A coding agent that came out of Sourcegraph. It runs both from the terminal (CLI) and inside VS Code-family editors, and once you give it a task it autonomously reads the codebase, edits files, and runs commands until the job is done. Its guiding principle is to get the most out of the latest models: rather than preserving backward compatibility and legacy features, it adopts whatever model and setup is smartest at the moment. After its May 2025 debut, a large architectural rebuild in May 2026 brought it to its current form, in which each thread can spin up its own cloud execution environment (Orbs). Work threads can be shared by URL, so you can show your team exactly what the agent thought and changed.
Key Features
- Works from both the CLI and the editor: In addition to the terminal CLI, it is available as an extension for VS Code and its derivatives (Cursor, Windsurf, and others). Keybindings, a non-interactive execute mode, and streaming JSON output make it easy to embed in scripts and pipelines
- Orbs (cloud execution environments): Each thread can launch a cloud machine preloaded with your code, tools, and plugins, and run the agent there. You can keep long-running tasks going without tying up your local machine, and run several tasks in parallel
- Multi-model setup: It is not locked to a single model; it mixes frontier models such as GPT-5.6 and Claude Fable 5 with faster, lighter ones. Which modes (reasoning intensity) you can use depends on your plan
- Thread sharing: Conversations with the agent can be saved and shared by link, which is useful for review, handoff, and spreading know-how inside a team
- Extensible via MCP: It supports the Model Context Protocol, so external services and in-house tools can be connected as tools for the agent
- Amp Free as a way in: There is a free entry point for trying it before paying. It used to be ad-supported and no longer shows ads, but capacity is limited, so check the official page for the current situation
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Amp Free | $0 | A free way to try it. Terms change, so check the official page |
| Megawatt | $20/month | $20 of included agent usage, 750 hours of orbs, low / medium modes (high mode when linked with a ChatGPT subscription) |
| Gigawatt | $200/month | $200 of included agent usage, 1,000 hours of xxlarge orbs, all modes including high and ultra |
| Unconstrained | Usage-based | Pay for what you use, priced per token for models and orbs. You can bring your own inference keys |
| Enterprise | Contact sales | SSO, directory sync, admin controls, minimal data retention, dedicated support |
| Students and teachers | $10/month | Requires proof of status |
Pricing is current as of August 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- The same agent works from both the terminal and the editor, so you can adopt it without changing how you work
- Orbs let you run long tasks without occupying your local machine, and make parallel work easy
- Thread sharing turns the agent’s process into something the whole team can learn from
- Because it continuously adopts the newest models, generational improvements show up directly in daily use
- A usage-based plan means you can start without committing to a fixed monthly fee
⚠️ Cons
- It deliberately drops backward compatibility, so features and behavior change quickly. That is a poor fit for locked-down, standardized workflows
- Available models change by mode and plan, and the pricing structure (included usage, orb hours, per-token rates) takes some effort to predict
- The terms of the free tier have shifted over time, so you cannot plan on it staying the same
- Japanese-language documentation and write-ups are scarce compared with English, so reading the official docs is essentially required
- Since you are delegating execution to an agent, reviewing and testing its changes remains as necessary as ever
Comparison with Similar Services
| Criteria | Amp | Claude Code | Cursor | GitHub Copilot |
|---|---|---|---|---|
| Provider | Sourcegraph lineage | Anthropic | Anysphere | GitHub (Microsoft) |
| Where you use it | CLI + VS Code-family extension | Mainly CLI | Dedicated editor | Editor extension (VS Code and others) |
| Models | Mixes several frontier models | Claude family | Choice of several models | Choice of several models |
| Cloud execution | Orbs (per-thread cloud environment) | Supports cloud execution | Supports cloud agents | Supports cloud coding agents |
| Sharing | Threads shared by URL | Limited | Limited | Via pull requests |
| Entry pricing | Free tier + from $20/month + usage-based | From $20/month | Free tier + from $20/month | Free tier + from $10/month |
Who Is It For
- Developers who work mainly in the terminal but also want the agent inside their editor
- People who want to hand long refactors or investigations to an agent without blocking their local machine
- Teams that want to share the agent’s working process for review and training
- Anyone who wants to start AI-assisted coding on usage-based pricing rather than a fixed monthly fee
- Those willing to accept fast-moving changes in exchange for always using the newest models
Summary
Amp combines an autonomous coding agent usable from both the CLI and the editor with per-thread cloud execution environments and sharing features. Its policy of prioritizing the latest models and aggressively cutting legacy behavior suits experimentation and individual or small-team development, while teams that need a fixed, long-lived workflow may find the pace of change burdensome. A practical approach is to start small with the free entry point or usage-based pricing, and consider a monthly plan once your usage pattern settles.