Amp vs Claude Code: mode-based agent against Anthropic's agent

What is the difference between Amp and Claude Code?

Amp is a coding agent where you choose how hard it should work, low, medium, high or ultra, and Amp picks the model, while its agents can run on remote machines called orbs. Claude Code is Anthropic's coding agent that runs Claude models in your terminal, IDE, desktop app or browser, with you choosing the model and permission mode. Both read your code, edit files and run commands, but Amp is built around routing and background work, and Claude Code around Anthropic's models and a deep local toolset.

They share more than you might expect. Amp falls back to CLAUDE.md and reads .claude/skills/, so a repo set up for Claude Code works in Amp with no changes. Everything below comes from the official docs as of 26 October 2026.

How does Amp work?

The Amp manual describes it as "a coding agent and development environment built for the frontier." It runs in the browser at ampcode.com, in macOS and mobile apps, and as a CLI:

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

Amp was built at Sourcegraph, but on 2 December 2025 it announced: "Amp is becoming a separate company. We're spinning out of Sourcegraph to become an independent research lab."

Three ideas shape it:

• The Dial. The Dial docs say it "controls how hard Amp works on a task." medium "is the default for most work," and high "is for hard tasks where a subtle miss would be expensive." Amp chooses the model, effort and tools for each mode.
• Orbs. The orbs docs call them "remote machines where Amp agents work without using your computer," and say "An agent in an orb keeps working while your laptop is closed."
• Threads. Each conversation is a thread you can reopen on any surface and share by link.

Our Amp Code guide covers setup, threads and settings in more depth.

How does Claude Code work?

The Claude Code overview calls it "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools." You install it and start it in a project:

curl -fsSL https://claude.ai/install.sh | bash
cd your-project
claude

It runs on Claude models, and the same engine powers a VS Code extension, a JetBrains plugin, a desktop app and the browser at claude.ai/code. The setup docs say it "requires a Pro, Max, Team, Enterprise, or Console account," and also works through Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry.

You control safety with permission modes. Auto mode, where background safety checks review actions, is the starting mode for interactive sessions in recent versions, and Shift+Tab switches to Manual, Edit automatically or Plan. Claude Code also supports subagents, hooks, MCP servers, skills and plugins. See the Claude Code setup guide.

How do Amp and Claude Code compare side by side?

How do the Dial and model choice compare?

This is the main design difference.

Amp asks you to choose "a mode based on how much the agent needs to figure out, not on which model happens to be behind it." On 26 October 2026 the modes page listed GLM-5.3-Flash for low, Opus 5.5 for medium, GPT-6 Astra for high and Fable 5.1 for ultra. Amp changes this routing over time. In the CLI you press Ctrl+S before your first message to turn the Dial, and a thread keeps its mode once started.

Amp can also use subscriptions you already pay for. If you sign in with ChatGPT, "OpenAI model usage is billed to your ChatGPT subscription instead of Amp credits," and a ChatGPT Only preset pins low, medium and high to OpenAI models. For Claude, the Dial docs say "The Claude Code mode in Amp lets you use Claude models with your Claude Subscription," running on the Claude Agent SDK. Amp's own modes are never covered by a Claude plan, and Amp-hosted orbs are billed by Amp.

Claude Code gives you Anthropic's models directly. You pick the model and effort, and it works the same way across every surface. If you want one vendor's models with full control over which one runs, that is simpler. If you would rather not think about models at all, Amp's Dial is built for that.

How do instructions, skills and MCP compare?

Instructions. Claude Code reads CLAUDE.md, and its memory docs say it reads AGENTS.md "as your project instructions" when there is no CLAUDE.md. Amp's AGENTS.md docs work the other way round: AGENTS.md first, and "If no AGENTS.md exists in a directory, but a file named AGENT.md (without an S) or CLAUDE.md does exist," Amp includes that file. Either file works in both tools, so pick the one that suits the rest of your team. The AGENTS.md guide and CLAUDE.md guide cover each format.

Skills. Claude Code loads skills from .claude/skills/ and ~/.claude/skills/. Amp's skills docs search .agents/skills/, .claude/skills/, ~/.claude/skills/ and several Amp folders, with the first skill of a given name winning. So skills you commit for Claude Code load in Amp too. Set amp.skills.disableClaudeCodeSkills if you want Amp to skip them. Amp also installs skills with amp skill add <repo-or-path>.

MCP. Both support MCP. Amp stores local servers under amp.mcpServers in ~/.config/amp/settings.json or a workspace .amp/settings.json, and workspace servers need approval with amp mcp approve before they run. The Amp MCP docs recommend "bundling MCP servers in skills" with a sibling mcp.json, so tools load only when the skill does. Claude Code connects servers per project or user; the Claude Code MCP guide covers the commands.

How do they handle background and scripted work?

Amp sends work to orbs. amp -ox "<prompt>" starts a thread in an orb from your terminal, and the orb sleeps a few minutes after you stop interacting, costing nothing until your next prompt wakes it. For scripts, execute mode "waits until the agent ends its turn, prints its final message, and exits":

amp -x "what package manager is used here?"
echo "summarize the changes on this branch" | amp -x

Claude Code runs cloud sessions at claude.ai/code that keep going while you are away, cloud scheduled tasks that run when your computer is off, and claude --teleport to pull a cloud session into your terminal. For scripts, claude -p runs one task and exits, and there are GitHub Actions and GitLab CI/CD integrations:

claude -p "summarize the changes on this branch"
git diff main --name-only | claude -p "review these changed files for security issues"

Both cover background work well. Amp ties it to one model of threads and orbs across every surface, and Claude Code spreads it across web sessions, scheduled tasks and CI. See Claude Code on the web for the cloud side.

Which should you pick?

Start with how much you want to manage models.

Amp tends to fit when:

• You want to pick a difficulty level and let Amp choose the model.
• You want agents on remote machines that keep working with your laptop closed.
• You want threads you can pick up on web, desktop and phone, and share by link.
• You already pay for ChatGPT or want to bring your own keys.

Claude Code tends to fit when:

• You want Anthropic's Claude models with direct control over which one runs.
• You want deep IDE integration in VS Code and JetBrains.
• You want Claude Code's permission modes, hooks, subagents and plugins.
• Your organization buys through a Claude plan or Bedrock, Google Cloud or Microsoft Foundry.

Because Amp reads CLAUDE.md and .claude/skills/, trying it on a Claude Code repo takes no setup. For how Claude Code compares with OpenAI's agent, see Codex vs Claude Code.

Keep one skill library for both

Amp and Claude Code both load skills from .claude/skills/, so one set of folders can serve both tools. Kipt keeps your prompts and Agent Skills as Markdown in a private GitHub repo you own, imports skills from a GitHub link, a folder or a .zip, and installs every skill into Claude Code with one npx skills add command. The free SKILL.md generator turns a prompt you reuse into a skill.