Amp

Amp is a coding agent whose big idea is the Orb: every thread gets its own remote machine with your code and tools, so you can send a prompt, close your laptop and let the agent keep working. It runs on the web, in apps and in a CLI. Hobby is free, Individual is $20 a month. Checked 31 October 2026.

At a glance

What Amp is

Amp's site calls it "the frontier agent and development environment that lets you build from anywhere." The manual says it is a coding agent and development environment, with four things to know.

1. Multi-model. Amp uses different models for different tasks, including fast ones, and the manual names GPT-5.6 and Claude Fable 5.1 among them. Amp's models page lists 109 models, so treat the manual's names as examples, not the full line-up.
2. Orbs. "Machines Amp creates for each thread, with your code, tools, and plugins preconfigured." The manual's summary: "Send a prompt, close your laptop, and the agent keeps working."
3. Everywhere. The web, macOS, iOS and Android apps, and the terminal CLI. The apps notify you when a thread is ready for your next message.
4. Threads. You can save and share your interactions with Amp.

The effect is a cloud agent with a developer-friendly shell around it. Compare Devin, which also works in a remote environment, and the cloud agents in Cursor and GitHub Copilot.

Orbs

The Orb is what makes Amp different. From the home page:

• A fresh machine per thread, with your code, plugins and tools. Orbs can wake on schedules, webhooks or events such as a new Linear issue.
• Portals. Run the app inside an Orb and preview the agent's changes live with reloading.
• Review and sync. Review diffs and browse files from any device. amp sync mirrors an Orb's changes to your local checkout.
• Multiplayer. Invite teammates into the same Orb, agent and running app.
• Terminal. A shell inside the Orb that you share with the agent.
• Self-hosted orbs and macOS satellites. The news section announces running orbs on your own infrastructure, and starting a Mac to build iOS and Mac apps.

An Orb is a virtual machine in the cloud, so it can run your dev server, tests and tools without touching your laptop. That is the difference from a local terminal agent such as Claude Code or OpenCode, where the work happens on your machine and stops when you close it.

The tradeoff is where the code and secrets live. An Orb has your repository and any credentials you give it. For sensitive work, the self-hosted orb option, which keeps the machine on infrastructure you run, is the one to ask about. See enterprise AI coding agent security.

The Dial

Instead of a long model list, Amp offers a setting called the Dial with low, medium and high positions. You pick how much capability and cost a task deserves, and Amp chooses the model. The site says linking a ChatGPT subscription makes the Dial's settings run on that subscription instead of Amp credits, and the link applies to the main agent, subagents, the thread reader and code review.

That is a clean idea: you stop choosing models and start choosing effort. It also hides which model ran, so if you need to know for compliance, check what Amp records for a thread. We found no statement on whether it shows which model handled each step.

CLI

The CLI runs agents locally or in CI, per the site:

• amp runs locally.
• amp sync mirrors an Orb's changes locally.
• amp --no-tui runs agents on servers and in CI without the terminal interface.

The install options listed are a curl script, a Windows route and Homebrew. We could not read the exact commands, because the manual pages we reached were behind sign-in. The Homebrew tap lives in the ampcode GitHub organisation.

The manual's quickstart has six steps: sign up, optionally sign in with ChatGPT for OpenAI models, add a project from a Git repository (or start without one), run an agent in an Orb, install the apps, and use the CLI. Sign-in options on the page include email, ChatGPT, passkey, Google and Apple.

Plans and prices

From ampcode.com/pricing on 31 October 2026:

Notes:

• The page says all tiers include no Amp token fees or limits and let you use tokens from other sources such as ChatGPT or Claude subscriptions. The home page says "Your Amp tier still includes agent usage and orb time" and that Ultra continues to use Amp credits. The pricing page lists Megawatt and Gigawatt and does not list an Ultra tier, and it does not explain how Amp credits relate to them.
• Both Megawatt and Gigawatt appear under Individual with one price and one minute allotment. The page gives no separate figures for each.
• No credit amounts, per-minute rates or token prices are published on the pages we read.

The cost model therefore has three parts: the tier fee, orb minutes (included, then pay as you go), and model tokens (your own key or subscription, or Amp credits). Add them before comparing with a flat plan such as GitHub Copilot Pro at $10 or Cursor.

What a day might look like

An illustration of the Orb workflow, built from the features the site lists, not a recording of a real session.

1. Morning. You open the web app, pick a project from your Git repository and start a thread: "Add pagination to the orders list, with tests."
2. The Orb starts. Amp creates a fresh machine with your code and tools. You close the laptop and leave for a meeting.
3. A notification arrives on your phone: the thread is ready for your next message. You review the diff in the app and ask for one change.
4. Event trigger. Overnight, a new Linear issue wakes another Orb that you configured to triage bugs.
5. At your desk, you run amp sync to pull the Orb's changes into your local checkout and run the app yourself.
6. Teammate help. Someone joins the same Orb to check the running app in the Portal.

The pattern matters more than the product: long jobs run remotely, you review from wherever you are, and local work is a sync away. That is the same direction Cursor and Claude Code have taken with their cloud sessions.

Questions to ask during a trial

Because pricing and terms are layered, use the free tier to answer these before you pay.

• How many orb minutes does a typical task consume on Hobby?
• What does it cost to run an Orb that waits for an event all week?
• Which models handled my threads at each Dial setting, and can I see that?
• Where do secrets live inside an Orb, and who on the team can read them?
• Can I delete a thread and its Orb data, and how fast?
• What does the code review feature cost against my subscription?

Write the answers down. They are the numbers you need to compare Amp with Devin, which sells a daily and weekly quota instead of minutes.

What it does well

• Work continues without you. Orbs run when your laptop is shut.
• Everywhere access. Web, mobile and CLI, with notifications.
• Teams features, including multiplayer in the same Orb, shared threads and SSO.
• Self-hosted option for the compute.
• Bring your own tokens. A ChatGPT subscription or your own keys can pay for model use.
• Event triggers. Orbs wake on schedules, webhooks and issues.

Limits and cautions

• Closed product. No source for the agent.
• Hidden pages. Parts of the manual sit behind sign-in, so some details need a trial to confirm.
• Naming churn. Tier names (Megawatt, Gigawatt, Ultra) and model names appear inconsistently across pages.
• Remote by default. Your repository sits on a remote machine, which matters for confidential code.
• Cost layering. Orb minutes plus tokens plus tier is harder to forecast than one price.
• Experimental features. The home page describes Recap as experimental and says it makes extra inference requests that incur extra cost.
• Data terms. We did not read the data-use terms. The pricing page's FAQ has a question on data retention and security commitments, so start there.

Amp versus the others

For the whole field see AI coding agents.

Who should pick it

Amp suits developers and teams who want agents that keep working while they are away, who like reviewing a diff on a phone, and who want to pay for model access with their own keys or subscriptions. It suits less well if you want everything on your own machine, if you need an open source agent, or if you want one flat monthly price.

How Amp compares on the unit of cost

Most agents sell either a flat plan or a usage allowance. Amp adds a third unit, Orb minutes, and lets model tokens come from your own subscription or key. That makes the comparison with a flat plan indirect: add the tier fee, any minutes beyond the included 45,000 on Individual, and the model cost you pay elsewhere. For a developer who already pays for a model subscription, the marginal cost may be small. For one who does not, Amp credits are the part to measure. Run a week on Hobby and total all three before deciding.

Our take

This is an editorial reading of the Amp site and pricing page, not a hands-on test.

Amp's design bet is that the agent should live on a machine of its own, with your laptop as one of several windows onto it. That solves a real irritation, the agent that stops when you close the lid, and it makes team features such as shared Orbs natural. The pricing page is also friendly to people who already pay a model vendor, since tokens can come from your own subscription or key.

Our reservations are about clarity. Tier and model names move between pages, parts of the manual are behind sign-in, and orb minutes add a unit that needs watching. Try Hobby with a non-sensitive repository, note how many orb minutes a typical task uses, and compare the total with what you pay today. If you handle confidential code, ask about self-hosted orbs before sending any of it.

Checklist

• Start on Hobby with a non-sensitive repository
• Decide where model tokens come from: your key, a ChatGPT link, or Amp credits
• Note the orb minutes used by three typical tasks
• Ask about self-hosted orbs for confidential code
• Try amp --no-tui in CI only after reading its manual
• Confirm data terms for your tier
• Re-check pricing on ampcode.com before subscribing