BetterSign™
Product: BetterSign ApproveGet the app
Ask for early access

Product · BetterSign Approve · in development

Your agents ask. You decide.

Signed, per-action approvals for AI agents.

Coding agents run commands, rewrite files and reach for the secrets in your .env. Most of the time that’s what you want. BetterSign Approve makes the risky steps wait for you: you see the exact command, and approve it with your fingerprint, on your desktop or your phone.

A hook in the agent’s tool sends each risky action to bs-gate, which asks your BetterSign identity over a local socket (desktop) or an end-to-end encrypted relay (phone). The app shows the raw action and answers with a decision signed by your identity key. No answer by the deadline is a denial.

  • See the real command, not the agent’s description of it
  • Approve once, for a few hours, or say no with a note
  • Keep your keys out of files: hand one to one command at a time
  • Per action kind: exec, write, net, tool; delete and multi-step plans in the protocol
  • Ed25519 receipts over the action’s hash; HPKE to a one-time key
  • Fail closed on timeouts, errors and bad signatures

The agent asks. You decide. Nothing risky happens without you.

In development

BetterSign Approve is built and being tested with Claude Code, with approvals on the BetterSign desktop app (macOS, Linux) and on Android phones. It isn’t released yet. Want to try it early? Tell us which agent you use.

Benefits

Why not just let it run?

Because you only find out what an agent did after it did it. Approve puts you in front of the few steps that matter, without making you watch the rest.

◎

See what will really happen

You see the exact command or file change, in full. The agent’s own explanation is shown too, clearly labelled, so a confused or tricked agent can’t talk you into something.

The sheet shows the raw action; the agent’s summary is untrusted and shown de-emphasised. The decision binds to the SHA-256 of the full action, so it can’t be swapped after approval.

⚡

Seconds, not supervision

Everyday reads pass straight through, and the agent keeps working. You’re only asked about the steps that change something, and you answer with a fingerprint.

Local lists decide first: a blocked list is refused on the machine; routine reads are left to the agent tool’s own permissions. Only the rest goes to the app, behind App Lock.

⏱

Say yes for a while

Approve a command once, or let the same exact command run for the next few hours. Web requests are always asked, and a shell command is never allowed “for anything”.

Rules are signed by your identity, scoped to an exact command (or prefix set in Settings), folder or session, capped at 24 hours, and re-checked every 15 minutes, so a pause takes effect quickly.

✎

Say no, and why

Deny with a note like “use the staging branch”, and the agent reads it and changes course instead of just stopping.

The note travels in the signed decision and is passed to the agent (for Claude Code, as the hook’s reason).

⚿

Keys out of files

Your database URL and API keys live in BetterSign, not in a .env file. Approve a command, and the key goes to that one command and nowhere else.

bs-gate secret run NAME -- cmd: the value is sealed to a one-time key, set only in the child’s environment, and wiped afterwards. Never written to disk or logged.

⛓

A signed record

Every request, approval and refusal is recorded, and each approval is signed by your identity, so you can show who allowed what, and when.

Approvals carry a receipt signed by your identity key over the action’s hash, valid for at most 10 minutes and checkable by anyone who can resolve your VLAD.

⏸

Pause everything

Pause one agent, or all of them, with one tap. Their requests are refused until you resume.

A paused grant answers every request denied/suspended without a prompt; the host then drops its cached rules.

∅

No answer means no

If you don’t respond, or anything goes wrong on the way, the action simply doesn’t happen.

Timeouts, relay errors, bad signatures and crashes all deny. Requests live at most five minutes.

What it covers

Every kind of step an agent takes.

Each request says what kind of action it is, and that decides what you see and what you can allow in advance.

When the agent wants to…You seeAllow for a while?
Run a shell commandThe exact command and the folder it runs inYes, for that exact command
Create or change a fileThe file and the changeYes, for this session
Fetch a web page or searchThe address or the searchNo, asked every time
Use another toolThe tool and what it was givenYes
Use a secretWhich secret, for which commandNo, asked every time

A delete inside a shell command is a shell command, so you see it in full. Approving a whole multi-step plan at once is built into the app and coming to the agent side.

How it works

Pair once. Then just answer.

There’s nothing to sign up for. The agent’s computer pairs with your BetterSign app, and from then on its requests come to you.

Pairing exchanges keys with a QR or pasted offer and a 6-digit check code. Each request is an envelope signed by the agent host and sealed to your identity’s encryption key; the answer is sealed to a one-time key the request carried.

  1. Run bs-gate pair where your agent works, scan or paste the code into BetterSign, and check that both show the same six digits.

  2. Add one line to your agent tool’s settings. For Claude Code, that’s a hook that runs before each tool call.

  3. When the agent reaches a risky step, BetterSign shows it to you, on your desktop if you’re there, or on your phone.

  4. Approve with your fingerprint, allow it for a while, or say no with a note. The agent carries on, or changes course.

Who holds what

YouYour BetterSign identity, your secrets, and every decisionDesktop app or phone, behind App Lock
The agentThe ability to ask, and nothing elseClaude Code, a script, a CI job
bs-gateA pairing key, your blocked and routine lists, and rules you signedOn the agent’s computer
The relayPasses sealed requests to your phone; can’t read themOnly for the phone; the desktop is asked directly

Approvals and rules are signed by your identity, so bs-gate can’t forge or widen one, and neither can the relay.

Straight answers

What Approve does, and what it doesn’t.

Approve decides whether an agent’s next step goes ahead. It’s not a sandbox, and we’d rather you knew where the edges are.

QuestionAnswer
Does it limit what an approved command does?No. Approving ./deploy.sh approves whatever that script runs. That’s why you see the full command first.
Can an agent get around it?An agent asks through its tool’s hook. Keep the hook’s settings where the agent can’t edit them, and run the agent with only the access it needs.
Does a secret stay hidden from the command?No: the approved command receives it, as it would from a .env file. It just never sits in a file, and only that command gets it.
Do git or my deploys check approvals?Not yet. Each approval is signed so they can; checks for git pushes and deploy steps are next.
Which agents work today?Claude Code, through its hooks. Any other agent, script or CI job can wrap a command with bs-gate run.

For developers

Pair, then put the hook in your user or managed Claude Code settings (not the project’s, which the agent can edit).

# pair this computer with your BetterSign app
bs-gate pair --label "Laptop · Claude Code"

# ~/.claude/settings.json
{ "hooks": { "PreToolUse": [ { "matcher": "Bash|Write|Edit|WebFetch",
    "hooks": [ { "type": "command", "command": "bs-gate hook claude-code", "timeout": 330 } ] } ] } }

# any command or CI step
bs-gate run -- ./deploy.sh production

# a secret for one command, from your vault
bs-gate secret run DATABASE_URL -- npm run migrate

Questions

Frequently asked.

Can I use it today?

Not yet. It’s built and in testing with Claude Code. Ask for early access and tell us which agent you use.

Desktop or phone: where do I approve?

Wherever you are. At your desk, requests appear in the BetterSign desktop app straight away, with no internet needed. Away from it, they come to your phone.

Won’t it ask me about everything?

No. Everyday reads like listing files or viewing a diff pass straight through, and you can allow a command you trust for a few hours. You’re asked about the steps that change things.

What if I’m not there to answer?

Then the step doesn’t happen. The agent is told it wasn’t approved and can wait, try something else, or stop.

How is this different from BetterSign API Delegate?

API Delegate gives an agent a token to use a service’s API within limits you set in advance, without ever holding your credential. Approve covers what an agent does on a computer, step by step, with you saying yes at the moment it matters. Many people will want both.

Can the relay see what my agent is doing?

No. Requests to your phone are sealed so only your BetterSign identity can open them; the relay only passes them along. On the desktop, nothing leaves your computer.

Let your agent work. Decide what matters.

BetterSign Approve is in development. Tell us which agent you use, and we’ll let you know when you can try it.