04 · Compare
How it compares
Features only, as each vendor describes them on its own public pages, with the date we read them. No judgements about anyone’s quality. Where a page says nothing, we write “not stated”.
| Feature | Ephemeral Code | Claude Code | OpenAI Codex CLI | Aider | Cursor |
|---|---|---|---|---|---|
| Your own API key | Yes — your provider key, held only by the gateway E | Yes, ANTHROPIC_API_KEY, or a Claude subscription; the CLI also supports third-party providers C1 | Sign in with ChatGPT, or an API key X1 | Your provider keys (e.g. --api-key); local models supported A1 | Own keys for chat models (OpenAI, Anthropic, Google, Azure, Bedrock); Tab uses Cursor’s models R2 |
| Where it runs | CLI on your PC (Windows, Linux x86_64); engine on your server E | Terminal, IDE, desktop app, web. Local sessions on your machine; cloud sessions in Anthropic VMs C1 C3 | Locally, in your terminal X1 | In your terminal A1 | An app for your computer R4 |
| Before a command runs | Manual: asks before each program. Auto (local engine only): a short read-only list runs without asking. Remote engine: Manual only; Auto coming E | Default Auto mode: a classifier reviews actions instead of you. Manual mode asks before edits and commands; read-only commands run without asking C2 | Auto preset (git folders): runs workspace commands without asking; asks before editing outside the workspace or using the network X2 | not stated | Run modes: Auto-review (“recommended default”), Allowlist, Run Everything R3 |
| OS sandbox for commands | Not on your PC today; a container mode is planned. Benchmark runs: rootless Podman E | Optional sandbox for bash commands (filesystem and network isolation), set up with /sandbox C2 | Yes, OS-level (Seatbelt on macOS, bwrap + seccomp on Linux); network off by default X2 | not stated | macOS (Seatbelt) and Linux (Landlock + seccomp); network blocked by default R3 |
| “Done” tied to tests | “Fixed and verified” only after the verification passes E | not stated | not stated | Can lint and test after every change A1 | not stated |
| Git result | A commit on the task’s own branch, after your approval; you merge E | Creates commits, branches and pull requests when you ask C1 | not stated | Commits each edit automatically; --no-auto-commits turns it off A2 | not stated |
| Telemetry to the vendor | None E | Usage metrics on by default with the Claude API; DISABLE_TELEMETRY=1 opts out C3 | Off by default; opt-in X2 | Opt-in analytics only A3 | not stated |
| Source code | Closed E | “All rights reserved” C4 | Apache-2.0 X1 | Apache-2.0 A4 | not stated |
| Price | Free during the private beta E | A Claude subscription or an Anthropic Console account C1 | Included in ChatGPT Plus, Pro, Business, Edu and Enterprise plans X1 | not stated | Free $0; paid plans $8–$200/month; Teams $30+/user/month R1 |
Sources — all read on 27 October 2026
- E — Ephemeral Code R3 (commit 85fdd861): source and release evidence, read 31 October 2026
- C1 — code.claude.com/docs/en/overview
- C2 — code.claude.com/docs/en/security
- C3 — code.claude.com/docs/en/data-usage
- C4 — github.com/anthropics/claude-code/blob/main/LICENSE.md
- X1 — github.com/openai/codex (README)
- X2 — learn.chatgpt.com/docs/agent-approvals-security
- A1 — aider.chat
- A2 — aider.chat/docs/git.html
- A3 — aider.chat/docs/more/analytics.html
- A4 — github.com/Aider-AI/aider
- R1 — cursor.com/pricing
- R2 — cursor.com/docs/settings/api-keys
- R3 — cursor.com/docs/agent/security/run-modes
- R4 — cursor.com/docs
Product names belong to their owners. Spotted something out of date? Tell us and we correct it, with the new date.