one harness is installed and signed in. A harness that isn't shows greyed out in the picker, with "Not set
  up here · open the setup steps". See [Harnesses on your own login](/setup/harnesses).
- For org work, your org's space is open and the repo is a folder in it. See [Products and
  repos](/setup/products-and-repos).

## Start one

The composer is on **Home** and on the **Sessions** page. Both are the same.

### Say what you want

Type the task in the box: "Describe the task, or type / for a command or skill". Type `/` to pick one of the org's
skills or the harness's own commands. Drop files on the box, or use the attach button, to send them with the first
message.

The **Quick starts** above the box fill it for common jobs: **Create a skill**, **Create an agent**, **Make an
artifact**, **Ask about the codebase** (read-only) and **Review a PR**.

### Choose Chat or Terminal

- **Chat**: a conversation with the assistant, rendered in EasyCode, with approvals as **Allow** and **Deny**.
- **Terminal**: the assistant's own interactive terminal. The first message is optional.

### Pick the harness and mode

Open **Harness** in the options tray under the box. If you have more than one account for that harness, **Account**
lets you choose which one runs the session.

**Mode** sets what the harness may do without asking. The choices are the harness's own:

| Harness             | Modes                                                                      |
| ------------------- | -------------------------------------------------------------------------- |
| Claude Code         | Plan, Ask before acting, Accept edits, Auto, Don't ask, Bypass permissions |
| Codex               | Read-only, Workspace, Full access                                          |
| Cursor              | Plan, Ask, Agent, Force                                                    |
| OpenCode            | Plan agent, Build agent                                                    |
| Gemini, Copilot, pi | Agent default: the agent reports its modes once the session starts         |

A mode with a red shield, such as **Bypass permissions** or **Full access**, skips the checks. Use it only in a copy
you can throw away.

### Pick the folder and product

- **Folder** (when the space has more than one) and **Working directory** choose where the harness starts.
- **Product**, in an org space, says which product the work is for. It starts on the folder's default; pick **No
  product** to leave it out.
- **From a ticket** starts the session on a Jira or Linear ticket, as quoted data, on its own branch. It shows when
  your org has [Jira and Linear](/setup/jira-and-linear) set up.
- **More** holds **Model**, **Reasoning effort** and **Task type**. Each is the harness's default until you set it.

### Start

Press Enter, or the arrow button. Shift+Enter adds a new line. The session opens in **Sessions**.

## When someone else is working in the folder

If another live session is in the same checkout, or someone edited it recently, the composer offers **Separate copy**. That
gives the session its own git worktree at the current commit, so it can't overwrite edits being made there, and you
bring its changes back later. **Keep this checkout** dismisses the warning for that checkout until midnight. A session
in its own copy says "Runs isolated, in its own copy of" the repo.

## What the session gets

In an org space, a session starts with the org's context, built for its harness. Nothing is written into your repo or
your harness's home folder.

- **The org's kit.** The org's instructions, skills, agents and rules, then the product's, at one published version
  of the library. Each harness gets them the way it reads them; see [Bring your existing kit](/guides/bring-your-kit).
- **The product map.** The product's repos, their roles and where they sit on your machine.
- **The org's code.** Read-only checkouts of the org's repos and the org map beside them. See [Set up knowledge for
  a repo](/guides/set-up-knowledge).
- **EasyCode's tools.** MCP tools to read the session's own history, work with tasks and documents, propose library
  changes, and drive the built-in browser. The full list is in [MCP tools](/reference/mcp-tools).

The chip under the composer, and in the session, shows what it launched with: the space and its library version, the
product, the org map's form and "this repo's" own files, such as `AGENTS.md`. "org context limited" means the harness
refused a channel, and the session ran with that part as pointers in its first prompt. If some skills weren't offered
for this product, the chip counts them and links to the library.

A **Personal** session gets no org context: it is just the harness on your folder.

## Next

- [Pick up a teammate's session](/guides/pick-up-a-session), or continue yours in another harness.
- [Workflows, gates and the board](/guides/workflows-and-the-board) to start a session from a task.
