Profiles and models

Irie’s settings live in one file. It says which agents and AI models Irie can start, where your projects are, and how Irie behaves.

The settings file

$ irie config path
/home/alice/.config/irie/config.toml

The file is written in TOML: [sections] and key = value lines. Irie doesn’t make it; until you do, every setting has its default.

Every setting is optional. A setting Irie doesn’t know stops it with an error that names the setting, and so does a value that makes no sense, such as a default profile that doesn’t exist.

$ irie config print

This prints the settings in effect: the defaults, with your file on top. Use it to check a change.

If you set XDG_CONFIG_HOME, the file is $XDG_CONFIG_HOME/irie/config.toml instead.

When a change takes effect

Irie reads the file when its core starts, and a running core keeps the settings it started with.

• ui.sidebar_width and ui.focus_key take effect at the next irie or irie up, which restarts the core for them.
• For any other setting, restart the core yourself. Your agents keep running.

tmux -L irie respawn-pane -k -t irie-agents:core

Run irie config print first: a core can’t start with a file that has an error in it, and the list then shows core offline until you fix it.

The list reads ui.red_on_done when it starts. After changing that one, restart the list as well:

tmux -L irie respawn-pane -k -t irie:0.0

Profiles

A profile names the program to start, the model to ask for and its settings. A task uses spawn.default_profile, unless its first line names another with +.

The built-in profile, claude, is Claude Code with Opus 5.5 at max effort in auto mode:

[profiles.claude]
harness = "claude-code"
command = "claude"
model = "claude-opus-5-5"
effort = "max"
permission_mode = "auto"
args = []
env = {}

To change it, set only the keys you want. This keeps the rest of claude and lowers its effort:

[profiles.claude]
effort = "high"

A new profile starts empty instead: Claude Code, with no model, effort or permission mode, so Claude Code’s own defaults apply. Give it the ones you want:

[profiles.fast]
model = "claude-sonnet-5-5"
effort = "low"

Then +fast fix the typo in the README uses it. spawn.default_profile = "fast" would make it the default.

What a Claude Code profile runs

claude <args> --model <model> --effort <effort> --permission-mode <mode> \
  --settings <file> --session-id <id> --name <name> <task>

• A setting left empty leaves its flag out.
• --settings adds Irie’s hooks for this agent only: small commands Claude Code runs to tell Irie how it’s doing. Your own settings files aren’t touched, and your own hooks still run.
• --name is passed only when claude --help lists it.
• If Claude Code turns down a value, such as a model your account can’t use, the agent fails with Claude Code’s own message. Irie never swaps in another value.

Profiles for local models

Claude Code can use a model on your own computer through a server that speaks Anthropic’s API, such as Ollama, LM Studio or llama-server. A profile points it there with environment settings. This one uses Ollama’s settings for Claude Code:

[profiles.local]
model = "gpt-oss:20b"
env = { ANTHROPIC_BASE_URL = "http://127.0.0.1:11434", ANTHROPIC_AUTH_TOKEN = "ollama", ANTHROPIC_API_KEY = "" }

Fetch the model first, with ollama pull gpt-oss:20b. Then start small jobs with +local, such as @shop +local add comments to cart.go.

Give it room. A model holds only so much text at a time, its context, counted in tokens: pieces of words. With under 24 GiB of video memory, Ollama gives a model 4,096 tokens and quietly cuts anything longer, so an agent can lose its instructions without a word.

Ollama asks for at least 64,000 tokens for coding tools. Start it with OLLAMA_CONTEXT_LENGTH=64000 ollama serve, or set it in the Ollama app’s settings: see Ollama on context length.

How well it works is up to the model: Anthropic doesn’t support Claude Code on other companies’ models.

Irie checking a local model’s context before an agent starts, and saying how to fix it, is planned. Its local setting for profiles isn’t accepted yet. Planned

Any other agent

A profile with harness = "generic" starts any program that runs in a terminal. Irie then sees only whether the program is still running: its row shows working until it ends.

[profiles.aider]
harness = "generic"
command = "aider"
args = ["--model", "ollama_chat/gpt-oss:20b"]
task = "paste"

• command is required, and the program’s own flags go in args. model, effort and permission_mode aren’t allowed here.
• task = "arg", the default, adds the task as the program’s last argument.
• task = "paste" types the task into the program once it’s running, then presses Enter. It suits programs that take their task at a prompt.

Fuller support for more agents, starting with Codex CLI, is planned after version 1. Planned

Other agents covers them in full, including a launcher or an agent your own team built.

Every setting

These are all the settings Irie knows, with their defaults. Durations are written like "30s", "5m" or "24h", and folders as absolute or ~/ paths.

[ui]

[spawn]

[profiles.<name>]

The defaults of a new profile. The built-in claude profile has its own values, shown above.

[repos]

[worktree]

[tmux]

[journal]

[hold]

[dog]

The jot dog sees each task you start through to the end. When a task’s turn ends with work left in its own worktree, files not committed or commits not yet in its repo’s land branch, it sends the agent back once to finish, and the task’s card says what is left until the work lands. A task you close is done: the dog never sends it back or lists it, whatever it left. It only reads: it never commits, merges or deletes anything.

Settings for later milestones aren’t accepted yet, and Irie would stop with an unknown-setting error. [notes] comes with milestones 5 and 7, [freeze] with milestone 6, [subagents] with milestone 8 and [env] with milestone 9.