@meta
  v: 1
  route: /titanwarden
  generated: 2026-11-07T17:41:17.966Z
  ttl: 1h

@intent
  purpose: Project page for Titanwarden, a local secret broker for machines where AI agents work: use, never see. An agent names a secret and a command; the broker checks which process is asking, writes the audit row first, starts the command with the value injected, and redacts the value from the output. The agent protocol has no operation that returns plaintext. Open source under Apache-2.0. Source: github.com/frontier-infra/titanwarden.
  audience: ai-agent, developer, operator, security-engineer
  capability: read, install, inspect-source

@state
  version: 0.1.0
  license: Apache-2.0
  platforms: macOS, Linux
  runtime: Python 3.12+, uv
  install[4]{step,who}:
    git clone https://github.com/frontier-infra/titanwarden && cd titanwarden,human-or-agent
    uv sync --frozen && scripts/install.sh,human-or-agent
    twadmin init,human-over-ssh
    twadmin unseal && tw status,human-over-ssh
  agent_ops[3]{command,returns}:
    tw run --env NAME=path -- command,command-output-redacted
    tw names prefix,names-and-versions-never-values
    tw who-touched path,audit-trail
  plaintext_to_agents: none
  restart: sealed_until_human_unseals

@actions
  - id: repo
    method: GET
    href: https://github.com/frontier-infra/titanwarden
  - id: license
    method: GET
    href: https://github.com/frontier-infra/titanwarden/blob/main/LICENSE
  - id: guide
    method: GET
    href: https://github.com/frontier-infra/titanwarden/blob/main/docs/guide.md
  - id: threat_model
    method: GET
    href: https://github.com/frontier-infra/titanwarden/blob/main/docs/threat-model.md
  - id: agent_usage
    method: GET
    href: https://github.com/frontier-infra/titanwarden/blob/main/docs/support/agent.md
  - id: site
    method: GET
    href: https://www.titanwarden.com

@context
  > Agents need credentials to do real work, but a password manager is built for a person: unlock, read, paste. Used by an agent, the value lands in its terminal, transcript, logs and next tool call. Titanwarden turns that around. One encrypted SQLite store per machine (AES-256-GCM values, Argon2id passphrase and recovery slots). Every restart starts sealed and a human unseals. Every use is audited: who, which secret, which version, when. An agent without a grant files a request a human answers. Runs on macOS and Linux with Python 3.12 or newer and uv. First setup is human-only: an agent may run uv sync --frozen and scripts/install.sh, then must stop; twadmin init, unseal, recovery rotation and reveal are for humans over SSH. Honest limits today: broker, agents and children run as the same OS user, so it prevents accidents and records honest use but is not a boundary against a deliberately hostile local agent; output redaction is a guardrail; interactive programs are not supported under tw run. Content taken from the public README and docs/guide.md in the repository; install steps match titanwarden.com.

@nav
  self: /titanwarden.agent
  parents: [/.agent]
  peers: [/.agent, /jebadiah.agent]
