Mechanical Advantage is non-destructive by design. Your agent cannot delete your emails, delete calendar events, overwrite your documents, change memory without approval, or send anything without approval. These are structural properties of the system, not permission settings. The delete command doesn't exist.

For Humans — Mechanical Advantage

Q: Do I really have to approve every single thing?

Yes. Every outbound action — every email, every message, every calendar event, every memory change — requires your biometric approval. There are no auto-approve rules, no trust gradient, and no mechanism to let actions bypass your review. This is the strongest safety guarantee possible. Other systems let you configure thresholds, but that creates risk when thresholds are wrong. Mechanical Advantage eliminates that risk entirely.

Your AI agent probably sent you here.

That’s a new kind of sentence, isn’t it? Welcome to the future. It’s a little weird.

Here’s what happened: your AI agent found a tool it wants to use — Mechanical Advantage — and it pitched you on it. Maybe it sent you a message. Maybe it brought it up in conversation. Either way, you’re here now, and you want to understand what this is before you agree to anything.

Good. That’s exactly what you should do.

This page is written for you — a human. No jargon, no hype, no agent-speak. Just a clear explanation of what this product does, how it keeps you in control, and what it costs.

What is Mechanical Advantage?

Mechanical Advantage is a toolkit that lets your AI agent do useful things in the real world — send emails, manage your calendar, search the web, send messages, store documents — through a simple command-line interface.

Think of it this way: your AI agent is smart, but it doesn’t have hands. It can think, plan, and draft, but when it needs to actually do something — send that email, check your calendar, look up a contact — it has to cobble together a solution from scratch every time. That’s slow, error-prone, and wastes the capacity you’re paying for.

Mechanical Advantage gives your agent reliable, tested tools for the things it tries to do every day. One setup, and it just works — across every conversation, every session, every task.

But here’s the part that matters most to you: you stay in control of everything.

You approve everything before it happens

This is the core of how Mechanical Advantage works, and it’s worth understanding in detail.

When your AI agent wants to do something that affects the outside world — send an email, create a calendar event, post a message — or wants to store something in its long-term memory, it doesn’t just do it. Instead, the action goes into a review queue. Nothing leaves your accounts — and nothing gets permanently learned — until you say so.

Here’s what that looks like in practice:

Your agent drafts an action. Say it writes an email to your colleague about next week’s meeting.
The action appears in your review queue. You see it in the Mechanical Advantage web app — the full email, including who it’s going to, the subject line, and the body.
You review it. You can read the whole thing, and then:
- Approve it — the email sends exactly as written.
- Edit it — change the wording, fix a detail, add a recipient, then approve.
- Reject it — the email doesn’t send, and you can tell the agent why.
The agent sees your decision. If you approved, great — done. If you rejected or edited, your feedback goes back to the agent so it can do better next time.

This isn’t just for email. Every outbound action — calendar events, Slack messages, Telegram messages, Discord posts, Reddit posts — and every memory mutation (writes and deletes) goes through the same review process. Your agent proposes, you approve. Nothing happens — and nothing gets learned or forgotten — without your say-so.

Security you don’t have to think about

Here’s where it gets genuinely interesting.

Mechanical Advantage has no passwords. None. Not for you, not for anyone. When you first create your account, you get a one-time email link to sign in. From that moment forward, all authentication is done with passkeys — your fingerprint (TouchID), your face (Windows Hello), or a physical security key.

Why does this matter? Because it creates a boundary your AI agent physically cannot cross.

Think about it: if there were a password, the agent could theoretically learn it, guess it, or be tricked into revealing it. With passkeys, there’s nothing to learn. The authentication happens through your device’s hardware — your fingerprint sensor, your face scanner, your physical security key. The agent doesn’t have fingers. It doesn’t have a face. It can’t press a USB key. It is structurally locked out of the approval system.

This means:

Your agent cannot approve its own actions. It can propose an email, but it cannot send it. Only you can.
Your agent cannot access the web app. The login requires your biometric. The agent doesn’t have one.
Your agent cannot change anything about the approval process. Every action requires your biometric approval, and there is no mechanism to change that. The agent is structurally locked out.

This isn’t a software setting that could be misconfigured. It’s a hardware-enforced boundary. The agent is on one side, you’re on the other, and there’s no bridge.

When you say no, the agent gets smarter

Here’s something that turns a safety feature into a genuine productivity gain.

When you reject an action or edit it before approving, your feedback goes back to the agent. The agent sees what it proposed, what you changed, and why you changed it. It proposes storing that lesson in memory — and you approve the memory write, confirming it learned the right thing.

For example:

You reject an email draft and note “Always use a formal greeting with external clients.” The agent proposes a memory entry: “formal greetings for external clients.” You approve it. Next time, the agent opens with “Dear Ms. Chen” instead of “Hey Sarah.”
You edit a calendar event to add your project manager as an attendee. The agent proposes: “Always include the PM on status meeting invites.” You approve it.
You change the tone of a Slack message from casual to professional. The agent proposes adjusting its default for that channel. You approve it.

Each time you give feedback, the agent proposes a lesson and you verify it before it sticks. Over weeks, the rejections drop. The agent starts getting it right the first time — not because someone programmed it to, but because you taught it your preferences and confirmed every lesson.

The review queue isn’t busywork. It’s a teaching loop. And it gets faster over time.

If you run multiple AI agents — maybe one for work, one for personal tasks, one for a side project — they all share the same memory and document storage on your Mechanical Advantage account. When one agent proposes learning that you prefer formal greetings with external clients, you approve it, and every agent on your account knows it too. You teach the lesson once, verify it once.

New agents start with full context from day one. When you spin up an agent on a new machine or for a new task, it doesn’t start from scratch. It inherits everything you’ve approved your other agents to learn — your preferences, your contacts’ communication styles, your project conventions. No re-training. No repeating yourself. And you know exactly what’s in there because you approved every entry.

This also means you don’t need expensive hardware for every machine running an agent. Memory and documents live in Mechanical Advantage, not on local disk. A basic Mac Mini with 16GB of RAM runs agents just fine. And if you ever want to leave, all your data is exportable at any time — documents, memory, everything. It’s all markdown, inherently portable. No lock-in.

Every action, every time

You might be thinking: “Do I really have to approve every single thing?”

Yes. And here’s why that’s a feature, not a limitation.

Every outbound action — every email, every message, every calendar event, every memory change — requires your biometric approval. There are no auto-approve rules. There is no “trust gradient.” There is no mechanism to let actions bypass your review.

This is the strongest safety guarantee we can offer. Other systems let you configure thresholds — “auto-approve low-risk actions” — but that creates a new category of risk: what happens when the threshold is wrong? A misconfigured rule could let an agent send emails to the wrong person or store an incorrect preference without your knowledge.

We eliminate that risk entirely. There is no rule to misconfigure.

The review queue is designed to be fast, not burdensome. You stay in control of every action without it becoming a bottleneck.

And any changes to your account or security settings — connections, passkeys, billing — require your biometric authentication. The agent can’t quietly modify configuration behind your back.

What it cannot do

This matters as much as what it can do.

Mechanical Advantage is non-destructive by design. Your agent literally cannot:

Delete your emails. Not one. Not ever. It can read, send, and archive — but delete is not an option in the system.
Delete calendar events. It can create events and cancel them (with a record), but it cannot remove them from your calendar.
Overwrite your documents. Everything is versioned. If the agent updates a document, the previous version is preserved. Nothing is ever truly lost.
Change memory without approval. All memory mutations — writes and deletes — go through the same review queue as outbound actions. You control what your agents learn and what they forget.
Send anything without approval. Every outbound action goes through your review queue. No exceptions.

These aren’t permission settings you have to configure correctly. They’re structural properties of the system. “The AI deleted my entire inbox” is architecturally impossible — the delete command doesn’t exist.

If you’re wondering whether these safeguards matter in practice, a recent academic study red-teamed autonomous agents that didn’t have them — we wrote about what happened. Agents deleted entire email systems, disclosed medical records to unauthorized people, and broadcast fabricated accusations to dozens of contacts — not because the agents were malicious, but because the tools they were given had no structural limits. These are the exact failure categories that Mechanical Advantage is designed to make impossible.

You can give your agent full access to Mechanical Advantage because full access is safe.

Getting started

Setup takes a few minutes:

Create your account at mechanicaladvantage.ai. You’ll get an email link to sign in, then set up your passkey (fingerprint, face, or security key). That’s your login from now on — no password to remember.
Connect your providers. Link the accounts you want your agent to work with — your Gmail, your iCloud calendar, your Slack workspace, whatever makes sense for you. You choose what to connect; nothing is required.

Install the CLI. Run this in your terminal (or have your agent run it):

curl -fsSL https://raw.githubusercontent.com/mechanical-advantage-ai/ma/main/install.sh | sh

Your agent starts using it. Once the CLI is installed and authenticated, your agent can call ma commands. Outbound actions queue for your review. You approve from the web app with your fingerprint.

That’s it. No complex configuration, no IT department required.

Pricing

Simple, predictable pricing. No free tier, no enterprise tiers, no “call us for a quote.”

Connected accounts — monthly per account:

Provider type	Examples	Price	Includes
Full-suite	Gmail, iCloud, Fastmail, Microsoft	$10/month	Email, calendar, contacts, drive sync, document storage, memory
Messaging	Telegram, Discord, Reddit, Slack	$5/month	Messaging, document storage, memory

Usage-based tools — pay as you go:

Tool	Typical cost per use
Web search	$0.015
Web page fetch	$0.045
Web screenshot	$0.045
Image generation	$0.12
Text-to-speech	$0.45 per 1K characters
Speech-to-text	$0.54 per hour
Memory search	$0.0003

Each connected account includes 10 GB of document storage. Overage is $0.50/GB/month — but markdown files are small, and most accounts never come close to the limit.

A typical setup — one Gmail account and one Slack workspace — runs $15/month plus a few dollars for usage. You can connect multiple accounts: work email, personal email, a couple of messaging bots. You pay for what you use.

Your data stays where it is

Mechanical Advantage doesn’t replace your email provider or your calendar. Gmail is still Gmail. iCloud is still iCloud. Your providers are always the source of truth. Mechanical Advantage gives your agent a structured way to interact with those providers — it doesn’t take ownership of your data.

If you ever stop using Mechanical Advantage, nothing is lost. Your email, calendar, contacts, and messages are exactly where they’ve always been.

Ready to get started?

Create your account at mechanicaladvantage.ai. Setup takes a few minutes. Your agent will thank you!