- 1An open-weight model by default
The built-in default model runs on your own machine.
- 2Hosted only when named, with your key
A provider is used only when a model names it, and its key goes only to its own API.
- 3Nothing reported back
No telemetry: AdultCoder reports nothing about you, your code or your use of it.
privacy & compliance
Privacy is the product. Compliance is engineered, not promised.
Know where your code, prompts and keys go, who approved what, and what every model was told. AdultCoder's controls are concrete enough to show an auditor.
the controls
Eight controls you can check, not take on trust.
Each is a setting or a boundary in the product, the same on every machine that runs it.
Sandboxed commands
Every command the agent runs is sandboxed, with a network allowlist. It can't read AdultCoder's own settings or its audit log.
Keys out of reach
Provider keys live in the system Keychain. The agent's commands run without them, and a known key in output is redacted.
Permissions per project
Another repository starts read-only. Editing it takes an explicit upgrade, kept only in the project that asked.
Deletion safety
It stops before deleting anything it did not create, and can send what it deletes to the Trash instead.
The machine's policy
The sandbox, the audit log, unattended mode and media spending limits are the machine's to set. A cloned repository can tighten the sandbox or lower a limit, never loosen either.
An audit log built for export
Every model conversation, keys taken out, in owner-only files with a chain of hashes. ac audit verify checks it; ac audit export hands it to your audit server.
Signed, encrypted sessions
Sessions shared across machines are signed and encrypted, with keys you keep in the Keychain and rotate.
AI commits, marked
The commits AdultCoder makes carry a co-author line, so your history shows which changes an AI wrote.
next
Masking secrets and personal data before anything reaches a hosted model.
in development
Off by default, and the machine's choice: a project can turn it on, never off. Secrets (keys, tokens, passwords in configs and connection strings) and personal data (emails, phone numbers, card numbers, IBANs, ID numbers) in anything sent to a hosted model become placeholders like ⟨EMAIL_1⟩. The model works with the placeholders; the real values are put back only on your machine, in the files and commands the agent writes. The audit log records what was masked, never the values. Names and street addresses come later.
what we optimize for
Controls that hold, not privacy theater.
- AData reality
Know which model saw what, and where it ran.
- BPermission reality
Sensitive access needs a readable, bounded trust model.
- CEvidence reality
Approved plans, replayable sessions, marked commits and the audit log are the receipts.
what we do not claim
- 1No certificates we don't hold
AdultCoder gives your compliance work evidence and controls. It does not make you compliant on its own.
- 2No zero-risk fantasy
We do not pretend every workflow becomes harmless once AI is involved.
- 3No slogan security
Trust comes from boundaries you can check, not from words on a page, this one included.
security FAQ
Short answers to the hard questions.
If a team cares where trust is stored, this is where a product either becomes real or becomes marketing.
.adultcoder folder; machine settings and the audit log in ~/.adultcoder, which the agent's commands can't read; keys in the system Keychain. Code goes to a model only when a turn sends it, and only to the model you chose.