What is the difference between Codex and Amp?
Codex is OpenAI's coding agent for the terminal, editor, desktop app and cloud, running OpenAI models under your ChatGPT plan or API key. Amp is an independent coding agent where you choose a difficulty mode instead of a model, and where agents often run on remote machines called orbs that you follow from the web, desktop or phone.
Both read AGENTS.md, use Agent Skills and MCP servers, have a one-shot mode for scripts, and can work remotely while your laptop is closed. They differ in model choice, billing and where the agent usually runs. Everything below comes from official docs read on 26 October 2026.
How does Codex work?
The Codex repository describes the CLI as a "Lightweight coding agent that runs in your terminal," licensed under Apache 2.0:
# macOS and Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"
# or with npm or Homebrew
npm install -g @openai/codex
brew install --cask codex
Run codex and choose Sign in with ChatGPT, or use an API key. Codex also runs as an IDE extension, inside the ChatGPT desktop app and as Codex Cloud, which works on OpenAI's machines from published environments.
You pick the model yourself. The models page recommends gpt-6.1-sol on Plus, Pro, Business, Enterprise and Edu plans when available, and gpt-6-luna for focused, repeatable tasks.
How does Amp work?
The Amp manual calls Amp "a coding agent and development environment built for the frontier." It runs on the web at ampcode.com, as native apps for Mac, iPhone, iPad and Android, and as a CLI:
curl -fsSL https://ampcode.com/install.sh | bash
amp login
amp
The CLI docs say the CLI "supports macOS, Linux, and Windows through WSL," and that "Amp can connect to VS Code, Cursor, Windsurf, Zed, and Neovim" from the editor's terminal.
Instead of a model picker, Amp has the Dial. The Dial docs advise: "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 routed low to GLM-5.3-Flash, medium to Opus 5.5, high to GPT-6 Astra and ultra to Fable 5.1. "A thread keeps the mode you chose when you sent its first message." The Amp Code guide covers threads, orbs and setup in more depth.
How do Codex and Amp compare side by side?
| Codex | Amp | |
|---|---|---|
| Maker | OpenAI | Amp Frontier Corporation |
| Main surfaces | CLI, IDE extension, ChatGPT desktop app, Codex Cloud | Web, Mac and mobile apps, CLI with editor connection |
| CLI platforms | macOS, Linux, Windows | macOS, Linux, Windows through WSL |
| Model choice | You pick a model, OpenAI by default | You pick a mode on the Dial; Amp routes to a model |
| Billing | ChatGPT plan limits, or API key rates | Amp credits, or your own keys and subscriptions |
| Remote work | Codex Cloud tasks from published environments | Orbs, Amp-hosted or self-hosted |
| Local safety | OS sandbox, network off by default | Approval through plugins (tool.call) and MCP tool settings |
| Project instructions | AGENTS.md |
AGENTS.md, falls back to AGENT.md or CLAUDE.md |
| Agent Skills | .agents/skills/, ~/.agents/skills/ |
.agents/skills/, .claude/skills/ and more |
| MCP config | [mcp_servers] in ~/.codex/config.toml |
amp.mcpServers in settings.json, or remote via amp mcp remote |
| Scripting | codex exec |
amp -x, or amp -ox to run in an orb |
How do Codex Cloud and Amp orbs compare?
This is where the two products feel most different.
Codex Cloud runs a task on OpenAI's machines from a published cloud environment: your GitHub repositories, dependencies, tools and network settings. The cloud environments docs say each new task "gets its own isolated workspace from the published environment." Default VMs range from 2 vCPUs and 8 GiB on Plus to 4 vCPUs and 16 GiB on Pro, Business and Enterprise, included in the plan. See Codex cloud environments for setup.
Amp orbs are, per the orbs docs, "remote machines where Amp agents work without using your computer." An orb pauses 5 minutes after the agent finishes or 20 minutes after your last interaction, whichever comes later, and "a sleeping orb costs nothing, even if it sleeps for weeks." Orbs are billed by the minute: the sizes page lists a1.tiny (1 CPU, 2 GB) at $0.08 an hour up to a1.3xlarge (16 CPUs, 44 GB) at $2.13 an hour, as of October 2026.
Setup differs too. Codex prepares the environment with you in a setup conversation and records an install script. Amp reads two optional executable files you commit, .agents/setup and .agents/resume, and reuses a matching project snapshot "for up to 72 hours." Amp also supports self-hosted orbs on your own infrastructure.
How do models and billing compare?
Codex uses your ChatGPT plan. As of October 2026, the Codex pricing page lists Plus at $20 a month with Codex on the web, CLI, IDE extension and iOS, Pro from $100 a month, and Business at $20 per user a month billed annually. Free and Go get GPT-6 Luna in the desktop app, subject to rollout. With an API key you pay API rates and lose cloud features.
Amp bills Amp credits for the models it serves, or routes through your own connections. The model routing docs list API keys, subscriptions such as ChatGPT, cloud platforms like Amazon Bedrock, and gateways like OpenRouter. As of October 2026, Amp's pricing page lists:
| Tier | Price | Highlights |
|---|---|---|
| Hobby | Free | Pay as you go Amp-hosted orbs, unlimited self-hosted orbs, bring your own keys |
| Individual (Megawatt) | $20 a month | 45,000 orb minutes on Amp-hosted orbs |
| Teams | No extra charge | Thread sharing, shared plugins and skills, pooled credits, SSO |
| Enterprise | Custom | SCIM, audit logs, spend limits, IP allowlisting |
If you already pay for ChatGPT, Amp can use it: "OpenAI model usage is billed to your ChatGPT subscription instead of Amp credits," and the ChatGPT Only preset keeps low, medium and high on OpenAI models.
How do instructions, skills and MCP compare?
Instructions. Codex reads AGENTS.md from your Codex home and from the project root down to your working folder. Amp's AGENTS.md docs include files in the working directory and parents up to $HOME, in subtrees as the agent reads them, and in ~/.config/amp/AGENTS.md. If a directory has no AGENTS.md but has AGENT.md or CLAUDE.md, Amp includes that instead.
Skills. Both scan .agents/skills/ in the project and ~/.agents/skills/. Amp's skills docs also read .claude/skills/, and the first skill with a given name wins. A skills folder committed for Codex works in Amp without copying. See Codex skills.
MCP. Codex stores servers as [mcp_servers.<name>] tables in config.toml. Amp adds them with amp mcp add to amp.mcpServers, and workspace servers in .amp/settings.json need amp mcp approve before they run.
How do they handle permissions?
Codex isolates the agent locally. The approvals and security docs say "By default, the agent runs with network access turned off," inside an OS-enforced sandbox, with an approval policy that decides when it stops to ask. See Codex approvals and sandbox.
Amp handles tool approval through plugins. The plugins docs describe a tool.call event that "fires before a tool runs" and can allow, reject, modify or synthesize a result. Remote MCP tools can be set to Allow (the default) or Ask First. For risky work, running the agent in an orb keeps it off your own machine entirely.
Which should you pick?
Codex tends to fit when:
- You already pay for ChatGPT and want Codex usage included in that plan.
- You want to choose the exact OpenAI model and reasoning effort yourself.
- You want an open source CLI with native Windows support and a local sandbox.
- You want cloud tasks without per-minute machine billing.
Amp tends to fit when:
- You prefer choosing a difficulty level and letting Amp pick the model.
- You want to mix providers, bring your own keys, or reuse a ChatGPT subscription.
- You want many agents on remote orbs you can check from your phone.
- You want self-hosted remote machines on your own infrastructure.
Both read AGENTS.md and .agents/skills/, so trying the other tool later costs little.
Keep your prompts and skills portable
Since both tools read .agents/skills/, one library can serve them both. Kipt keeps prompts and Agent Skills as Markdown in a private GitHub repo you own, with Cmd+K search, and installs every skill into Claude Code or Codex with one npx skills add command. The free SKILL.md generator turns a prompt you keep reusing into a skill folder.