Amp review

Amp is a coding agent and development environment from the founders of Sourcegraph ^25,27. Sourcegraph launched it in July 2025 and spun it out as its own company in December 2025, with Quinn Slack as CEO ^26,27. Amp's terms now name Amp Frontier Corporation ¹⁹, and it calls itself an independent agent research lab that also runs Amp Labs, which embeds engineers inside companies such as Westpac ^18,24.

Amp is deliberately opinionated. It chooses models for you through four effort modes, drops features it stops using, and pushes work onto remote machines it calls orbs ^1,11. Agents run in orbs, on your own runners, or locally in the CLI, and you can follow them from the web, the macOS and iOS apps, or Slack through the Puck assistant ^1,10,13.

You start a thread on the web, in the CLI or by asking Puck. The Dial sets effort: low, medium, high or ultra, each with its own model, prompt and tools, fixed for the life of the thread ¹¹. Medium runs on Claude Opus 5.5 and ultra on Claude Fable 5.1, high uses GPT-6 Astra, and an oracle model gives second opinions while subagents handle search, library research and delegated tasks ^8,12,20. In an orb, the agent gets a fresh machine with your repository cloned, installs dependencies, runs the app and tests, and shares previews through portals ^5,20. Each user can start 20 orbs in a burst, then one every five minutes ⁵.

When the work is done, changed files appear in the Changes pane. Ship commits, rebases, runs the full test suite and pushes to the base branch by default; Push to Branch opens a pull request instead, and Custom Ship runs your own prompt ⁶. Experimental GitHub webhooks can continue a thread on new PR comments or spawn an orb for a new issue ⁷. Locally, the CLI connects to VS Code, Cursor, Windsurf, Zed and Neovim ¹⁰. Amp runs tools without asking; plugins can allow, reject or rewrite each tool call, and admins can push them to the whole workspace ^8,9.

Amp is free if you bring your own model access. The Hobby tier uses a linked ChatGPT subscription or your own API keys with no Amp token fees, and agents run free on your own runners ^2,21. You pay Amp for orbs, billed per minute from $0.08 an hour for a1.tiny to $2.13 for a1.3xlarge, and for any model usage you buy through Amp at provider list price with no markup ^3,4. Megawatt costs $20 a month and includes $20 of credits plus 45,000 orb minutes; Gigawatt costs $200 and includes $200 of credits plus 1,000 hours of a1.xxlarge orb time ². Included usage resets monthly and does not roll over, and bought credits expire after 12 months ³. A developer on Megawatt who uses $60 of models through Amp in a month pays $20 + $40 = $60, unless a ChatGPT subscription covers the OpenAI share. Teams cost nothing extra; Enterprise is quote-only ².

Amp holds a SOC 2 Type II report and pentests its whole stack every year ¹⁴. It runs on Google Cloud in the US, and its model providers, including Anthropic, OpenAI, Google, Amazon Bedrock, Baseten and Fireworks, serve from US servers ¹⁴. Traffic is encrypted with TLS 1.2+ and stored data with AES-256, and secrets are redacted before they reach threads or providers ¹⁴. Neither Amp nor its subprocessors train on your data unless you opt in, and Enterprise workspaces cannot opt in at all ¹⁴. Minimal Data Retention limits what model providers keep, but it is not zero retention: safety, legal and some model-specific holds of up to 30 days apply ¹⁵. Enterprise adds SSO, SCIM, authentication audit logs and MCP registry allow-lists; there is no self-hosted option ^14,16.

Amp has no independent rating profile to cite and publishes no customer case studies. Its about page quotes individual developers, including the creators of pi and OpenClaw and OpenAI's Codex lead ¹⁸. The only named enterprise relationship is an Amp Labs partnership with Westpac ²⁴. Tessl reported that Amp claimed profitability at spin-out ²⁶. An independent review by Ry Walker calls cost the most common community complaint and points to features removed in 2026 ²⁸.