Amp

Amp is a coding agent that started at Sourcegraph and announced in December 2025 that it was becoming its own company. It runs as a CLI that connects to your editor, has apps for desktop and mobile, and can farm work out to cloud sandboxes it calls Orbs, either hosted by Amp or on your own runners.

Its distinctive feature is billing: you can use tokens from subscriptions you already pay for, such as ChatGPT or Claude. On November 12, 2026 it replaced its low/medium/high/ultra “Dial” with a per-thread model and effort selector.

Who it’s for

Developers who already pay for one or more AI subscriptions and want a single agent that can use any of them, and people who want to run many agents in parallel in the cloud.

What stands out

• Bring any subscription. Use ChatGPT, Claude or SuperGrok plans, or API keys, with no Amp token fees on Hobby and Megawatt.
• Huge model list. About 115 models, picked per thread with an effort setting, and the selector shows which subscription or key will be billed.
• Orbs. Each thread can get its own cloud machine, and since October 29 Orbs can start a Mac to build and test iOS and Mac apps.
• Puck. Launches and coordinates agents for you.
• Threads on every device. Start a thread on your phone, check it on your Mac and follow up in the terminal.

Drawbacks and things to watch

• Switching models costs extra. Changing model mid-thread re-reads the whole thread without the cache, so the next message costs more.
• Old Dial settings are gone. They didn’t carry over to the new selector; new threads start on the Staff Pick model.
• Enterprise pays for self-hosted Orbs. Since November 13 they’re billed per minute on Enterprise, though still free on other plans.
• No local models. Custom endpoints must be reachable from Amp’s servers.

Privacy and security

Amp is SOC 2 Type II certified with annual penetration tests. It stores threads in Google Cloud, which can include code snippets or whole files, but says it doesn’t store or index your entire codebase. It doesn’t train on your data unless you opt in, and on Enterprise training can never be turned on. Deleted threads are removed within 30 days.

Getting started

On Mac, Linux or WSL, install the CLI with curl -fsSL https://ampcode.com/install.sh | bash, then connect it to VS Code, Cursor, Windsurf, Zed or Neovim as the CLI docs describe, or download the desktop or mobile app.

Plans and prices

Amp FAQ

Is Amp free?

Yes, if you bring your own models and compute. The Hobby plan includes all features, unlimited self-hosted Orbs and no Amp token fees when you use your own subscriptions or API keys. Amp-hosted Orbs are pay-as-you-go, and the $20 Megawatt plan includes 45,000 hosted Orb minutes.

Can I use my ChatGPT or Claude subscription with Amp?

Yes. Amp connects ChatGPT, Claude and SuperGrok subscriptions through browser sign-in, and also accepts provider API keys billed to your own account.

What are Amp Orbs?

Per-thread cloud machines with your code, tools and plugins set up, hosted by Amp or on your own runners. Hosted Orbs cost from $0.08 to $2.13 an hour depending on size.

Does Amp train on my code?

Not unless you opt in, and on Enterprise training can never be enabled. Amp stores your threads, which can include code, and says it doesn't store or index your entire codebase. It's SOC 2 Type II certified.

Can Amp use local models?

Not directly. Custom OpenAI- and Anthropic-compatible endpoints are supported, but requests come from Amp's servers, so the endpoint must be reachable from the internet.

Latest changes

Sources

• Amp pricing
• Amp news: Bye Bye, Dial
• Amp news
• Amp manual
• Amp docs: Orb sizes and costs
• Amp models
• Amp security
• Amp docs: model routing

Facts last checked November 14, 2026.