← All posts

Safety ·

Approving Everything Isn't a Safety Model. Here's What TRMNLY Does Instead.

TRMNLY's agent approvals come in three postures, with READ and WRITE tags on every tool, rules scoped to a host or folder, a destructive-command gate that fires even on Auto-accept, a signed audit log and per-write snapshots you can revert.

Small glass tiles on a dark track approaching a glowing glass arch, one tile paused at the gate

In Stack Overflow's 2025 Developer Survey, 81% of developers who use AI agents said they have concerns about the security and privacy of their data.1 Most agent tools offer two choices: approve every call, or approve nothing. TRMNLY gives you the settings in between.

Pick a posture

The chip in the agent sidebar header shows the session's posture. Press ⇧Tab to cycle it:2

  • Approve (the default). Tools that write show an approval card. Tools that only read run without asking.
  • Auto-accept. Every tool runs without a card.
  • Observe. Read-only. Writes and shell commands are blocked.

Posture is saved per project. Since 0.1.96 that includes Auto-accept, so a new tab on a production repo can't start in Auto-accept because a scratch project asked for it.2,3

Every tool is READ or WRITE

The Safety panel tags every tool READ or WRITE, plugin and MCP tools included, and WRITE tools ask by default. Turn on Always prompt for a single tool, like bash or browser_navigate, and it keeps asking even on Auto-accept.2

Rules as narrow as the call

An approval card can do more than approve once. You can allow or deny the tool for the whole project, or allow this path, allow this directory, or allow on <host>. The most specific rule wins, so a host rule beats a tool rule. A directory rule stops at a folder boundary: a rule on proj never covers proj-secrets.2,4

Since 0.1.108, deny rules also catch commands hidden behind sudo, env, subshells, quoting and bash -c.5

Some commands always ask

A short list of commands you can't undo, including rm -rf, git push --force, git reset --hard, DROP TABLE and mkfs, asks for approval the first time each one appears in a session, even on Auto-accept.2 Since 0.1.96 the same check covers calls from Claude Code and other MCP clients. Denying one of these commands no longer uses up its once-per-session prompt, either.3

A log that can't be quietly rewritten

The Audit Log records every tool call, approval, denial and rule change in an append-only file. Each entry is chained with an HMAC signature, and Verify integrity rechecks the whole chain. If you realize you approved too much, Revoke on any row locks that tool for the project and clears the rules tied to it.2,6

Undo for the agent's own edits

Before the agent writes a file with write_file or edit_file, TRMNLY saves a snapshot of it. To roll back, use the controls on the write card in the chat, or ask the agent to revert its snapshot. The revert itself needs approval, and it works on files git isn't tracking.7,8

The point

Trust isn't all or nothing. Run Auto-accept on a scratch repo, Approve on the one that deploys, and leave the destructive-command gate on everywhere.

Sources

  1. 1.Stack Overflow, 2025 Developer Survey, "AI," section "Challenges with AI agents." survey.stackoverflow.co
  2. 2.TRMNLY docs, "Agent safety." trmnly.dev/docs/safety
  3. 3.TRMNLY changelog, 0.1.96, "An approval means what the card said" (October 6, 2026).
  4. 4.TRMNLY changelog, 0.1.85, "Approving a tool call runs what the card shows, once" (September 13, 2026).
  5. 5.TRMNLY changelog, 0.1.108, "The agent sidebar, audited again: about 300 fixes" (October 17, 2026).
  6. 6.TRMNLY changelog, 0.1, "Safety gates and audit log" (May 13, 2026).
  7. 7.TRMNLY docs, "Orientation." trmnly.dev/docs/orientation
  8. 8.TRMNLY Claude Code skill docs, "trmnly-automation," section "Write snapshots."