Which AI CLI tool has the best subagent support?

Claude Code has the most mature subagent system as of April 2026, supporting Agent Teams with direct inter-agent messaging via a shared task list. All four tools support subagents, but Claude Code's implementation allows isolated context, custom system prompts, and parallel agent execution.

Which CLI tools support MCP (Model Context Protocol)?

All four CLI tools support MCP. Claude Code supports stdio and SSE. Codex CLI supports stdio and streaming HTTP, and can also run as an MCP server itself. Gemini CLI supports stdio, HTTP, and SSE. GitHub Copilot CLI includes GitHub MCP built-in plus custom servers.

Which AI coding CLI is open source?

Codex CLI (OpenAI) and Gemini CLI (Google) are both open source under Apache 2.0. Claude Code and GitHub Copilot CLI are proprietary.

Which AI CLI tool has the largest context window?

Gemini CLI has the largest context window at up to 1 million tokens. Claude Code supports 200K tokens, Codex CLI approximately 192K tokens, and Copilot CLI varies by the model selected.

Claude Code vs Codex CLI vs Gemini CLI vs Copilot CLI — Full Feature Comparison 2026

Last verified: April 3, 2026

CLI Agent
Feature Comparison

All four tools are now full agentic terminal coding agents. The real differentiation is model ecosystem, pricing, open-source status, and depth of specific features — not feature presence/absence.

Core Agent Capabilities

Feature	Claude Code	Codex CLI (OpenAI)	Gemini CLI (Google)	GitHub Copilot CLI
Advanced Logic / Thinking	✅ Sonnet 4.5 default	✅ GPT-5.4, GPT-5.3-Codex, codex-mini; adjustable reasoning	✅ Gemini 2.5 Pro / 3 / 3.1 Pro	✅ Sonnet 4.5 default + GPT-5.3, Gemini 3 Pro, Opus 4.6
Running Code	✅ Bash, runs scripts	✅ Shell commands, file edits, test runs	✅ Built-in shell commands, file ops	✅ Commands, tests, builds
Bash Execution	✅ Yes	✅ 3 modes: read-only / auto / full-access + `--yolo`	✅ Built-in `run_shell_command`	✅ User confirms; `--yolo` for auto
Web Search	✅ WebSearch + WebFetch tools	✅ Built-in `web_search` tool	✅ Google Search grounding + `web_fetch`	✅ `web_fetch` tool (built-in MCP)
Image Input	✅ Paste images in chat	✅ Screenshots, wireframes, design specs	✅ Multimodal: images, PDFs, clipboard paste	✅ Yes (images sent to models)

Extensibility & Skills

Feature	Claude Code	Codex CLI (OpenAI)	Gemini CLI (Google)	GitHub Copilot CLI
Skills	✅ Markdown-based skill files	✅ AGENTS.md + custom commands	✅ Agent Skills (`.md`, `/skills install`)	✅ Agent Skills (`.md`), shared with Copilot cloud agent & VS Code
MCP Support	✅ Native (stdio, SSE)	✅ Native (stdio + streaming HTTP); can also run AS an MCP server	✅ Native (stdio, http, sse)	✅ GitHub MCP built-in + custom servers
Hooks / Lifecycle	✅ PreToolUse, PostToolUse, Notification hooks	✅ ExecPolicy rules for command control	✅ Hooks for customizing behavior	✅ PreToolUse, PostToolUse, sessionStart, subagentStart hooks
Plugins / Extensions	✅ Plugin system	❌ Not yet (AGENTS.md covers most use cases)	✅ Extensions (`/extensions install`)	✅ Plugins (`/plugin install owner/repo`)

Multi-Agent & Delegation

Feature	Claude Code	Codex CLI (OpenAI)	Gemini CLI (Google)	GitHub Copilot CLI
Subagents	✅ Isolated context + custom system prompts, tools, permissions	✅ Via `config.toml` roles; parallel execution in git worktrees	✅ Custom agents (`.gemini/agents/`), generalist agent for routing (experimental)	✅ Built-in agents (explore, task, code review, plan) + custom `.agent.md` files
Parallel Agents	✅ Agent Teams (Opus 4.6): teammates communicate via shared task list & direct messages	✅ Multi-agent in parallel worktrees; orchestration via OpenAI Agents SDK	✅ Subagents can run in parallel (experimental); YOLO mode by default	✅ `/fleet` coordinates subagents in background; multiple sessions in VS Code
Cloud Delegation	❌ No (local only)	✅ `codex cloud exec` — delegates to Codex Cloud sandbox; results applied locally	❌ No (local only)	✅ `&` prefix delegates to Copilot cloud agent; `/resume` switches between local & cloud

Memory & Persistence

Feature	Claude Code	Codex CLI (OpenAI)	Gemini CLI (Google)	GitHub Copilot CLI
Project Context File	✅ `CLAUDE.md` (hierarchical: global → project → subdirectory)	✅ `AGENTS.md` (project root)	✅ `GEMINI.md` (hierarchical: global → project → subdirectory)	✅ `.github/copilot-instructions.md` + `AGENTS.md`
Persistent Memory	✅ Subagent memory directories (`~/.claude/agent-memory/`); file-based memory tool	✅ Session JSONL logs; config-based persistence	✅ `save_memory` tool appends to `~/.gemini/GEMINI.md` automatically	✅ Repository memory (learns conventions cross-session); cross-session memory; `/chronicle` for insights
Session Resume	✅ `/resume`, `--continue`	✅ `codex resume`, `codex exec resume --last`	✅ `--resume` flag, `/resume` browser, `/chat save/resume <tag>`	✅ `--resume`, `--continue`, `/resume`; sessions shared between CLI & VS Code
Auto-Compaction	✅ Context compaction when approaching limit	✅ Auto-compaction for long sessions	✅ Context management with session limits	✅ Auto-compaction at 95% of token limit

Headless & Automation

Feature	Claude Code	Codex CLI (OpenAI)	Gemini CLI (Google)	GitHub Copilot CLI
Headless Run	✅ `-p` flag	✅ `codex exec` (dedicated non-interactive mode)	✅ `gemini -p "prompt"`	✅ `-p` / `--prompt` flag
Streaming from Headless	✅ JSON output + stdout streaming	✅ `codex exec --json` streams JSONL events (turns, commands, reasoning)	✅ `--output-format stream-json`	✅ `--output-format=json` JSONL stream
JSON Response	✅ Via prompts	✅ `--json` flag + `--output-schema` for validated structured JSON	✅ `--output-format json`	✅ `--output-format=json`
CI/CD Integration	✅ Can be scripted	✅ First-class: `codex exec --full-auto` with API key auth for CI	✅ Can be scripted with headless mode	✅ Can be scripted with `-p` + `--autopilot`

Platform & Ecosystem

Feature	Claude Code	Codex CLI (OpenAI)	Gemini CLI (Google)	GitHub Copilot CLI
Open Source	❌ No	✅ Yes (Apache 2.0, built in Rust)	✅ Yes (Apache 2.0)	❌ No (public repo, proprietary)
Sandboxing	✅ Permission modes (ask, auto-accept, etc.)	✅ OS-level: macOS Seatbelt, Linux Landlock; network blocked by default	✅ Sandboxed execution environment	⚠️ Permission-based (no OS-level sandbox)
Multi-Model	❌ Anthropic models only	⚠️ OpenAI models only (+ local Ollama via `--oss`)	❌ Google models only	✅ Anthropic + OpenAI + Google models
Context Window	200K tokens	~192K tokens	Up to 1M tokens	Varies by model selected
Pricing	Anthropic API / Claude Pro subscription	Included with ChatGPT Plus/Pro/Business/Enterprise; or API key	Free tier: 1,000 req/day; paid via Google AI Pro or Vertex AI	Included with any Copilot plan (Free/Pro/Pro+/Business/Enterprise)
IDE Integration	✅ VS Code extension	✅ VS Code / Cursor / Windsurf extension + Desktop app	✅ Gemini Code Assist (VS Code) — shared technology	✅ Native in VS Code Chat; sessions sync between CLI & editor

Key Differentiators

Claude Code

Most mature subagent system (Agent Teams with direct inter-agent messaging); strongest reasoning on complex tasks; Anthropic-only models.

Codex CLI

Best CI/CD story (codex exec + --output-schema for validated JSON); cloud delegation; OS-level sandboxing; open source (Rust).

Gemini CLI

Largest context window (1M tokens); most generous free tier; multimodal (images + PDFs); Google Search grounding; open source.

Copilot CLI

Only true multi-model CLI (Anthropic + OpenAI + Google); deepest GitHub integration; repository memory that learns across sessions; cloud agent delegation.