Irie at work

Each person runs Irie on their own computer, and their agents use your company’s AI accounts and rules. Nothing of Irie runs anywhere else.

Irie works the same at a company as anywhere: on each person’s own computer, with the agents and AI accounts they already use. This page is for whoever decides whether a team may use it, and for whoever sets it up.

Short answers

Question	Answer
What does it cost?	Nothing, and there’s no paid version. The Apache 2.0 license lets your company use it, change it and share it, with no warranty, as is usual for open source. Your agents cost what they would without Irie.
Is there a server, an account or a sign-up?	No. Each person runs Irie on their own computer, and nothing of Irie runs anywhere else.
What does it send over the internet?	Nothing of its own. Each agent talks to its AI company, or your company’s provider, as it would without Irie. The details
What does it keep?	Each person’s sessions, and a log of their agents’ steps for 30 days, on their own computer, in folders only they can open. The details
Does it see our credentials?	No. Agents sign in as they would without Irie, and Irie never reads their logins or sends them anywhere.
Does it change our Claude Code setup?	No. It adds its hooks to each session it starts, and your company’s managed settings still win. Below
What runs on our computers?	One program, `irie`, which needs tmux and git, and the agents you choose. The Mac app is optional. The details
Is it ready?	It’s early. The first release, 0.1.0, has no download yet, so each person builds it from its source code.
Who supports it?	Its community, in public on GitHub. Security problems go to the maintainers privately. There are no support contracts or promised response times.

Your company’s Claude Code settings

Many companies set Claude Code up for everyone with managed settings: rules an administrator deploys, which nobody’s own settings can override. Irie works inside them.

Irie gives each Claude Code session it starts a settings file of its own, with its hooks, the small commands Claude Code runs at each step, and its status line. Managed settings rank above that file, so wherever the two differ, your company’s rules win.

If your company sets	Then in Irie
`allowManagedHooksOnly`	Claude Code runs only the hooks and status line your company deploys, so Irie’s don’t run. Agents still work, but Irie can’t tell how they’re doing: each row turns stalled after 30 seconds, and shows no cost.
`disableAllHooks`, in managed settings	Every hook and custom status line is off, Irie’s too, with the same result.
`disableAutoMode`	Sessions Irie starts in auto mode start in default mode instead. Agents then ask before more of their steps, so their rows turn red more often.
Allow, ask and deny rules	They apply to every agent Irie starts. Irie never approves a step for an agent.
`availableModels`	Claude Code takes only the models on the list, whatever a profile asks for. Give each profile a model from the list.

To see which managed settings apply to you, run /status in Claude Code: its Setting sources line names them.

Your company’s AI provider

If your company set Claude Code up for its provider, such as Amazon Bedrock or a gateway of its own, in Claude Code’s settings or in managed settings, Irie’s agents use it too. There’s nothing to change in Irie.

If your provider wants its own names for the models, give them to Irie’s built-in profiles. On Bedrock, for one:

[profiles.claude]
model = "us.anthropic.claude-opus-5-5"

To pick a provider for one profile instead, put its settings in the profile’s env. This profile runs Claude Code on Amazon Bedrock, with the settings from Claude Code’s Bedrock page:

[profiles.bedrock]
model = "us.anthropic.claude-opus-5-5"
env = { CLAUDE_CODE_USE_BEDROCK = "1", AWS_REGION = "us-east-1" }

Claude Code signs in to AWS in the usual ways, such as your AWS profile or single sign-on. Irie never sees the credentials.
For a gateway, set ANTHROPIC_BASE_URL to its address instead, as Claude Code’s page on LLM gateways says. Google Cloud and Microsoft Foundry have pages of their own.
A new profile like this one leaves the permission mode to Claude Code. Add permission_mode to choose one.
+bedrock runs one task there. To run every task there, set spawn.default_profile = "bedrock" in place of auto, which chooses among Irie’s built-in profiles.

The cost Irie shows for a session is Claude Code’s own estimate, at list price. Your company’s bill comes from its provider.

Proxies and certificates

Behind a company proxy, Claude Code reads HTTPS_PROXY and NO_PROXY, and NODE_EXTRA_CA_CERTS for your company’s certificates, as its network page says.

Put them in Claude Code’s own settings or in a profile’s env, not only in a terminal. Irie’s agents start with the environment of the terminal that first started Irie, so a variable you set in another terminal later doesn’t reach them.

A company launcher

If your company starts Claude Code through a launcher, a program of its own that sets Claude Code up first, make the launcher a profile’s command. A launcher that starts Claude Code says what it has to do.

Your team’s own agents

An agent your company built runs in Irie too, whatever it’s built on. Other agents says how, and what Irie can see of it.

Setting it up for a team

Each person runs their own. Everyone builds Irie and runs it on their own computer, as Install shows. There’s nothing to run for the whole team.
Share one settings file. A team can hand out a config.toml with its profiles, project folders and choices. A setting Irie doesn’t know stops it with an error that names it, so a mistake shows at once.
Choose the permission mode. Irie’s built-in cloud profiles start Claude Code in auto mode, where Claude Code decides which steps are safe to take without asking. Set permission_mode to your team’s choice.
Keep tasks off local models, if you’d rather: leave local out of route.profiles.
Turn off what you don’t want. route.delegate = false stops Irie telling sessions they may hand subtasks to cheaper models, subtasks that edit files without asking. journal.tool_input = "none" keeps the agents’ steps out of Irie’s log.

A team’s file that does all of that could read:

[profiles.claude]
permission_mode = "default"

[profiles.sonnet]
permission_mode = "default"

[profiles.haiku]
permission_mode = "default"

[route]
profiles = ["haiku", "sonnet", "claude"]
delegate = false

[journal]
tool_input = "none"

Each person puts it at ~/.config/irie/config.toml, checks it with irie config print, and restarts Irie’s core, as Profiles and models shows.

What isn’t there

No downloads yet. Each person builds Irie from its source code, and updates it with git pull and a new build.
No central console, policies or audit trail across a team. Each person’s Irie keeps its own records, on their own computer.
No accounts, single sign-on or roles, because Irie has no accounts at all.
No Windows: Macs and Linux only.
No support contracts, promised response times or certifications. Help happens in public on GitHub, and SECURITY.md says how to report a security problem privately.

Check it yourself

What Irie is lists every part of Irie, every file it keeps, every connection it makes and every word it says to an agent, with commands to check them on your own computer.

Irie’s code, its design and every source it was built from are public.