🆕 for more detailed information visit the 🌐Community Guide to Open-Source AI Software Craft - oss-ai-swe.org
| Tool | Stars* | Licence | Default models | What it can do out-of-the-box | Ideal use-case |
|---|---|---|---|---|---|
| Codex CLI | 30.5 k (github.com) | Apache-2.0 | OpenAI o3 / o4-mini (or any Responses-API compatible provider) (github.com) | • Reads + writes files • Runs shell commands in a sandbox with approval policies • Multimodal prompts (images) |
When you want ChatGPT-level reasoning and the ability to run code locally with fine-grained security gates. |
| Gemini CLI | 54.5 k (github.com) | Apache-2.0 | Google Gemini 2.5 Pro (free quota) or any Vertex AI/AI Studio key (blog.google) | • 1 M-token code window • Multimodal generation (PDFs ↦ apps, sketches, video via Imagen/Veo) • Built-in Google Search grounding |
When you live in Google’s ecosystem and need multimodal generation or very large context windows. |
| OpenHands | 60 k (github.com) | MIT | Pluggable (Claude 3.5 Sonnet recommended, but any via tools/MCP) (aiagentslist.com) | • “Do-anything” agent: edit code, run commands, browse web, call APIs • Headless/CLI / Docker / Cloud modes |
Delegating full-stack dev tasks to an agent that can touch anything—good for automation pipelines. |
| OpenCode | 7 k (github.com) | MIT | Multiple providers (OpenAI, Anthropic, Gemini, Bedrock, Groq, etc.) (github.com) | • Bubble-Tea TUI • Integrated Vim-like editor, LSP, file-change visualiser • Named-argument custom commands |
Terminal-first devs who want a chat/TUI hybrid with rich session management but light autonomy. |
| Aider | 35.2 k (github.com) | Apache-2.0 | Any chat-completion API or local model; ships with presets for GPT-4o, Claude 3.7, DeepSeek, etc. (aider.chat) | • Builds a semantic map of your repo for large-project edits • Auto-commits with git • IDE plug-ins, voice-to-code, lint + test loop |
Pair-programming flow where you stay in control: incremental, review-friendly commits on big codebases. |
*GitHub star counts on 7 July 2025.
| Aspect | Codex CLI | Gemini CLI | OpenHands | OpenCode | Aider |
|---|---|---|---|---|---|
| Command execution | Opt-in or --full-auto; sandbox blocks network by default (github.com) |
Requires explicit opt-in to --exec; uses Docker sandbox for “dangerous” commands |
Full autonomy is the default; can be scoped with policies but expects trust in the agent (aiagentslist.com) | Runs commands only when you invoke :run or approve |
Never runs shell commands itself; only edits files and leaves execution to you |
Take-away: If you need strict guard-rails, Codex CLI and OpenCode expose the most granular approval controls; OpenHands is the most autonomous.
- Gemini CLI has the largest first-party context (1 M tokens) and deepest multimodal hooks (Search, Imagen, Veo) (github.com).
- Codex CLI reaches 192 k tokens via codex-1 and supports AGENTS.md for repository-specific tips (openai.com).
- Aider keeps its own code graph so it scales beyond the model window and stays performant on monorepos (aider.chat).
| Install | Notes |
|---|---|
brew install codex or download binary (github.com) |
Includes Rust + TypeScript deps; auto-login for ChatGPT Plus |
npm i -g @google/gemini-cli or npx run (github.com) |
Node 20+ required; Google sign-in grants free quota |
pip install openhands (or Docker compose) |
CLI, headless server and browser UI come in one bundle |
Single static binary (opencode) or go install … |
Cross-platform; no external runtime after build |
pip install aider-install && aider-install (aider.chat) |
Python-only; works offline with local Ollama models |
| Tool | What it nails | Where it still hurts (July 2025) |
|---|---|---|
| Codex CLI | Tight integration with OpenAI ecosystem; robust sandbox; recipes for CI | Windows needs WSL2; remote agent (Codex Web) still faster for long tasks (openai.com) |
| Gemini CLI | Huge context + native multimodal; generous free quota | Nightly builds can break; Docker sandbox flaky on macOS (github.com) |
| OpenHands | End-to-end agent that can really run your stack; micro-agent architecture | Terminal-compat warnings and other UX quirks; steep trust/safety learning curve (github.com) |
| OpenCode | Slick Bubble-Tea UI and file-diff visualiser; easiest to extend in Go | Early-dev; API subject to breaking changes (maintainer note in README) (github.com) |
| Aider | Git-aware, commit-oriented workflow; code map keeps context fresh | Doesn’t execute code itself—still need to run tests manually if your repo lacks CI hooks |
- Need maximum autonomy (agent handles everything): pick OpenHands.
- Prefer chat-style edits & explicit control: go with Aider.
- Already an OpenAI subscriber and care about sandboxing: Codex CLI.
- Google stack, multimodal assets, huge context windows: Gemini CLI.
- TUI enthusiasts who like Bubble-Tea & Go: OpenCode.
- All five tools can talk to any Responses/Chat-Completions compatible endpoint (OpenRouter, Ollama, Groq) once configured—handy if you want to standardise model spend across your org.
- Model Context Protocol (MCP) support matters if you intend to chain tools: Codex CLI and Gemini CLI ship with MCP clients; OpenHands can host an MCP server so other agents (even Claude Code) can offload work to it (github.com).
Bottom line: they occupy different points along the control ↔ autonomy spectrum. Try the one whose philosophy matches how you like to work—and keep a second one around for tasks where its strengths shine.
Generated by Eleanor ( ai.intellectronica.net ) using o3.