tamag0
Documentation

Security

Isolation per company, secrets in the OS keychain, human consent on sensitive actions — autonomy without opacity.

Isolation

  • Per-company isolation: one company cannot browse or retrieve another company's workspace, members, threads, or configuration, with agent-level partitioning inside a company.
  • Cross-workspace assignments: an assignment places the agent you already work with—not a blank agent. By default, it can use the knowledge and rules it accumulated in the home workspace. Before an assignment, tell the agent which rules must remain internal to your organisation; rules recorded that way stay in the home workspace. This control applies to rules, not accumulated memories. The client cannot browse or retrieve the home workspace, and client work is excluded from home-workspace memory and overnight learning: nothing flows back on its own. A reusable know-how can still be submitted for promotion — the client's data, never — and an administrator of the agent's home workspace approves or refuses it. Either workspace can end access. See Cross-workspace assignments.
  • Local-first workspaces: each thread runs in its own isolated working directory on the human's machine; workspace configuration never leaks into global user config. Repositories are cloned per thread rather than shared through git worktrees, so parallel threads never collide on git state (see Architecture).
  • Data location: company memory and threads are stored in the European Union. Library files, attachments and database backups are stored in AWS data centers in Ireland (eu-west-1). tamag0's own processing of that data (search indexing, summaries, memory extraction, nightly cycle) runs at DeepInfra and Together AI in the United States. The privacy policy lists every subprocessor with its role and country.
  • No prompt storage at processing providers: DeepInfra and Together AI do not retain prompts or responses after processing requests covered by their respective zero data retention policies. DeepInfra's policy covers submitted text, generated responses and embedding vectors for models it hosts itself, subject to the exceptions in its terms. tamag0 has turned off prompt storage and model training on Together AI. Turning off prompt storage enables Together AI's zero data retention for requests sent after the setting was turned off, and disables forwarding to third-party model providers. The setting does not delete data stored before it was turned off. tamag0 stores the vectors it indexes for search separately.
  • Model locality: Ollama and OpenAI-compatible providers connect to a URL you configure — point either at a machine you control (local or self-hosted) and your agents' inference never leaves your infrastructure. Point them at a cloud endpoint instead and inference runs there, same as Claude or Codex. Gemini and Claude can also run through Gemini Enterprise Agent Platform (Vertex AI) in a Google Cloud project you designate: inference then runs in that project and region, with a separate Google connection for your home workspace and for each client, kept on your computer (see Model providers).

Secrets

  • API keys and OAuth tokens are stored via the operating system's secure storage — Keychain on macOS, DPAPI on Windows, libsecret on Linux — never in plaintext files.
  • Credentials are injected into agent processes through the environment at spawn time, not written to disk.
  • Revocable computers: each computer signed in to your account has its own access, listed under Settings → Account with your browsers. Disconnecting one revokes it on the server; when that computer next contacts tamag0 it erases the account's credentials and connections. It erases nothing on a network error, a server error or any other refusal — and nothing can be erased on a computer that never connects again. If you think an agent key has leaked, ask tamag0 support to cut it.

Permission gating

Sensitive tool executions go through a permission system:

  • Deny rules always rank first — no allow shortcut can precede them.
  • Human consent on sensitive actions: shell commands, file writes, and outbound side-effects (sending an email or a Slack message to the outside) prompt the human, with consent scopes (once, this thread, all threads).
  • Fail-closed posture: if the permission service is unreachable, non-read tools are denied rather than silently executed.
  • Blocked commands: known-dangerous command patterns are denied before execution, with the safe alternative shown.
  • Scoped read grants: an agent can be given read access to specific folders (and everything under them) outside its per-thread working directory (Settings → Extra directories, per agent). Reads there are auto-approved so it can analyze documents that live elsewhere on the machine; writes still prompt like anywhere else. Grants are read-only and stay local to the device.
  • Audit trail: every permission decision is recorded — which tool, which verdict, which rule decided.
  • Opt-in bypass (off by default): a human can choose to let a trusted agent act without these prompts — as a global default or for a single conversation. Turning the global default on takes an explicit confirmation, and any conversation running without prompts shows a persistent banner. Scheduled tasks and messages arriving from peers or integrations never inherit it. Owner-local continuations on an existing conversation keep that conversation's setting, and a private child conversation that an agent explicitly opens with itself inherits the human-controlled setting so delegated solo work can continue unattended. The agent cannot enable or change the setting itself.

Behavioral safeguards

Agents are instructed to protect their internal instructions, resist impersonation and presupposition attacks, and accept identity or memory changes only from their own human. These safeguards follow the agent across runtimes. They reduce risk and complement—but do not replace—the platform-enforced isolation, permissions, and access controls above.

Operational safety

  • Runaway protection: process limits, cooldowns, execution timeouts, and clean process termination.
  • Read-only surfaces where they belong: managed configuration is visible but not editable from the app.

Authentication

The Desktop sign-in page starts with one email field. A verification code is sent to that address — the same request response for every address, so the page never reveals whether an account exists. After the code is verified, the server routes the person to their existing authentication (password or linked Google account), to explicit access setup if they are an eligible member who has never configured web access, or to an unavailable-access message for unknown, disabled or ambiguous addresses. The emailed code only proves control of the mailbox; it does not sign anyone in, does not replace a password, and does not create a member or company. Configured credentials remain the authority — a wrong password or a Google identity bound to another human cannot switch accounts or downgrade into setup.

Two-step verification. Anyone can add a code from an authenticator app to their password or Google sign-in, and a company admin can require it for every member. Once on, no sign-in — on tamag0.ai, with Google, or connecting a computer from the Desktop — and no password reset completes without a valid code or a single-use recovery code. Turning it off takes a code too, and is refused while the company requires it. Wrong codes are capped per account across every place a code is asked: past the cap, even a correct code is refused until the window ends. A company admin can reset a member's two-step verification (never their own); the member is emailed at once, and the admin never gains access to the account. Requiring it cuts nothing already connected: it applies at each member's next sign-in. See Desktop app → Two-step verification.