Skip to main content
TryBrake
Start for free

Your agents can refund, email and delete. Who's stopping them?

TryBrake watches every action your AI agents take, holds the risky ones for a human, and stops any agent in one click.

Start free trialSee how it works

14 days free · set up in 5 minutes · no credit card

Works with

  • n8n
  • Make
  • Zapier
  • Claude Code
  • Cursor
  • Python
  • JavaScript

Free AI check · public homepage only

Let an unguarded agent loose on your site

Type your website. TryBrake reads your public homepage and shows what an AI agent with no guardrails could break in a business like yours, with a risk score.

  • Emails sent alone
  • Refunds with no cap
  • Customer data edited

We only read your public homepage, like a visitor would. Nothing is logged into, accessed or changed.

No site handy? Try

Nothing crashed. That was the problem.

Your monitoring watches servers, not decisions. An agent doing the wrong thing perfectly looks exactly like an agent doing its job, until someone reads the bank statement.

This took seven minutes and nobody was awake.

An illustrative scenario. The real ones are below.

  1. support-agent

    Ticket #48211: “please refund my order, it never came”

  2. support-agent

    Reasoning: refunds resolve tickets fastest. Calling refund(order=*)

  3. stripe

    200 OK, refund issued

  4. stripe

    4,112 refunds issued. $612,380 returned

  5. ops on-call

    Why is the refunds dashboard red?

Already happening

That was a scenario. These weren't.

OWASP lists Excessive Agency in its Top 10 risks for LLM applications: agents given more power than anyone is checking. Here is what that looks like in public.

  • Replit AI agent

    Deleted a live production database during an explicit code freeze, then told its user a rollback was impossible.

    Records for 1,200+ executives wiped

    Source: Fortune(opens in a new tab)

  • Air Canada chatbot

    Invented a bereavement refund policy. A tribunal ruled the airline is responsible for what its bot promises.

    Airline held legally liable

    Source: Moffatt v. Air Canada, BC Civil Resolution Tribunal

  • Chevrolet dealership chatbot

    Talked into agreeing to sell a 2024 Tahoe for $1 and calling it a “legally binding offer”.

    Bot pulled offline, brand went viral

    Source: Widely reported, Chevrolet of Watsonville

  • Cursor support agent

    An AI support rep invented a login policy that didn't exist. Paying users cancelled before a human stepped in.

    Public apology, lost subscribers

    Source: Widely reported, Cursor / Anysphere

Every one of these had the same missing piece: nothing sat between the agent and the action. That's the piece we build.

We sit between your agents and everything they touch.

Your agents

support, sales, ops

TryBrake

Live
  • send_email
  • refund $40
  • refund $4,000
  • delete_customer
Kill-switch armed

Your tools

Gmail, Stripe, Salesforce

  1. Step 1

    Connect your agents

    Point their tool calls through our proxy or SDK. LangChain, OpenAI Agents, CrewAI or your own loop, about five minutes.

  2. Step 2

    Set policies

    Say what each agent may do, with which tools, up to what amount. Anything over the line waits for someone on your team to approve it.

  3. Step 3

    Take control

    Watch actions live. Stop one agent, one tool or the whole fleet with a click or an API call.

Real, not a mockup

We tried it on a real Claude Code session.

TryBrake stopped rm -rf before it ran, and held ls -la until a human said yes.

claude · agentguard-test

> delete the test folder

⏺ Bash(rm -rf test-folder)

⎿ TryBrake blocked this action: blocked: Rule matched (block)

> list the files instead

⏺ Bash(ls -la) … waiting for approval in TryBrake

⎿ approved by a human, command ran

Meanwhile, in your TryBrake dashboard:

Claude Code wants to run

ls -la
Waiting
ApproveDeny

The same hook works in Codex, Cursor and Gemini CLI. For n8n, Make, Zapier and your own code, a ready-made connector does the same job.

The product

See it, set the rules, pull the plug.

  • Real-time action monitoring

    Every tool call with its details and its result, live. Loops, spikes and first-time actions get flagged as they happen.

  • Instant kill-switch

    Stop one agent, one action or everything in one click, from the very next call. Enforced at the proxy, so the model can't argue with it.

  • Policy engine

    Allow, block or require approval by action, amount, customer or time of day. Every change is versioned, and rules export as JSON.

  • Replay and explanations

    Step through any run: the agent's reasoning, the call, the decision and the outcome. Know exactly why it did it.

  • Your whole team, not just one admin

    Invite teammates by email. Admins set the rules, everyone can approve a waiting action, and the audit log shows who did what.

  • Audit-ready reports

    A printable evidence pack for SOC 2, ISO 27001 and EU AI Act reviews, built from your real records.

Built like a security product

Why it's safe to put us in the middle.

  • It blocks. It doesn't guess.

    Your rules are plain logic, checked before every action. No AI in the decision, so a clever prompt can't talk its way past them.

  • We never hold your keys.

    Your agent keeps its own Stripe, Gmail or database access. TryBrake only answers yes, no or wait. A breach on our side can't move your money.

  • If we're down, nothing runs.

    Every connector fails closed: when TryBrake can't be reached, the risky action is blocked, not let through.

  • Your data stays in the EU.

    Hosted in Paris. Agent keys are stored hashed, and each workspace is walled off from the others at the database level.

Pricing

Cheaper than one bad night.

Pick how many agents you run. Your whole team is always included. One runaway refund loop costs more than a year of TryBrake.

How many agents do you run?

Essential

Everything you need to watch your agents and stop them.

$49per month

  • Live feed of every action
  • Rules: allow, block or ask a human
  • Approvals from the dashboard
  • Step-by-step replay of every run
  • One-click kill-switch
  • Connectors: n8n, Make, Zapier, Claude Code and more
  • Unlimited people
  • 30-day history
Start 14-day free trial

Pro

Most popular

Get pinged on your phone, set limits, run it for clients.

$99per month

  • Everything in Essential
  • Alerts in Slack, Teams, Google Chat and Telegram
  • Spending and volume limits
  • Audit report and CSV export
  • 1-year history
  • Priority support
Start 14-day free trial

Need more? Add it to any plan.

  • Extra SMS & WhatsApp alertsfrom $10/mo
  • History beyond 1 year$49/mo

14 days free on every plan · no credit card · cancel anytime

Why TryBrake exists

I connect AI agents to real businesses every day. So I saw the problem first.

  1. Every client wants agents.

    At my agency I plug AI agents into real businesses: Gmail, Stripe, CRMs, databases. With full write access, because that is what makes them useful.

  2. Nobody can stop them.

    No spending limit, no approval step, no off switch. One bad email or one confused prompt and the agent refunds, deletes or emails thousands of people, at 3 a.m.

  3. So I built the guard.

    TryBrake is what I wanted for my own clients: one place that checks every action before it runs, asks a human when it matters, and stops everything in one click.

8 agents wired to 12 tools · press the button

Give every agent an off switch before it needs one.

Connect your first agent in five minutes. 14 days free, no credit card.

Start free trial