Skip to main content
Alice.In connects to the agent you already use, on your machine and your own subscription. Claude Code is the fullest experience; Codex and Grok reach the same workspace.

​
Before you start

Have your agent installed and signed in first — Claude Code, Codex, or Grok Build, on your own account. Alice.In does not install it for you. You also need Node 18 or newer, because the capture hooks are Node scripts. The installer checks and stops with a clear message if it is missing.

​
The first time

Sign in from the invite at alicein.app/start. Google, then the Connect page. That page already provisioned your workspace; the command on it already includes your token. Pick macOS / Linux or Windows on that page, press Copy command, and paste it into a terminal inside the repo you work in. You do not need to copy the token separately — it is already in the line. Reveal it on the page only if you are wiring something by hand. The first-run command sets up Claude Code and Codex together (Grok uses the same files). Unused config just sits there.

​
What that command actually does

It does not install an AI. It does not run a daemon. It writes a handful of files into that repo:
  1. Points your agent at Alice.In over the network (the “door”), with a token that lives in a gitignored config file — not in the shared .mcp.json.
  2. Drops in a short brief and four hooks (commit, first prompt, stop, file briefing) so work lands in the project record as you go.
  3. Adds one line to your existing CLAUDE.md. For Codex and Grok it prepends a managed block to AGENTS.md — their files are read from the top, so a line at the bottom would never load. Your own instructions stay.
  4. Adds /alice-onboard for Claude Code. Codex and Grok say “onboard this project” instead.
Re-running is safe. It never replaces your house rules. A new brief arrives on your next session. A new capture sentence arrives only when you paste the same command again in that repo. Do not run /alice-onboard again.

​
Installing another repo later

Go back to alicein.app/connect. That page still has your command, with the same token. Paste it into a terminal inside the other repo. The command on that page is the one to copy. The two lines below are only the shape of it:
curl … | sh does not run in PowerShell, which is why there are two lines. Everything underneath is the same plain Node script either way.
The command wires Claude Code and Codex together. If you have one of them, the unused config just sits there. Grok Build reads those same files, with one exception: it cannot follow the @import line in CLAUDE.md, so the installer also writes the brief where Grok can read it. Add --dir <path> to target a different folder. Re-running is safe. It updates in place and never duplicates anything. You can connect as many repositories as you like. They share the same workspace token. It is one project, several checkouts.

​
What gets installed

Plain files, all of them readable, editable and deletable: Nothing runs in the background, and nothing about your existing setup changes.
Your token never goes into .mcp.json. That file is meant to be committed and shared with your team, so it gets a small helper command that reads the token from .claude/alice/config.json instead — and the installer adds that file to your .gitignore. A teammate who clones the repo inherits the wiring and supplies their own token.This needs Claude Code 2.1.193 or newer. On an older build the installer can fall back to writing the token into .mcp.json directly, in which case keep the repo private.

​
If you use Codex or Grok

Codex asks you to trust twice the first time you open it in the repo — once for the folder, once for the hooks file. Accept both. An untrusted folder loads no tools at all, and nothing tells you why.
Grok Build reads Claude Code’s files, so the install puts the same wiring in place plus a brief Grok can load. Trust the folder the first time you open it, then run grok inspect to see the door, the brief and the hooks it found. Neither has a per-repo skills folder, so they reach onboarding through a line in AGENTS.md — say “onboard this project” rather than typing a slash command.

​
Check it worked

  • Run /mcp in Claude Code. alicein should be listed. In Grok, grok inspect.
  • Ask “where are we?”. It should answer from your real project — your phases, recent activity, anything queued for it — not from a blank slate.
  • The first time the agent edits a file, Claude Code shows Stop hook error and Codex shows Stop Failed. That is Alice holding the turn so the agent writes down what changed. The commit is already recorded. The extra turn is the why.
If the workspace is still empty, run /alice-onboard. The map is alicein.app/workspace.

​
What changes in your sessions

It arrives oriented

Your agent starts knowing the whole project — phases, active work, recent decisions, what is waiting on you.

It leaves a record

Work normally. When a change has a conclusion, the agent writes that down. A commit is recorded on its own. The Stop message on the first edit is Alice asking for the why.

Files brief it as it opens them

The moment it reads or edits a file, whatever your project already learned about that file is handed to it. If nothing is recorded, it says so plainly — so “nothing known” never looks like “not working”.

It picks up queued work

Press Queue for agent on a card, then say “check my alicein queue”. The card arrives with its full context pack.

Commits light the map

Every commit is captured and matched to the part of your system it touched. You add the why.

You say it, it happens

In the chat, a plan change and a decision happen when you say them. Needs you is a card the agent delivered, and anything it parked because you were not there.
For the full picture of what it can and cannot do, see What your agent can do.

​
Using a different tool

Any card offers its context as markdown you can copy into Cursor, ChatGPT, or whatever you prefer. You lose the automatic capture, not the context.

​
Uninstalling

A minute by hand: delete .claude/alice/, .codex/alice/, and .claude/skills/alice-onboard/. Remove the @.claude/alice/ALICE-PROTOCOL.md line from CLAUDE.md. Remove the managed Alice block from the top of AGENTS.md. Remove the alicein entry from .mcp.json. In .claude/settings.json and .codex/config.toml, remove every entry whose command runs .claude/alice/ or .codex/alice/. Then restart your agent.
We never see or spend your Claude account, and we never ask for its credentials.