Security
Where your code, credentials and agents live, what the server can ask your computer to do, and what your computer refuses.
Your agents run on your own computers, never on a Sculptured server. Each computer keeps its own rules: which folders agents may start in, and how much they may do. The server can ask for less than a computer allows, never more.
Where things live
flowchart LR
subgraph PC["Your computer: your trust"]
B[Browser]
R["Sculptured runner and its rules"]
A["Claude Code / Codex"]
F[("Folders you allowed")]
end
subgraph SV["Sculptured account: server"]
G["Gateway: sign-in, computers"]
I["Your own Sculptured: teams, runs, history"]
end
B -- "HTTPS, session cookie" --> G
G -- "only your requests" --> I
R -- "outgoing secure link, device key" --> G
I -- "launch, message, stop, folder and git requests" --> R
R -- "starts with your access level" --> A
A -- "reads and writes" --> F
| What | Where it lives |
|---|---|
| Your code and files | In the folders you allowed, on your computer. They reach the server only as results you ask for, such as a folder listing, Git status or a diff. |
| Claude Code and Codex sign-in | In the CLIs' own settings on your computer. Sculptured never reads or stores them. Sign in only starts the CLI's own sign-in. |
| Teams, runs, messages and transcripts | In your own Sculptured. With an account, that is one private instance per user on the server. Running it yourself, it is ~/.sculptured/ on your computer. |
| Your computer's key | ~/.sculptured/runner/device.json on that computer, readable only by your user. The server keeps only a hash. |
| Your account password | Stored as an argon2id hash. Google and GitHub sign-in store no password. |
See Privacy for what is recorded about how you use Sculptured.
Running it yourself
sculptured serve listens only on 127.0.0.1 and accepts only requests addressed to 127.0.0.1 or localhost, with a matching origin. That stops other websites from driving it through your browser. The runner connects to it with a private token that agents never see.
Connecting a computer
- A one-time connect keyThe Computers page copies a command that carries a connect key. It works once, for 30 minutes, and only a signed-in browser can create one.
- Traded for a device keyThe installer trades it for this computer's own device key. The server stores only a hash of it.
- An outgoing linkThe runner opens a secure link (
wss://) out to your account. Nothing connects in to your computer, so no port is opened. - RevocableDisconnect on the Computers page drops the live link at once and refuses that key from then on. The runner stops and says to run
sculptured connectagain.
A computer can only connect as itself. It cannot pretend to be another of your computers.
If a computer connects with a pairing code instead of a key, the approval page warns: only approve a code you just saw on your own computer. It shows when and from which address the code was requested.
What your computer refuses
The runner checks every launch, restart and local request against its own rules, whatever the server asks.
| The server asks to… | Your computer… |
|---|---|
| Start an agent in a folder you did not allow | Refuses: "… is not in a folder this computer allows." |
| Start an agent in a folder that doesn't exist | Refuses: "… does not exist on this computer." |
Reach outside a folder with .. or a symbolic link | Resolves the real path first, then refuses. A sibling folder such as /home/you/proj2 is not inside /home/you/proj. |
| Run a different program, add environment variables or extra arguments | Refuses. A computer only runs its own Claude Code and Codex. |
| Give an agent more access than the computer allows | Lowers it to the computer's level. |
| Run a check command when checks are turned off | Refuses, and the check is recorded as failed with the reason. |
A refused launch shows on the agent: it needs attention, with the reason.
Allowed folders
Allowed folders decide where an agent may start, not what it can do once running. The installer allows your home folder. Add more with sculptured allow <folder>, or choose them when you connect with --allow.
Access level
| Access | Claude Code | Codex | Set with |
|---|---|---|---|
| Full (default) | Runs as the team asks. An unrestricted agent skips permission prompts. | Full access | sculptured connect … --full-access |
| Sandboxed | Always restricted: risky actions need your approval in the browser. | Writes only inside its folder | sculptured connect … --sandboxed |
What the account may do on a computer
Each computer turns these on or off with --enable and --disable on sculptured connect:
| Capability | What it allows | Full access | Sandboxed |
|---|---|---|---|
git | Git status, diffs and commits from the Git panel | on | on |
copies | Agents' own copies of a repository | on | on |
checks | Running a team's check commands, such as tests | on | off |
cli-sign-in | Starting Claude Code's or Codex's own sign-in | on | on |
plan-usage | Reading how much of your Claude or ChatGPT plan is used | on | on |
Checks are off when sandboxed because a check runs a command from the team's design outside the agent's sandbox.
The Computers page shows each computer's folders, access and capabilities as the computer reports them. The computer enforces its own copy either way.
Agents and their tools
- Each agent gets its own token for Sculptured's tools. It works only while that agent's process runs, so one agent cannot use another's tools.
- The runner's own secret is removed from every agent's environment.
- An invalid tool call is recorded as evidence and answered with an error. It never interrupts the agent.
- A restricted agent is never launched unrestricted. If a runner cannot honour restricted, the launch is refused.
- Safety rules refuse the tool calls you name, for Claude agents, with no model call. They are a speed bump, not a sandbox: an agent can still reach the same result another way, so they add to the computer's own limits rather than replace them.
- The integrations an agent is given (Kindex, Signet-Eval) are programs the runner finds on its own computer; the server never sends a program to run.
- Sculptured's own files in a project folder are under
.sculptured/: the project map, agents' notes, and lessons you confirmed. When agents work in their own copies, Git also records those copies in the project's.gitfolder. Skills and Signet-Eval rules are written in the runner's own data folder, never in the project.
Accounts and sessions
- Invite-only. Joining needs an invite. Invites are single use, expire and can be revoked.
- Sessions end after 30 days without use; each use renews them. Sign out from the Account page.
- Every change from the browser must come from the Sculptured site itself and carry a CSRF token tied to your session.
- Password sign-in is rate-limited, and unknown emails get the same answer as wrong passwords.
- Google and GitHub sign-in are bound to the browser that started them and cannot be replayed.
- Your Sculptured accepts only requests the gateway signed for you. You reach only your own teams and computers.
- Admins manage invites, users and computers. They cannot open your teams or your files.
Community
Community holds published blocks, saves and public profiles, in its own database. It is served by Sculptured's own server for your signed-in account: your browser never calls another site for it, and Community never sees your session cookie. It has no access to your runner, transcripts, files or Git. Publishing strips machine-specific folders, model overrides and workflow input, but a job's text can still contain private details. Review it before you publish. See Community.
Not covered
- The device key is kept in a file readable only by your user. Keeping it in the system keychain is planned.
- A hostile server could ask your computer for anything its rules allow: starting Claude Code or Codex in an allowed folder, with your access level. Allowed folders and Sandboxed are how you limit that.