Amp

• Starting several agents in separate orbs from a phone or browser, then reviewing their diffs and proof of work later
• Sending a screen recording of a UI bug and having the agent reproduce and fix it in an orb
• Asking the agent to start a dev server and share a portal link so a teammate can review a preview
• Turning a Mac mini or GPU box into a shared runner so the team can spawn agents that need that hardware
• Running `amp -x` in scripts or CI to execute a single prompt and exit

Amp's distinguishing idea is that every agent should get its own machine. Orbs, portals, and runners make parallel background work practical, and the September 2026 pricing change removes the subscription barrier for anyone who already pays for model access. Multiplayer threads, shared portals, and shared runners are built for teams working on the same agent threads. The cost is stability: Amp says plainly that it drops features and ignores backward compatibility, modes are fixed per thread, and the native apps are still in beta. It suits developers who want to be on the frontier of agent workflows more than organizations that want a fixed, procurement-friendly tool.

About Amp

Amp was built inside Sourcegraph and spun out in December 2025 as Amp Inc., founded by Sourcegraph co-founders Quinn Slack and Beyang Liu. It is a separate product from Sourcegraph Cody. Its core unit is the thread, one conversation with the agent that opens the same way on ampcode.com, in the CLI, and in the macOS and iOS apps. Each thread can run in an orb, a fresh isolated Linux machine that Amp sets up with your code, tools, and plugins. The agent can install dependencies, run services, drive a browser or an interactive desktop, and prove its changes work while your own computer is off. Orbs are billed by the minute and pause 5 minutes after the agent stops or 20 minutes after your last interaction, whichever comes later. Portals expose web servers inside an orb as authenticated URLs you can open, annotate, and share. You can also run agents on your own hardware through runners, which are free to use, can serve many directories and create Git worktrees, and since September 2026 can be shared with a workspace. The Dial replaces model picking with four modes, low, medium, high, and ultra. Each mode bundles a model, reasoning effort, tools, and an Oracle for second opinions, across models such as GPT-6 Astra, GPT-5.6, and Claude Fable 5.1. Specialist subagents (Search, Oracle, Librarian, Read Thread) handle focused work, and Puck is an assistant that starts, monitors, and coordinates agents, including from Slack and by voice. Teams get shared threads, multiplayer threads where colleagues prompt the same agent, shared portals, and pooled credits. Since September 13, 2026, Amp has been free when you bring your own compute and model access: a ChatGPT subscription, other subscriptions, or your own API keys, with no Amp token fees. You pay for orbs, pay-as-you-go or through the Megawatt ($20/month, 45,000 orb minutes) and Gigawatt ($200/month) tiers, and inference bought through Amp is charged at provider API prices with no markup for individuals and non-enterprise workspaces. The tradeoff is Amp's stated philosophy: it ships fast, removes features it does not use, and keeps no backward compatibility, and a thread's mode cannot be changed once it starts.