Amp

Amp is a coding agent and a development environment at the same time, run by an independent research lab for AI agents. That is how the team describes itself on its About page. The idea behind Amp: you give an agent a task, and it keeps working in its own prepared environment without your own machine having to stay open. Anyone looking for an AI assistant for software development that is not tied to a single editor and uses several models side by side will find a product with a distinct point of view here.

What Amp does day to day

According to the documentation, Amp is the same agent with the same threads on every surface. You can start in the browser at ampcode.com without installing anything, in the apps on a Mac, iPhone, or iPad, or in the terminal with the command line. The apps notify you when a thread is waiting for your next message, and there is also dictation and sharing from other apps. The command line runs on your own machine with local files and tools, and can also be started without an interface on servers or in CI runs. A thread is a saved, shareable conversation with the agent.

Orbs: a cloud machine per thread

The distinctive part is what the vendor calls orbs. According to the vendor, an orb is a machine that Amp creates for each thread, with the project's code, tools, and plugins already set up. You send a prompt, close the laptop, and the agent keeps working. If you like, you can mirror changes into your local copy with one command. Portals let you open a development server inside the orb and comment directly on the page. With automations, a thread can even wake itself up again, for example to check metrics after a rollout. According to the security document, an orb pauses when idle, is stored encrypted, and is deleted when you delete the thread.

Models, MCP, and extensions

Amp relies on several models and picks a suitable one for each task. In its security document the vendor names Anthropic, OpenAI, Google, Meta, and other hosting services as inference providers. You can bring your own subscriptions and keys, such as a ChatGPT or Claude subscription. Amp can be extended through the Model Context Protocol, skills, plugins, and an AGENTS.md file in the project. Many MCP servers and skills from the Skill Road catalog therefore also fit Amp.

Plans and cost

The pricing page lists four tiers. Hobby is free: you pay for orb time on a usage basis or use your own runners, and model access through your own subscriptions or keys is possible without an Amp markup. The Individual plan is a monthly plan with a fixed allowance of orb minutes and unlimited repositories. Teams add shared threads, permissions, shared portals, plugins, and skills, plus SSO and pooled billing. Enterprise adds, among other things, SCIM, audit logs, network and MCP controls, and individual contract terms. Current amounts are on the official pricing page.

Privacy and security

According to the vendor, Amp is SOC 2 Type II certified and is penetration-tested externally every year. The infrastructure runs on Google Cloud in the US; per the vendor, the client does not read, store, or index the entire codebase but sends snippets as context. Secrets are redacted automatically, though the vendor explicitly says this is best effort. According to the vendor, models are trained on your data only after an explicit opt-in, and never on Enterprise. Self-hosting is not offered. If you work under strict requirements, read the model providers' retention rules and the security document first.

Where it sits

Unlike open-source agents such as goose or OpenHands, Amp is a hosted service that you do not operate yourself. In return, orbs, apps, and team features come from a single design. Compared with editor assistants such as Zed or Cline, the emphasis is on tasks you hand off and review later.

Who Amp fits

Amp suits developers and teams who want to hand tasks to agents in parallel, check in while away, and review the results in the repository. If you need full control over operations or self-hosting, an open-source alternative serves you better. A good way to start is the free Hobby plan with a small repository and a clearly scoped task, such as a test for a single function.