arca
/ˈar.ka/ · Latin: a strongbox for keeping valuables under lock
age-encrypted secrets, kept safely in front of AI agents.
/ˈar.ka/ · Latin: a strongbox for keeping valuables under lock
age-encrypted secrets, kept safely in front of AI agents.
Per-value age ciphertext with cleartext metadata in one JSON store. Reuses your existing sops key. No daemon, no account.
A command execs with a secret while the value stays out of stdout and the model's context — and if the command prints it anyway, arca redacts it from the output.
Every access is hash-chained and signed per session, so edits and deletions are detectable. arca log --verify proves the log is intact.
Plant realistic decoys with canary --template stripe. Any use trips a loud, signed alert of a leak detection.
Bind a secret to what an agent does: a grant scoped to a command, a use count, and a time window — not just whether it can see the value.
An arca mcp server exposes audited tools, plus --no-print, approval gates, and TTLs a detected agent can't bypass.
The value lands in the command's environment, never stdout. If the command prints it, arca swaps it for a marker before an agent can read it.
$ arca exec --only DB_PASSWORD -- sh -c 'echo connecting with $DB_PASSWORD' connecting with «arca:DB_PASSWORD» # the value never reaches the agent
Plant a realistic decoy. The moment anything uses it, a signed alert fires and the trip is recorded in the tamper-evident log.
$ arca canary AWS_PROD_KEY --template aws # ... later, if anything reads it ... ⚠ CANARY TRIPPED: "AWS_PROD_KEY" was accessed by malicious-agent
A require-grant secret is unusable until you scope it to a command, a use count, and a window — then it auto-expires.
$ arca grant DEPLOY_KEY --command 'terraform *' --uses 3 --ttl 15m $ arca exec --only DEPLOY_KEY -- terraform apply # allowed (1 of 3)
Every event is hash-chained and signed per session. One command walks the chain and every signature.
$ arca log --verify
audit log OK: 128 event(s) chained, 128 signed
Every command, flag, and the policy model — the day-to-day reference.
no-print, approval gates, canaries, grants, and rate limits for agents.
Expose audited, capability-scoped tools to a model over stdio.
Actors, components, trust boundaries, and the security assessment.