Amp

Notes from reading the Amp manual
Published

November 2, 2025

I’m constantly trying new coding agents. One that’s stood out for me over the last few months is Amp. I finally took some time to carefully read their manual and took notes on some of features I’ve missed.

Distinctive Features of Amp

Amp has all the trappings of a coding agent you might expect: support for MCP, ability to set project, folder, and global context (AGENTS.md), a permissions system, headless mode, and more. But there are some distinctive features that set it apart. Most users are already familiar with these, but they’re worth mentioning.

  1. Threads: This is my favorite feature. Amp stores all your threads on the server. You can resume a thread (or a colleague’s thread) anytime. You set different visibility levels (public, unlisted, workspace-shared, or private). Reading your team’s threads helps you see what’s going on and how others are using Amp. For example, here’s a public thread where I had Amp help with SEO optimization—turning a FAQ into standalone pages. Since threads are stored on the server, I can start work on my laptop and finish it later on a server or phone. Old sessions are always accessible regardless of which device I used.

  2. Model Selection is Abstracted: Amp picks the right model for each task. Their philosophy is that they tune the prompts and the tools to get the best performance for each model. I was initially skeptical about this, but it works. You also do have some control (see below).

  3. Oracle: Exception to model abstraction: you can nudge Amp to use a more expensive model. Ask it to “use the oracle” for GPT-5 reasoning.

  4. Pay Per Token (and can be expensive): You pay per token and it’s not cheap. However, I like it enough to where I think it’s worth it because it saves me time, and I’ve gotten better at using it efficiently. I haven’t tried the free version yet.

The rest of these notes are features I discovered or didn’t completely understand until I read the manual.

1. Handoff

If you’re familiar with /compact from Claude Code, handoff is a better version. Just like Claude, you can provide a prompt explaining what you want to focus on or what’s important. Amp generates a new thread prompt with the context summarizing the current thread that you can edit before starting fresh. Showing you the new prompt helps you guide the transition. This strikes the right balance and I find that handoff keeps you in control while still being efficient.

2. Message Queueing

By default, sending a message interrupts the agent. But you can queue messages to be sent once the agent finishes its turn without interrupting its current work. In the editor extension, use Cmd-Shift-Enter (macOS) or Ctrl-Shift-Enter (Windows/Linux). In the CLI, use the queue command from the command palette.

3. Scoped Instructions

You can use globs in YAML front matter of markdown files to scope instructions to certain files. For example, create a typescript-conventions.md file with language-specific coding rules that only apply when working with TypeScript:

---
globs:
  - '**/*.ts'
  - '**/*.tsx'
---
Follow these TypeScript conventions:
- Never use the `any` type
- Use strict mode

When you bring files with this frontmatter into context (ex: @typescript-conventions.md), the context will only apply if Amp has read a file matching any of the globs. This also works by reference: if you have See typescript-conventions.md in your AGENTS.md, it will respect the globs too. Read more.

4. Editing, Reverting & Forking Messages

Press up arrow in the CLI to edit prior messages. You can edit the message, delete & restore (removes the message and reverts Amp’s changes after it), or fork (creates a new thread from that point).

5. Toolboxes

Toolboxes are similar to skills in Claude Code. I like them because they have less abstraction: you just write simple executables that describe themselves and handle inputs. When Amp starts, it runs each executable with TOOLBOX_ACTION=describe, and the tool writes its description to stdout. When Amp decides to use the tool, it runs it again with TOOLBOX_ACTION=execute.

6. Librarian

The Librarian is great for working with newer libraries or private codebases. It acts like a specialized subagent that aggressively searches through repositories or documentation sites to gather context. The inner workings aren’t public, but it’s effective at priming the agent with the information it needs before you start coding.

Tips for Using Amp

  • For the best experience, use the Ghostty terminal. I was using iTerm and had some issues.
  • Features may be slightly ahead in the VS Code extension, like the handy code review tool. Try both. I recently discovered I like the VS Code extension over the CLI for this reason.
  • CLAUDE.md compatibility: Claude Code is my second favorite coding agent. One frustrating part of using multiple agents is maintaining separate rules files. Amp handles this by falling back to CLAUDE.md when no AGENTS.md exists in a directory, so you can use the same rules across both.

The manual’s at ampcode.com/manual. Worth reading if you’re using Amp regularly.