Amp Code: install, the Dial, orbs, AGENTS.md, skills, MCP and pricing

What is Amp Code?

Amp is a coding agent that runs from the web, desktop and mobile apps and a terminal CLI, and can work on remote machines while your laptop is closed. You pick how hard it should work, give it a task, and it reads code, runs commands and edits files. The Amp manual describes it as "a coding agent and development environment built for the frontier."

People search for "Amp Code" because the product lives at ampcode.com. It started inside Sourcegraph, but on 2 December 2025 Amp announced: "Amp is becoming a separate company. We're spinning out of Sourcegraph to become an independent research lab." The company is Amp Frontier Corporation.

Key ideas in today's Amp:

• The Dial. You choose a mode by difficulty, not by model, and Amp routes to the model it picked for that level.
• Orbs. Remote machines per thread, so an agent keeps working after you close your laptop.
• Threads. Every conversation has a URL you can reopen on any surface, share or reference from another thread.
• Open conventions. AGENTS.md, standard SKILL.md folders and MCP servers.

This guide follows the Amp docs as of 26 October 2026.

How do you install Amp?

The web app at ampcode.com needs no install. For the terminal, the CLI docs give this installer for macOS, Linux and WSL:

curl -fsSL https://ampcode.com/install.sh | bash
amp login
amp

amp login adds an account, and amp starts the interactive CLI. The docs also show Windows and Homebrew options, and if your workspace has its own install instructions you can see them after signing in.

The install page offers a macOS app (Apple silicon, macOS 26 or later), an iPhone and iPad beta through TestFlight, and an experimental Android app.

To use Amp beside your editor, the CLI docs say "Amp can connect to VS Code, Cursor, Windsurf, Zed, and Neovim." Open your editor's terminal, run amp, press Ctrl+O and select ide connect. Neovim also needs the Amp Neovim plugin.

How does the Dial choose a mode and model?

The Dial docs advise choosing "a mode based on how much the agent needs to figure out, not on which model happens to be behind it." There are four settings:

Amp changes the model behind each mode over time, so check the modes page for the current routing. In the CLI, press Ctrl+S before your first message to turn the Dial, or press Ctrl+O and type "mode." A thread keeps the mode you chose when you sent its first message, so start a new thread to switch.

Amp also uses subagents. The models and subagents page lists Task, which "carries out delegated work end to end and reports back," and Read Thread, which "reads and summarizes other Amp threads."

What are orbs and threads in Amp?

The orbs docs define orbs as "remote machines where Amp agents work without using your computer." Each orb thread gets its own environment with your code, plugins and tools. "An agent in an orb keeps working while your laptop is closed," and an orb sleeps a few minutes after you stop interacting, costing nothing until your next prompt wakes it. In the CLI, amp -ox "<prompt>" starts an execute-mode thread in an orb, or press Ctrl+O and select "thread: new in orb".

A thread is one conversation: your prompts, the agent's replies, its tool calls and the files it changed. Visibility can be Private, Workspace, Group (enterprise only) or Unlisted, meaning anyone with the link. When a thread gets long, ask the agent to "Handoff and ..." to start a fresh thread that carries over the context that matters. To reuse earlier work, paste a thread URL or type @T-... in a prompt.

For scripts and CI, execute mode runs one prompt, prints the final message and exits:

amp -x "what files in this folder are markdown files? Print only the filenames."
echo "what package manager is used here?" | amp -x

How does Amp use AGENTS.md?

The AGENTS.md docs say Amp reads AGENTS.md from:

• The current working directory or editor workspace roots, plus parent directories up to $HOME
• Subdirectories, when the agent reads a file in that subtree
• $HOME/.config/amp/AGENTS.md and $HOME/.config/AGENTS.md, always, if they exist
• System-wide paths such as /etc/ampcode/AGENTS.md

If a directory has no AGENTS.md but has AGENT.md or CLAUDE.md, Amp includes that file instead, so a repo set up for Claude Code works without changes.

You can pull in other files by @-mentioning them, including globs like @doc/*.md. A mentioned file with globs frontmatter loads only after Amp reads a matching file, which keeps framework-specific rules out of context until needed:

---
globs: ["**/*.tsx"]
---

Use function components and keep each component in its own file.

How do skills and MCP servers work in Amp?

Skills. Amp uses standard Agent Skills folders with name and description frontmatter. The skills docs list the search order, where the first skill with a given name wins. It includes:

• ~/.config/agents/skills/, ~/.agents/skills/ and ~/.config/amp/skills/
• .agents/skills/ in the project and its parents
• .claude/skills/ in those directories, ~/.claude/skills/ and ~/.claude/plugins/cache/
• Extra paths in amp.skills.path, your personal and workspace skills repositories, and official Amp skills

Set amp.skills.disableClaudeCodeSkills to skip the Claude Code locations. The model sees every skill's name and description and reads the rest of SKILL.md only when it uses the skill. To install one:

amp skill add <github-repo-or-path>
amp skill add <github-repo-or-path> --global
amp skills list

Project skills land in .agents/skills/ so you can commit them, and --global installs under ~/.config/agents/skills/.

MCP. The MCP docs store local servers under amp.mcpServers in ~/.config/amp/settings.json or a workspace .amp/settings.json. Workspace servers need approval with amp mcp approve <name> before they run.

amp mcp add context7 -- npx -y @upstash/context7-mcp
amp mcp remote --personal add Linear https://mcp.linear.app/mcp

Remote definitions are stored by ampcode.com, and many remote servers sign in through OAuth in your browser. The docs recommend bundling MCP servers inside skills with a sibling mcp.json, because "tools stay hidden until the skill loads." That keeps the tool list small. The MCP servers roundup has servers worth trying.

How much does Amp cost?

As of October 2026, Amp's pricing page lists:

The pricing docs say paid credits can be added without a paid tier, and for individuals and non-enterprise workspaces Amp adds no markup to provider API prices. Purchased credits expire twelve months after purchase. The Dial docs add that if you sign in with ChatGPT, OpenAI model usage bills to your subscription, and the ChatGPT Only preset pins the main agent of low, medium and high to OpenAI models. Ultra keeps Amp's routing and is not covered.

How does Amp compare with Claude Code and Codex?

Amp's skill search covers both .claude/skills/ and .agents/skills/, so skills committed for Claude Code or Codex load in Amp without copying. The Codex vs Claude Code comparison goes deeper on those two.

Amp suits people who want agents running in the background on remote machines and threads they can pick up from any device, without managing model choice. Claude Code and Codex suit people who want a specific vendor's models and their tighter local workflows.

How do you keep one skill library for Amp and other agents?

Amp reads skills from the same folders as Claude Code and Codex, so one library can feed all three. Kipt stores your prompts and skills as Markdown in a private GitHub repo you own, with Cmd+K search and imports from a GitHub link, a folder or a .zip. Because Amp reads .agents/skills/ and .claude/skills/, any skill folder you copy from that library into a project is picked up in your next session.