Agents and profiles
Adroit runs Claude Code, Codex CLI, Pi and OpenCode, each in the terminal interface it ships and with the program and login you already have for it. No agent comes first: every one is run the same way, side by side, and any of them can work in the same repository.
Profiles
Section titled “Profiles”A session starts from a profile, which names its agent, where the agent keeps its login, how it is launched and the agent’s own options. Profiles are kept in Settings › Agents. You can keep several for one agent, such as a work and a personal Claude Code, and choose which one new sessions start from.
On a draft, choose the profile, and with it the agent; Ctrl+g p and Ctrl+g P move through them. The draft also offers a model and permissions for that launch. A running session keeps the profile and login it started with, whatever the settings say since.
Without any profile set up, Adroit offers one for each agent, which runs it as it runs in your terminal, with Claude Code as the default.
Logins
Section titled “Logins”A profile runs either in its agent’s own environment, sharing the login and history your terminal uses, or with a login of its own: a folder where the agent keeps a separate login, history and settings. Log in from the profile’s page in Settings, which runs the agent’s own login in a terminal. Adroit keeps no credentials itself.
In settings.toml
Section titled “In settings.toml”Settings are kept in ~/.config/adroit/settings.toml, which the Settings
window writes. Profiles look like this:
profile = "claude-work"
[[profiles]]id = "claude"name = "Claude Personal"agent = "claude"
[[profiles]]id = "claude-work"name = "Claude Work"agent = "claude"home = "/Users/me/.local/share/adroit/profiles/claude-work"
[profiles.claude]model = "opus"
[[profiles]]id = "codex-review"name = "Codex Review"agent = "codex"
[profiles.codex]sandbox_mode = "read-only"profile names the default. home gives a profile a login of its own. The
settings reference
lists every setting.
Read-only and custom permissions
Section titled “Read-only and custom permissions”A draft can start its session read-only, or with permissions of your own. What that keeps a session to depends on its agent:
- Claude Code loses its file-editing tools and runs its shell commands in Claude Code’s sandbox, which keeps them from writing.
- Codex uses its read-only sandbox.
- Pi asks before no tool call and has no sandbox, so a read-only session is given only the tools that read files.
- OpenCode runs as its plan agent, with only the tools that read.
How each agent behaves
Section titled “How each agent behaves”Claude Code gets the most: Adroit can take its suggested prompt into the
prompt with Tab, a ! line runs in its bash mode, and the details
list the subagents it sets to work. It is reminded of its session’s name, and
of an empty timeline, with each prompt you send.
Codex reports its working and waiting states, including permission and question prompts, through the terminal. Its menus and approvals take keys directly while they are open; Ctrl+g i passes every key to it for its own commands.
Pi recalls the prompts sent in the session’s directory, as Pi keeps
them, and renaming a session in Adroit renames it in Pi with /name. A Pi
profile can name a model of a server of your own, such as ollama/qwen3:8b.
OpenCode needs version 2.1 or later of @opencode/cli. Each session runs
with a server of its own. It can use a local model too.
Every agent’s sessions list the prompts sent to them, and the links, files and images their conversation brought up, in the details; every one records a timeline, names its session, and can open sessions and talk with the others.
In the console, Ctrl+Enter, which sends the prompt, reaches the agent as Enter.
Screen rules
Section titled “Screen rules”When an agent’s program reports no state, Adroit reads how the session stands from its screen: working, idle, or waiting on an approval or a choice, and whether a dialog of the agent’s own holds the keys. A request on the screen wins over a report that it is working, and work still running after a turn ends, such as subagents or shells, shows the session as in the background.
It reads the screen by a set of rules for each agent, written as data. A file
~/.config/adroit/screen-rules/<agent>.toml on the machine where the
sessions run replaces the whole set Adroit carries for claude, codex,
pi or opencode, and is read again as soon as it changes. Adroit’s own sets,
under crates/agents/rules/ in its source, are a starting point:
agent = "codex"
[lines]chevron = '^\s*›'interrupt = '^\s*• .*esc to interrupt'
[regions]composer = { last = "chevron" }above = { before = "composer" }footer = { within = "above", since = "chevron" }
[[rules]]id = "interrupt_hint"state = "working"when = { in = "footer", has = "interrupt" }
[[rules]]id = "composer_idle"state = "idle"default = trueshows = "composer"A file that cannot be used is set aside, and Adroit’s own set reads the
screen instead. adroit session explain <session> names the set in force, the
rule that decided the state and the rows it went by. The
format
is described in full in Adroit’s specification.