Setup guide / Claude Code
Set up curio for Claude Code.
Claude Code reads skills from .claude/skills. One command puts curio's skill there, and from then on Claude Code works on your workspace through curio, in a session of its own.
Claude Code · the session the steps below set up, played on this page
Setup / 01
Four steps, once per project.
The agent gets a curio session and nothing more. Its edits wait in Sudo's sandbox, built against the workspace, until a person approves the commit.
-
01
Install curio
On the machine the agent works on. Skip this if curio is installed already.
Install script
The self-contained build for your system, checked against the release and put on your PATH. No .NET needed. Run it again to update.
$ curl -fsSL https://curiosity.sh/install.sh | bash $ curio tuiPS> irm https://curiosity.sh/install.ps1 | iex PS> curio tuiInstalls to
~/.curiosity/bin· read the script.NET tool
With the .NET SDK installed, curio is a global tool.
dotnet tool updatekeeps it current.$ dotnet tool install --global Curiosity.Shell $ curio tuiOr take the single file for your system from the releases
-
02
Sign in to your Studio
The browser signs you in with any login the workspace supports, SSO included. It needs a system administrator account. Leave the approve checkbox on the sign in page unticked, and every commit the agent stages waits for a person.
$ curio login --server https://my-workspace.curiosity.ai $ curio whoami -
03
Install Claude Code
Anthropic's native installer puts
claudeon your PATH and keeps it updated. Run it in your project and sign in with your Claude account.$ curl -fsSL https://claude.ai/install.sh | bash $ cd my-project $ claudeOn Windows:
irm https://claude.ai/install.ps1 | iex. Or with Node.js:npm install -g @anthropic-ai/claude-code. -
04
Give it curio's skill
In the project,
curio skills installwrites.claude/skills/curio/SKILL.md. With--globalit goes to~/.claude/skillsand every project on this machine has it.$ curio skills install # this project $ curio skills install --global # every projectRun it again to refresh the index of Sudo's skills. A copy you edited is kept unless you pass
--force.
In the cloud / 02
Claude Code on the web
A cloud session runs in a fresh container, with no keyring and no browser to sign in with. Give the environment what curio needs to sign in by itself. Use a token without the approve right, so approval stays with a person.
01
Install curio in the environment's setup script: curl -fsSL https://curiosity.sh/install.sh | bash.
02
Set CURIOSITY_SERVER and CURIOSITY_TOKEN as the environment's variables.
03
Allow the workspace's domain in the environment's network access.
Use it / 03
Ask in your words. Approve in your browser.
Start claude in the project and ask for the change in your own words. Claude Code picks up the skill when the request is about the workspace.
You ask
Add a code endpoint to the workspace that lists the open supplier notices for a part. Build it and stage the commit.
The agent reads the skill, then works through curio, one command at a time. You see each command; the last one waits for you.
$ curio skills search "code endpoint"
$ curio skills show add-code-endpoint
$ curio cat /workspace/config/schemas/nodes/Part.cs
$ curio put open-notices.cs /workspace/code/endpoints/open-notices.cs
$ curio run -- build
$ curio run -- commit -m "Open supplier notices for a part"
$ curio commit approve --no-wait # a person reviews it
$ curio commit wait --timeout 120 # until they answer
01
Let it run curio without asking
Add "Bash(curio:*)" to permissions.allow in .claude/settings.json. Approval still waits for a person: the right to approve is the token's, not the agent's. To keep it out of the full screen interface, add "Bash(curio tui:*)" to permissions.deny.
02
Exit codes mean something
curio run passes the sandbox command's exit code through. 77, 78, 75 and 64 mean not signed in, not allowed, conflict and usage, so the agent can tell a failed build from a missing sign in. 79 is a commit waiting for a person, and 124 is curio giving up waiting.
03
JSON, and no waiting forever
The skill has the agent pass --json for one line of JSON, --yes where a command asks, and --timeout on every call. A long build runs with curio run --detach and is picked up with curio run wait.
04
Never the interface
The skill tells the agent to never run curio tui. It is the full screen interface for people, and it would hang the agent's terminal. Plain curio only prints the help.
05
Re-read before overwriting
Sudo or an admin may edit the same session. curio get prints the file's hash, and curio put --expected-hash <hash> refuses the write if the file changed since.
06
Approval stays with a person
Signed in without the approve right, the default, curio commit approve --no-wait opens Sudo's review screen in the browser and exits 79 at once. curio commit wait then hears the answer, whatever the agent does.
Other coding agents / 04
The same curio, from another harness.
OpenAI Codex
Skill in .agents/skills
Codex reads skills from .agents/skills in the repository, in the CLI, the IDE and its cloud tasks.
GitHub Copilot
Skill in .github/skills
Copilot reads .github/skills. For the cloud agent, add the workspace token as an Agents secret and the workspace to its firewall allowlist.
Google Antigravity
Skill in .agents/skills
Antigravity reads workspace skills from .agents/skills, in the app, the IDE and the agy CLI.
Mistral Vibe
Skill in .agents/skills
Vibe reads .agents/skills in a trusted folder, and can run offline on a local Devstral model.
Devin
Skill in .agents/skills
Devin reads .agents/skills in every repository it works on. Its VM signs curio in from two secrets, CURIOSITY_SERVER and CURIOSITY_TOKEN.
Any other harness
Skill in any folder
If it can run a shell command, it can run curio: JSON output, exit codes that mean something, and one skill that says how.
Setup guide