# Profiles

> Run one CLI as several accounts: a program, arguments and environment.

A profile is an account: the program to run, the arguments to add and the
environment to run it with. It is the shape a lot of people already had as a
shell function — a `claude` with seven variables in front of it — with
somewhere to keep it.

## Making one

In **Settings → Providers**, press **New profile** in the **Installed** list.
For most accounts the form is two fields:

| Field | What it is |
|---|---|
| **Name** | what you call this account, e.g. `GLM` |
| **Command** | what you type in a terminal to start it — an agent such as `claude`, or a function of your own such as `claude-work` |

A function or alias of your own is run once in your interactive shell, with a
stand-in for every agent program, and read back: the agent it starts, the
arguments it adds and the variables it sets. Only a function or an alias is
run this way; a program or script on your `PATH` is kept as a program and runs
when the profile does.

**Details** holds the rest, filled in by that reading or by hand:

| Field | What it is |
|---|---|
| **Agent** | the CLI it is based on |
| **Program** | optional. Left empty, the base CLI's own launch line is used |
| **Arguments** | one line, added to that launch, e.g. `--permission-mode bypassPermissions` |
| **Config directory**, **Endpoint**, **Token**, **Models** | Claude Code only: `CLAUDE_CONFIG_DIR`, `ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`, and the models the chat offers instead of Opus, Sonnet and Haiku |
| **Environment** | name and value pairs, with **Add variable** and **Remove**. Called **Other variables** for Claude Code |

**Save** and it becomes a row of its own in the list. *The program runs with
these variables set and these arguments added.*

## Models

A Claude Code profile, opened in **Settings → Providers**, has three more rows:

- **Models** — each model it offers, with a switch, all on by default. A model
  switched off leaves the model picker in [chat](/docs/chat). `default`, what a
  session runs when none is picked, always stays on. A model of your own is
  still added in the **Models** field above.
- **Advisor model** — none by default. devpit starts every session on this
  profile with it: chats, terminals, board steps and idle compaction. It goes
  in the settings devpit starts the profile with, never in the profile's own
  `settings.json`.
- **Subagent model** — for a profile you made, the model its sessions'
  subagents run on, as `CLAUDE_CODE_SUBAGENT_MODEL`. It is saved with the
  profile.

## Compacting an idle conversation

A long Claude Code conversation left alone outlives its prompt cache, and the
turn after that reads the whole context again, uncached. **Compact idle
conversations**, in a Claude Code profile, runs the CLI's own `/compact` first:

| Option | When |
|---|---|
| **Never** | the default: nothing changes until you pick another |
| **When you return** | on your next message, when the conversation has been idle at least five minutes and its context is over the threshold |
| **Just before expiry** | after four minutes idle, while the cache still holds; a turn that is running makes it wait |

The threshold is 200,000 tokens unless you change it. A compaction leaves a
quiet line in the thread — *Compacted at 212k tokens* — in the window and on
the phone, and its cost goes on the conversation. One that fails leaves the
conversation as it was.

<!-- Screenshot — the profile editor, with an environment variable masked. -->

## Secrets

A variable whose name looks like a secret — anything with `TOKEN`, `KEY`,
`SECRET`, `PASSWORD` or `AUTH` in it — is drawn as dots, with **Show `<name>`**
to reveal it. The reveal follows the variable's name and not its row, so
removing a row above it cannot slide the reveal onto a different secret.

What devpit handed out this way is taken back out of a [run's](/docs/runs)
output before it is stored.

## Where a profile is used

- **Default agent**, so a new terminal opens that account.
- The **Account** field of an [agent](/docs/agent-steps) or
  [session](/docs/session-steps) step. A step naming a profile that has gone
  fails rather than quietly falling back to the default: the difference between
  two accounts is somebody's bill.
- The account and model picker in [chat](/docs/chat). Once a conversation has
  spoken, its account is fixed.
- **Runs as**, the account an [orchestrator](/docs/orchestrators) speaks as.
  Only a profile based on Claude Code can be one.

A profile can do what its CLI can. One based on Claude Code is
[driven](/docs/claude-code); one based on any other CLI opens in a terminal and
nothing more.

A profile that sets `CLAUDE_CONFIG_DIR` is a separate installation of the CLI,
with its own skills and its own MCP servers — which is why those panes ask
which installation you mean. See [Skills and MCP](/docs/skills-and-mcp).

## Removing one

**Remove** in the editor. It is refused while a board step still names it, and
the refusal names the steps and their projects, because "a step uses it" on a
machine with six projects says there is a problem without saying where.
