How odette works
Understand what stands between an agent and your machine, and whether odette fits how you work. This page explains the parts, follows one request through them, and compares odette with other ways of running an agent unattended.
The parts
Section titled “The parts”- A session is one Claude Code agent in a container. It sees your project folder and its own home, nothing else of yours, and it has no network of its own.
- The daemon runs on your machine. Its proxy is the session’s only way out, its gates decide what passes, and it adds your real credentials to a request on the way out. The session only ever holds a placeholder.
- The dashboard is the web page where you see every session and answer what is waiting on you.
One request, start to finish
Section titled “One request, start to finish”- The agent runs
git fetch,npm installorcurl. The request leaves the container and reaches the proxy; there is no other route. - The proxy checks your rules: hosts you allowed or denied for this project or for every project (Approvals).
- A request no rule covers goes to the Egress gate. In
judgemode an AI judge allows it, denies it, or holds it for you when it is unsure. Inmanualmode it is always held. - A held request waits in the dashboard. You allow it once, allow it for the project or for every project, or deny it.
- An allowed request goes out. If it is for Claude, Gemini or Google, odette adds the real credential here.
A command an agent asks to run on your machine takes the same path through the Commands gate. Those run as you, outside the sandbox, so decide them with care (Security).
Where things live
Section titled “Where things live”| Where | What |
|---|---|
~/.odette/ |
Your settings, the allow and deny lists, stored secrets (encrypted). |
.odette/ in a project |
The container, extra mounts and the project’s allowed hosts. Read-only inside a session. |
| A project’s Claude home | Conversations, memory and settings for one profile in one project. |
odette keeps all of these on your machine.
When odette fits
Section titled “When odette fits”There are four common ways to run an agent without watching every step. The trade-off is what you set up yourself against what you accept as given.
| Approach | You get | You do yourself |
|---|---|---|
| The agent’s own permission prompts | A question before each action | Answer every one, in the terminal, at every step; a long run stops while you are away |
| A sandbox around the agent’s shell commands | Limits on what those commands can read, write or reach | Handle everything the agent does outside the shell; many tools, hooks and plugins run outside it |
| A container you build yourself | A boundary you control | Network rules, credential handling, approval flow, and one setup per project |
| odette | A container, an egress and command gate with an AI judge, credentials kept out of the session, one dashboard for many agents | Install it, keep an SSH key loaded, and answer what the judge holds |
odette fits when you run agents unattended or in parallel, when the agent needs credentials you do not want it to hold, or when a team needs one place to decide what agents may reach.
It fits less well for a quick edit you are watching live, for platforms it does not run on (Getting started), or where you cannot run a container runtime.