An AI DevOps agent that deploys, checks, and asks first

A customer has been offline since last night's migration. The crew reads their account in your billing database, signs in to the router with a certificate that expires in five minutes, reads the line's log, and finds the cause: the line still points at a plan the migration removed. Then it stops, because the fix is a change, and it asks you first. You tap Approve on your phone. That's crew bot as an AI DevOps agent: it does the reading and the legwork on your own servers, routers and databases, and you say yes to whatever changes them.

What the AI DevOps agent does

Signs in to your servers and routers over SSH, with five-minute certificates.
Works in your databases. Postgres and MySQL are built in. Ask a question in plain words and get the answer from your own data.
Finds what's wrong and tells you the fix, or makes it once you approve.
Deploys to your own server. It builds the release, starts it, checks it answers, then switches over, on your domain with HTTPS, and keeps the release before so you can go back.
Works on a schedule. Routines run when you set them, and webhooks let your other systems hand it work.
Reaches the rest of your stack: GitHub, web apps you sign it in to, and any app with an MCP connector.
Uses a computer of its own, and yours when you say so: its browser, its screen and a terminal.

If you want the background first (what an AI DevOps agent should and shouldn't be allowed to do), read our guide to AI DevOps agents and what they can safely do on your servers.

It asks first

The crew asks before it sends an email, changes an account or runs anything that writes, unless you've told it to go ahead. In practice a job reads first, as much as it needs, and then stops at the one step that changes something:

Approve from anywhere. The request comes to your phone with buttons. You can also talk to the Lead on Telegram, by text, voice note or photo; only your own Telegram account is listened to.
Rules for a team. Add team members and set approval rules.
A budget on every job. Fifty US cents by default. If a job needs more, it stops and asks.
Every step on record. Each job keeps what it read, clicked, ran and found, so you can check its work.

Why this matters, and how to set the line between "do it" and "ask me": human-in-the-loop AI.

Your logins stay out of reach

A vault the AI can't read. Logins, keys and tokens for your servers, routers and databases are encrypted (AES-256-GCM) and kept out of the crew's computer. No model ever sees them.
Five-minute certificates for SSH, so there's no long-lived key for the crew to leak.
Your passwords stay yours. You type passwords into websites yourself.
Kept apart. Connections use HTTPS or SSH, and each workspace's data is kept apart from every other workspace's.

An agent that reads logs and web pages will meet text written to manipulate it. Here's how that works and what protects you: prompt injection, and how to stay safe.

A fix, step by step

Here's the offline customer from the top of this page, as the job ran:

Support desk: a customer at Block C has been offline since the migration

› read the customer's account · read only
  plan home-20M · edge-router-2 · line 4471
› ssh edge-router-2 · 5-minute certificate
› show line 4471 and its last log
  rejected 6× since 14:02
  profile old-10M not found
✓ Found it: the line still points at a plan the migration removed

Needs your OK: move line 4471 to plan home-20M on edge-router-2
One change. Everything before it only read.

After the approval, the crew moved the line to the new plan and told the Support desk what was done. The router's login stayed in the vault the whole time.

A deploy, step by step

When the crew builds something, it can put it live on your own server. Here's release 14 of a shop's API:

A deploy with checks: build, ship to your own server, confirm the app answers, then go live, or roll back.

Uploaded release 14 from the job's files.
Built it and started it as a service of its own.
Checked it answers: 200, in 140 ms.
Made it current, on your domain over HTTPS.

A release that doesn't answer never goes live, and the one before keeps running. If something's wrong later, Roll back to release 13 puts the previous one back.

Your database, in plain words

Ask the question you'd ask a colleague:

You asked: how many repeat customers did we have in September?

› select count(distinct customer_id) from orders
  where placed_at >= '2026-09-01' and customer_id in (
    select customer_id from orders group by 1 having count(*) > 1)

✓ 214 repeat customers, up from 187 in August

Reading happens freely; writing waits for you. Marking an order paid in your shop admin, for example, happens after you approve it.

Where it works

Its own workspace computer. On paid plans, each workspace gets a small cloud computer for the crew, a Firecracker microVM that sleeps when idle:

Plan	Workspace computer
Free	None
Start, $19 a month	2 CPUs, 2 GB memory, 5 GB storage
Growth, $49 a month	2 CPUs, 4 GB memory, 10 GB storage
Scale, $149 a month	4 CPUs, 8 GB memory, 25 GB storage

More storage is $2 a month per 10 GB.

Your own computer, when you say so. With the crew bot app for Windows or Mac, the crew can use that computer's browser, screen and terminal, and your own AI agents. You can lend it the coding agents you already use, such as Claude Code or Codex, and what runs on them isn't counted against your plan. On a Mac you watch every move live and can take over at any time.

Your servers. Working on your servers, routers and databases is on every plan, Free included.

A crew, not one bot

The Lead plans a job and splits it between specialists who work at the same time. A skeptic attacks every finding before it reaches you. And the crew remembers how you like things done, so you explain a job once.

It grew out of running real operations: billing, networks and client support. Two jobs it has done:

Fixed a client's connection, live in a meeting. It read the issue in WhatsApp, found the client's account, logs and devices, opened the router's management app, signed in and fixed the fault, then told the colleague who'd asked.
Turned a client's issues into a workplan. Each issue became a task. It checked the codebase for where each one stood and drew up the plan.

Good to know

The workspace computer is modest. Up to 4 CPUs, 8 GB and 25 GB on Scale, and none on Free. For heavy builds, point the crew at a bigger server of your own.
Models from several AI labs, picked per job, rather than one company's own model. Or bring your own keys.
Apps: in your browser, and on Mac and Windows today. Linux, iPhone and Android are coming.

See how crew bot compares with other always-on agents, including where they're ahead: compare.

Pricing

Servers, databases, email, browser and your apps are included on every plan, as are routines, webhooks and approvals on your phone. Paid plans add the workspace computer and more AI to work with, measured at what the AI providers charge with no markup. Plans are Free, then $19, $49 and $149 a month per workspace. See pricing.

FAQ

Can an AI DevOps agent deploy without asking?

crew bot asks before anything that writes, including putting a release live, unless you've told it to go ahead for that kind of job. Every job also has a budget, and every step is on record.

Does the AI see my server passwords?

No. Logins, keys and tokens are encrypted, kept in a vault out of the crew's computer, and never shown to a model. SSH uses five-minute certificates.

Which servers and databases does it work with?

Servers and routers you reach over SSH, Postgres and MySQL databases, GitHub, web apps you sign it in to, and any app with an MCP connector.

What happens if a deploy fails?

It doesn't go live. The crew checks that a new release answers before switching over; if it doesn't, the release before keeps running. You can roll back to the previous release at any time.

Do I need a paid plan to work on my servers?

No. Working on your servers, routers and databases is on every plan, Free included. Paid plans add a workspace computer of the crew's own and more AI to work with.