Claw field notebook
● last updated 2026-05-15 edit on GitHub colophon
§ 7 Compare / § 7.2 · 8 min read

CLI coding agents — Claude Code · Codex CLI · Gemini CLI · Copilot CLI

Four shell-native AI coding CLIs. Same shape, different identities. Where each wins, where each lags, when to switch — and why most thoughtful people end up using two.

What this comparison is#

Four shell-native AI coding CLIs. You install a binary, run a command in your project, and the CLI reads files, runs commands, edits code, and talks to MCP servers — all from your terminal.

From a distance they look the same. Once you actually use them, they diverge fast — different default models, different auth, different sandboxing philosophies, different plan modes. This is the “which CLI fits which job” cut. Not a benchmark. Not a winner table. The honest take at the bottom is which one I’d pick for which kind of work.

Dimension Claude Code Codex CLI Gemini CLI GitHub Copilot CLI
Install (recommended) curl -fsSL https://claude.ai/install.sh | bash npm install deprecated; brew + WinGet + PowerShell installer also supported npm install -g @openai/codex brew install --cask codex or binary release also supported npm install -g @google/gemini-cli brew install gemini-cli or npx (no install) also supported curl -fsSL https://gh.io/copilot-install | bash Renamed from gh-copilot (deprecated Oct 2025) to standalone copilot-cli; brew/winget/npm also supported
Auth Subscription OAuth (Pro/Max/Team/Enterprise) or API key Also: Bedrock, Vertex, Foundry env vars; OAuth tokens; apiKeyHelper for rotating creds Sign in with ChatGPT (Plus/Pro/Business/Edu/Enterprise) or API key Also: device code (beta), enterprise access tokens, custom model providers Personal Google OAuth, Gemini API key, or Vertex AI OAuth path has no API key — just run `gemini` GitHub OAuth (browser /login) or fine-grained PAT with Copilot Requests scope Subject to org/enterprise Copilot policy
Free tier None — paid plan required Pro from $17/mo billed annually ($20/mo monthly); or pay-per-token via Anthropic API None — ChatGPT Plus or API key required The old $5 trial credit no longer exists in current docs 1,000 req/day on personal Google OAuth (60 req/min cap) or AI Studio API key Both free-tier paths give the same daily cap; OAuth uses no API key, AI Studio key gives model-selection control Available with GitHub Copilot Free (limited premium requests) Each prompt counts against the plan's monthly premium-request quota
MCP support Client (stdio · SSE · streamable-HTTP) Project-scoped `.mcp.json` + user-scoped `~/.claude.json`; not exposed as an MCP server itself Client (stdio); built-in MCPs as first-class runtime servers (v0.130.0+) Destructive MCP tool calls always require approval Client (stdio · SSE · streamable-HTTP) Configure in `~/.gemini/settings.json`; resource tools added v0.40.0; OAuth-aware Client (stdio); ships with GitHub's MCP server by default Custom servers via /mcp; OAuth client-credentials grant for headless auth; experimental MCP Tasks
Sandboxing Permission modes (no container/VM) default · acceptEdits · plan · auto · dontAsk · bypassPermissions; protected: .git, .claude/ OS-enforced sandbox: workspace-write · read-only · danger-full-access Network off by default; web search cached/live/disabled; auto_review checks data exfiltration Opt-in sandbox: macOS Seatbelt · Docker/Podman · Windows Native · gVisor · LXC Disabled by default; enable via `-s` flag or `GEMINI_SANDBOX` env var Trust + per-tool approval (no container) Per-directory trust; --allow-tool · --deny-tool · --allow-all-tools; read-only `gh` cmds auto-approved
Plan mode First-class — Shift+Tab cycles default → acceptEdits → plan Also /plan prefix, --permission-mode plan, or settings default No dedicated plan mode Closest: --sandbox read-only --ask-for-approval on-request First-class plan mode (enabled by default) Routes Pro for planning + Flash for implementation; --yolo flag bypasses (CLI-only) First-class — Shift+Tab cycles ask/execute → plan Plus experimental Autopilot mode via /autopilot or Shift+Tab cycle
Trusted folder behaviour No folder-trust dialog Uses permission rules + per-project MCP server approval instead Detects version-control state; recommends Auto for git repos Read-only mode recommended for non-VC folders; no formal trusted-folders file Off by default; opt in via settings.json When enabled: first-run dialog; untrusted runs in safe mode (no MCP, no extensions, no .env) Prompted on first use per directory Choose session-only · remember this folder · exit; persistent across sessions since v1.0.37
Default model Claude Sonnet 4.6 (auto-mode uses Opus 4.7 on eligible plans) Fast-mode default bumped to Opus 4.7 in 2.1.142 (CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1 pins back to 4.6); switch via /model · --model · ANTHROPIC_MODEL env gpt-5.5 (recommended); gpt-5.4 (alt high-end); gpt-5.4-mini (fast); gpt-5.3-codex (coding specialist) Also gpt-5.3-codex-spark research preview (Pro subs only). Switch via /model · -m flag · config.toml gemini-2.5-pro (stable channel); gemini-3-pro-preview (preview channel) Per Gemini CLI docs: Auto (Gemini 3) routes between gemini-3-pro-preview and gemini-3-flash-preview; on the API side, gemini-3-pro-preview redirects to gemini-3.1-pro-preview since 2026-03-09. Gemma 4 also enabled by default via Gemini API since v0.42.0; plan mode auto-routes Pro for planning, Flash for implementation Claude Sonnet 4.5 (per README) Also GPT-5 and others via /model picker; default subject to change
OS support macOS · Linux · Windows native (+ WSL) macOS · Linux · Windows macOS · Linux · Windows (incl. Windows Native Sandbox) macOS · Linux · Windows (PowerShell 6+) + WSL

The four identities#

If you only remember one thing per CLI:

  • Claude Code — Anthropic’s premier coding agent. Subscription-first (Pro/Max/Team/Enterprise OAuth). The most polished plan/permission model of the four (plan / acceptEdits / auto modes cycled with Shift+Tab). Recently moved off npm to a curl-installer as the recommended path.
  • Codex CLI — OpenAI’s terminal coder. OS-enforced sandbox (workspace-write default, network off, web search defaults to a cached index). Strong default safety posture, less mode-switching ceremony. ChatGPT-account-first; the old $5 free credit is gone.
  • Gemini CLI — Google’s terminal coder. The only one of the four with a meaningful no-cost tier — 1,000 req/day on a personal Google account (no card needed), or the same daily cap via an AI Studio API key with explicit model-selection control. The widest sandbox backend menu (Seatbelt, Docker, Podman, Windows Native, gVisor, LXC). Plan Mode is first-class and routes Pro for planning + Flash for implementation.
  • GitHub Copilot CLI — GitHub’s terminal coder. Renamed in October 2025 — the old gh-copilot extension is deprecated; the current product is the standalone copilot-cli binary. Defaults to Claude Sonnet 4.5 (not a GitHub-owned model). Ships with GitHub’s MCP server pre-wired. The CLI that’s easiest to use if your work already lives on github.com.

Where each wins#

Claude Code#

  1. The plan/approval model. plan mode is a first-class permission state — you can ask Claude to research and propose a multi-file edit before it writes anything, then approve, refine, or hand-edit the plan with Ctrl+G. Shift+Tab cycles through default → acceptEdits → plan mid-session. Of the four, this is the one that feels most like working with a thoughtful contractor rather than a junior who just starts typing.
  2. The breadth of the auth model. Six different ways to authenticate (subscription OAuth, API key, env-var token, OAuth long-lived token, apiKeyHelper for rotating creds, plus Bedrock/Vertex/Foundry routing). If you need to point Claude Code at a specific model gateway or rotating-credential setup, it’s the most accommodating.
  3. Auto mode with a classifier. On Max/Team/Enterprise/API, auto mode runs autonomously with a background classifier model checking each action for destructive or exfiltration-style behaviour. The other three either prompt for everything or rely on a sandbox.

Codex CLI#

  1. The strongest default sandbox. workspace-write mode plus network-off-by-default plus cached web-search plus an auto_review reviewer agent that watches for data exfiltration. You don’t have to think about safety knobs; the defaults are tight.
  2. Custom model providers as a first-class concept. If you want to point Codex CLI at a non-OpenAI model (your own deployment, a self-hosted endpoint, another vendor via a proxy), the config supports it cleanly — Chat Completions API or Responses API, with auth via env var, OpenAI keyring, or no auth (for local models).
  3. The codex-spark research preview. A specialist model optimised for near-instant iteration on small tasks. Available on Pro subscriptions only at time of writing — niche, but if you do a lot of small typed-prompt-typed-prompt-typed-prompt iteration, the latency difference shows.

Gemini CLI#

  1. The free tier that’s actually free. A personal Google account (OAuth) gets you 1,000 requests/day — no card, no plan. An AI Studio API key gives you the same daily cap with explicit model-selection control. Either way, of the four, this is the one a student or a beginner can pick up without paying anything.
  2. The widest sandbox menu. macOS Seatbelt for native isolation. Docker/Podman for cross-platform. Windows Native Sandbox (uses icacls to set Low Mandatory Level on files — note that changes are persistent after the sandboxed session, unlike VM snapshots). gVisor for Linux user-space-kernel isolation. LXC for full-system containers. If your security posture cares about which mechanism is doing the isolating, Gemini CLI lets you choose.
  3. Plan Mode with intelligent model routing. Plan mode is enabled by default; when active, Gemini routes the heavier Pro model for the planning phase and the cheaper Flash for execution. You get plan-quality reasoning without paying Pro rates for every line of generated code.

GitHub Copilot CLI#

  1. It already knows your GitHub world. Ships with GitHub’s MCP server pre-wired — issues, PRs, repo metadata, releases, workflow runs — without configuration. If your code lives on github.com, this is the lowest-friction CLI of the four for “agent that knows about my repo.”
  2. Plan mode plus experimental Autopilot. Shift+Tab cycles into a structured plan mode where Copilot asks clarifying questions before writing. The --experimental flag (or /autopilot) adds autopilot mode — agent runs continuation after continuation until the task is done (capped at 5 by default with --max-autopilot-continues). Closer to “kick off and walk away” than the other three out of the box.
  3. Free plan availability. Copilot CLI ships with every GitHub Copilot plan — including the Copilot Free plan (which has a limited monthly premium-request quota). Of the four, this is the one with the lowest “I want to try a real agent CLI” cost for someone who already has a GitHub account.
  4. Custom agent profiles + repo-level instructions. AGENTS.md at the repo root, .github/copilot-instructions.md, and ~/.copilot/agents/ for personal agent profiles give it a richer customisation surface than the other three for “make this agent behave consistently across my team.”

Where each lags#

Claude Code#

  • No free tier of any kind. Pro is $17/mo annual minimum. Or pay-per-token via the Anthropic API (which works fine, but the OAuth-subscription experience is what most users come for).
  • The npm install is deprecated. You’re meant to use the curl/PowerShell installers or Homebrew/WinGet. If you have habit-installed it via npm, your install isn’t recommended any more.
  • No built-in sandbox. The permission-mode system is sophisticated, but if your security posture wants OS-level isolation, you’re running Claude Code inside Docker or a VM yourself.

Codex CLI#

  • No first-class plan mode the way Claude Code and Gemini CLI have. The closest equivalent is --sandbox read-only --ask-for-approval on-request, which gets you safe browsing but not the “Claude proposes a multi-step plan and you approve it” UX.
  • The old $5 free credit is gone. You need a ChatGPT Plus subscription (≈$20/mo) or an API key (pay-per-token) to use it.
  • No Shift+Tab mode cycling. Modes are set with flags at startup or with /permissions mid-session — less ergonomic than the cycle-and-go model.

Gemini CLI#

  • Trusted Folders is off by default. Some readers expect a “this folder is trusted” prompt on first launch — Gemini CLI doesn’t ask one unless you opt in via settings.json. Documented behaviour, but easy to misread as “no trust system.”
  • Windows Native Sandbox is persistent. Unlike Docker or Seatbelt, the icacls integrity-level changes survive after the sandbox session ends. Useful but surprising; readers should know.

GitHub Copilot CLI#

  • Still iterating fast. v1.0.48 as of mid-May 2026 — stable enough for daily use, but the changelog has had some breaking-ish renames (the gh-copilot deprecation in October 2025 wasn’t smooth — docs.github.com/copilot/how-tos/use-copilot-for-common-tasks/use-copilot-in-the-cli now redirects to a deprecation notice for the old extension). Newer features (Autopilot, MCP Tasks) are flagged experimental and shift between releases.
  • Default model is not GitHub’s. Claude Sonnet 4.5 is the default, with GPT-5 and others available via /model. GitHub reserves the right to change the default. This is fine, but worth knowing if you assumed Copilot CLI was running a GitHub-owned model.
  • No OS-enforced sandbox. Like Claude Code, the safety story is trust-prompts and per-tool approval. Good for productivity, less good if you want OS-level isolation.

Decision sketch#

Do you need a meaningful no-cost tier (no card, no plan)?
  ├─ Yes → Gemini CLI (1,000 req/day on personal Google account)
  └─ No → continue
       │
       Is your code already on github.com (issues/PRs/Actions central)?
         ├─ Yes → GitHub Copilot CLI (ships with GitHub MCP server)
         └─ No → continue
              │
              Do you care most about plan-before-execute UX?
                ├─ Yes → Claude Code (best plan/approval model of the four)
                └─ No → continue
                     │
                     Do you care most about default sandbox tightness?
                       ├─ Yes → Codex CLI (workspace-write + network-off + auto_review)
                       └─ No → pick by default model preference

They overlap more than they compete#

For most builders, the right answer is two of the four, not one. Common combinations worth knowing:

  • Copilot CLI for daily repo work + Claude Code for the gnarly multi-file refactors. Copilot CLI knows the repo + the PR/issue surface; Claude Code’s plan mode is better when you need a deliberate multi-step rewrite.
  • Gemini CLI for personal projects + a paid CLI for work. The free tier is genuinely usable for personal side-project velocity; the paid tools have higher reliability and more polished plan UX for production code.
  • Codex CLI as the “safe one” in CI/autonomous flows. The default-tight sandbox plus the auto-review reviewer agent make it a reasonable choice for “run this CLI inside a runner without a human watching.”

Picking one CLI to rule them all is a choice that costs you. Try two; switch on context.

What we are NOT claiming#

  • Which CLI writes the best code. Quality varies by language, framework, model selection, and how you prompt — across all four. We have no benchmark data and would not trust a benchmark even if we had one. The CLI shape doesn’t predict generated-code quality.
  • Specific pricing in foreign currencies. Pricing is in USD here and tracks the vendor’s listed page; check your billing dashboard for local rates.
  • Anything about model quality differences. That’s out-of-scope per Claw’s scope guardrails. We compare tool shapes, not model leaderboards.
  • What will be true in six months. All four vendors ship fast. Versions in this comparison are current as of 15 May 2026. Free-tier numbers and default models in particular move.

Honest take#

If you’re Sush — Microsoft Copilot SE, code lives mostly on github.com, builds Hugo and Astro sites against MCP servers all day, doesn’t want to run a Docker container around the agent — GitHub Copilot CLI is the daily driver. It’s been the daily driver. This site was built with it.

For the work where you want to think before you cut — a multi-file refactor, a careful migration, anything where the agent needs to plan before it acts — Claude Code’s plan mode wins. Worth the second subscription for the maybe-twice-a-week jobs that benefit from it.

If you’re a student or new builder with no budget, Gemini CLI’s free tier is genuinely the right starting place. 1,000 requests/day on a personal Google account is enough to learn what these tools can do without paying anything.

Codex CLI is the one I reach for least, but I’d pick it first if I were standing up an autonomous coding agent in CI with a security-conscious org behind me. The defaults are tighter than the other three’s, which matters when no human is watching.

Sources

See also

Related entries