Read more about: #agents#claude-code#infrastructure#llms#cli

Every CLI coding agent, compared

agents · cli · claude-code · tool-design · infrastructure

The terminal is where agents got serious. Not IDE plugins. Not web chatbots. The CLI.

Claude Code, Codex CLI, Gemini CLI, OpenCode. These aren’t toys. They read your codebase, edit files, run tests, commit code. Some run for hours without human intervention. Some spawn sub-agents. Some sandbox themselves so thoroughly they can’t access the network.

There are now 36 CLI coding agents. I’ve mapped the entire landscape.

The big four

The frontier labs all have terminal agents now. But an open-source project is outpacing them all.

AgentStarsLicenseLocal ModelsFree Tier
OpenCode97.5KMITYes (75+ providers)Free (BYOK)
Gemini CLI93.6KApache-2.0No1000 req/day
Claude Code64KProprietaryNoNone
Codex CLI59KApache-2.0Yes (Ollama, LM Studio)None

OpenCode exploded to 97.5K stars. It’s the free, open-source alternative to Claude Code with 650K monthly users.

Gemini CLI has the most generous free tier. 1000 requests per day with just a Google account. No API key required. But no local model support.

Claude Code is locked to Claude models but has the richest feature set. Jupyter notebook editing, sub-agent orchestration, the deepest permission system.

Codex CLI is the only one written in Rust. OpenAI rewrote it from TypeScript in mid-2025 for performance.

The full landscape

Sorted by GitHub stars.

First-party (major labs)

AgentMakerStarsLangLicenseKey Feature
Gemini CLIGoogle93.6KTSApache-2.01M token context, generous free tier
Claude CodeAnthropic64KTSProprietaryCreated MCP, Jupyter editing, deepest features
Codex CLIOpenAI59KRustApache-2.0Rust performance, model-native compaction
Qwen CodeAlibaba18.1KTSApache-2.0Ships with open-weight Qwen3-Coder
Trae AgentByteDance10.7KPythonMITSOTA on SWE-bench Verified
Copilot CLIGitHub8KShellProprietaryGitHub ecosystem integration
Kimi CLIMoonshot AI5.9KPythonApache-2.0First Chinese lab with CLI agent
Mistral VibeMistral3KPythonApache-2.0Only European lab CLI agent
Junie CLIJetBrains31TSProprietaryDeep JetBrains integration, CI/CD native
Amazon Q CLIAWS1.9KRustApache-2.0Deprecated, now Kiro (closed-source)

Community & independent

AgentStarsLangLicenseKey Feature
OpenCode97.5KTSMIT75+ providers, 650K users
OpenHands67.5KPythonMITFull platform, Docker sandbox, $18.8M raised
Open Interpreter62KPythonAGPL-3.0Runs any code, not just file edits
Cline CLI57.6KTSApache-2.0IDE agent that added CLI mode
Aider40.3KPythonApache-2.0Pioneer, git-native, tree-sitter repo map
Continue CLI31.2KTSApache-2.0JetBrains + CLI, headless CI mode
Goose29.9KRustApache-2.0MCP-native architecture, Block-backed
Warp25.9KRustProprietaryFull terminal replacement with agents
Roo Code22.1KTSApache-2.0Multi-agent orchestration (Boomerang)
Crush19.5KGoCustomBeautiful TUI, from Bubble Tea team
SWE-agent18.4KPythonMITResearch-grade, NeurIPS paper
Plandex15KGoMITDiff sandbox, git-like plan branching
Kilo Code14.9KTSApache-2.0500+ models, zero markup
Claude Engineer11.2KPythonMITSelf-expanding tools
AIChat9.2KRustApache-2.0Swiss Army knife CLI
DeepAgents8.9KPythonMITLangChain’s agent harness
Pi6.6KTSMITOnly 4 tools, self-extending
ForgeCode4.6KRustApache-2.0300+ models, Rust performance
Kode CLI4.3KTSApache-2.0Multi-model collaboration
gptme4.2KPythonMITOG agent (2023), still active
AutoCodeRover3.1KPythonSource-Available$0.70/task on SWE-bench
Codebuff2.8KTSApache-2.0Multi-agent architecture
Codel2.4KTSAGPL-3.0Docker sandbox built-in
Grok CLI2.3KTSMITxAI/Grok in terminal
Agentless2KPythonMITNo persistent agent loop
AmpN/ATSProprietaryMulti-model per-task (Sourcegraph)

Agent orchestrators

These don’t write code themselves. They run multiple CLI agents in parallel.

ToolStarsWhat it does
Claude Squad5.9KParallel agents via tmux + git worktrees
Toad2.1KUnified TUI for multiple agents (by Rich creator)
Superset1.2KTerminal command center for agent teams
Emdash1.2KYC-backed, Linear/GitHub/Jira integration

Feature comparison

The features that actually differentiate them.

AgentMCPSandboxSub-agentsHeadlessPlan ModeProject Memory
OpenCodeYesDockerYesYesYesAGENTS.md
Claude CodeYesSeatbelt/BubblewrapYesYesYesCLAUDE.md
Codex CLIYesSeatbelt/LandlockYesYesYesAGENTS.md
Gemini CLIYesSeatbelt/DockerYesYesYesGEMINI.md
Qwen CodeYesDocker/SeatbeltYesYesYesQWEN.md
AiderNoNoneNoYesNoNone
GooseYesDocker (MCP)YesYesYes.goosehints
OpenHandsYesDockerYesYesYesNone
Continue CLIYesNoneYesYesNo.continue/rules
Cline CLIYesCheckpointsYesYesYes.clinerules
WarpYesNoneNoYesYesWARP.md (reads all)

Warp reads everyone’s memory files: WARP.md, CLAUDE.md, AGENTS.md, and GEMINI.md. If you switch between agents, it just works.

New features to watch

The latest wave of CLI agents added several differentiating features:

FeatureWho has itWhat it does
LSP SupportClaude Code, OpenCode, Crush, ClineLanguage Server Protocol for IDE-grade code intelligence
Skills/Prompt TemplatesClaude Code, Gemini CLI, OpenCode, Pi, Kilo CodeReusable capability packages loaded on-demand
HooksClaude Code, Gemini CLI, Goose, Mistral Vibe, CrushPre/post tool execution event handlers
Voice InputGemini CLI (experimental), Cline, Aider, GooseSpeech-to-text for hands-free coding
Checkpoints/BranchingClaude Code, Plandex, Gemini CLI, Kilo Code, ClineGit-like state snapshots for plan exploration
Multi-agent OrchestrationClaude Code, Roo Code (Boomerang), Claude Squad, EmdashCoordinate multiple specialized agents
Tree-sitterAider, Claude Code, Plandex, Cline, Kilo CodeAST-based code understanding

Sandboxing approaches

I wrote about sandboxing strategies in detail, but here’s the CLI agent reality:

AgentLinuxmacOSNetwork
Claude CodebubblewrapSeatbeltProxy with allowlist
Codex CLILandlock + seccompSeatbeltDisabled by default
Gemini CLIDocker/PodmanSeatbeltProxy
GooseDocker (optional)NoneVia MCP
OpenHandsDockerDockerIsolated
CodelDockerDockerIsolated

Claude Code and Codex CLI both use OS-level primitives. No Docker required. This matters for CLI tools — users won’t install Docker just to use an agent.

How to pick

You want the most features. Claude Code or OpenCode. Sub-agents, hooks, skills, updated almost daily, LSP support. Claude Code has the deepest permission system. OpenCode is open-source with 75+ providers.

You want free. Gemini CLI. 1000 requests/day, no API key, 1M token context, skills, hooks, checkpoints. Hard to beat.

You’re in the OpenAI ecosystem. Codex CLI. OS-level sandboxing, Apache-2.0, written in Rust. Native GPT integration.

You want local models. OpenCode, Aider, or Kilo Code. All support Ollama. Kilo Code has 500+ models; Aider has tree-sitter repo maps.

You’re building your own agent. Pi. Four core tools, great component library, extensions, solid philosophy. A clean base to fork.

You want plan branching. Plandex. Git-like branching for plans, diff sandbox, tree-sitter repo maps.

You love Charmbracelet. Crush. From the Bubble Tea team, written in Go, LSP-aware.

You’re on JetBrains. Junie CLI. JetBrains’ own agent, deeply integrated, works headless in CI.

Thirty-six agents. Four that matter for most people: OpenCode for open-source, Claude Code for features, Gemini CLI for free, Codex CLI for performance.

The rest solve specific problems — browse the full list above.

A year ago, none of this existed. Now there’s a CLI agent for every workflow. Pick one and start shipping.


Full dataset with all 36 agents, features, and metadata: cli-agents.json