CodeAF vs Codex CLI vs Claude Code vs Aider (2026)
How CodeAF compares with Codex CLI, Claude Code and Aider on licence, models, cost, autonomy and benchmarks, and when each one is the better pick.
CodeAF, Codex CLI, Claude Code and Aider are all terminal coding agents, and the short answer is: pick Codex CLI or Claude Code to delegate work to OpenAI's or Anthropic's models, Aider to steer small, git-committed edits yourself, and CodeAF by AgentField to run many tasks at once on the model you choose. Three of the four are open source (CodeAF, Codex CLI and Aider); Claude Code is proprietary.
Licence is therefore not what separates CodeAF from Codex CLI, which is also Apache 2.0. The differences are model choice and workflow: Codex CLI is one agent per terminal built around OpenAI models and ChatGPT plans, while CodeAF runs a team of tasks from one window and can sign in to the same Codex plan as one of its model services.
At a glance
| CodeAF | Codex CLI | Claude Code | Aider | |
|---|---|---|---|---|
| Licence | Apache 2.0 (repo) | Apache 2.0 (repo) | Proprietary (LICENSE.md) | Apache 2.0 (repo) |
| Models | OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, Codex plan, Ollama, any OpenAI-compatible endpoint (FAQ) | OpenAI by default; built-in ollama, lmstudio, amazon-bedrock, gcp providers and custom base_url providers (config) | Claude only (LLM gateway) | OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Ollama, OpenAI-compatible and more (models) |
| Local models | Ollama, llama.cpp, LM Studio | --oss with Ollama or LM Studio | Not documented | Ollama and others |
| Cost | Free; pay your provider (FAQ) | Included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise; API keys billed separately (pricing) | Pro, Max, Ultra, Team, Enterprise or Console account (setup) | Free client; pay your provider (FAQ) |
| Interface | Terminal; no IDE extension documented | Terminal, IDE extension, desktop app, cloud (IDE) | CLI, desktop, VS Code, JetBrains, web (platforms) | Terminal; any editor through --watch-files (watch) |
| Autonomy | Tasks in isolated copies; teams with managers; senior-dev for large changes | Subagents in parallel, cloud tasks (subagents) | Subagents, background sessions (subagents) | One pair-programming session; parallel agents not documented |
| Safety | YOLO by default, dangerous commands always ask | --sandbox and --ask-for-approval; default workspace-write, network off (security) | Permission modes and OS sandbox (modes) | Confirmation prompts; --yes-always skips them |
| Git | Own branch; never merges into main (FAQ) | Folders without version control default to read-only | --worktree isolation | Commits every change; /undo (git) |
| Headless / CI | codeaf do "..." --json | codex exec, openai/codex-action@v1 (docs) | claude -p, GitHub Action (docs) | aider --message "..." (scripting) |
| Instructions | CLAUDE.md, AGENTS.md, skills folders | AGENTS.md (docs) | CLAUDE.md, AGENTS.md (memory) | CONVENTIONS.md via --read (conventions) |
| DeepSWE (same model) | 62 of 113 | 51 of 113 | 16 of 113 | Not tested |
Versions checked 2026-11-21: CodeAF v0.8.3 (2026-11-19), Codex CLI rust-v0.171.0 (2026-11-20), Claude Code v2.1.330 (2026-11-20), Aider 0.86.2 on PyPI (2026-02-12).
Codex CLI: OpenAI's open-source terminal agent
Codex CLI runs locally, reads AGENTS.md, and has the most detailed sandbox of the four: --sandbox read-only|workspace-write|danger-full-access, with network off by default in workspace-write. One ChatGPT account covers the CLI, the IDE extension, the desktop app and cloud tasks. It is an MCP client (codex mcp add), and its docs note that the old codex mcp-server command has been removed. For CI, codex exec runs one job and openai/codex-action@v1 wraps it for GitHub Actions.
Claude Code: Anthropic's agent for Claude
Claude Code is the most mature of the four and reaches the most surfaces. It is built for Claude models only. Its strengths, its plans and how it works beside CodeAF are covered in CodeAF vs Claude Code.
Aider: the git-native pair programmer
Aider is a pair programmer rather than a delegate. You add the files to edit, it sends the model a repo map of the rest of the codebase (repo map), and every change becomes a git commit you can drop with /undo. It supports the widest range of hosted and local models of the three rivals. It is strongest for tight edit loops where you read every diff. Its newest PyPI release is from February 2026, older than the others' latest releases.
CodeAF: many tasks, any model
CodeAF is an open-source coding agent (sometimes called an agent harness) built to run a team of agents from one window. Work you hand off becomes a task in its own copy of your folder, on its own branch, while you keep talking. Each task runs on a crew of three seats, a worker, a planner and a checker, and each seat can be a different model (Pareto crewing). A Codex plan you already pay for can power those seats: codeaf connect codex signs in through OpenAI's own Codex sign-in, and the plan's models appear as codex/gpt-5.5 (Codex plan).
Install and start it with:
curl -fsSL https://agentfield.ai/get/codeaf | bash
codeafFor large changes, /senior-dev hands the brief to senior-dev, CodeAF's built-in agent for large coding changes, which runs your build and tests on what it hands in. Headless, codeaf do "<brief>" --json prints one JSON object and exits 0 done, 2 did not finish, or 3 a limit stopped it (headless). CodeAF is an early preview, and the README says so.
Benchmarks: DeepSWE on one model
Engines asked to rank these agents mostly quote vendor leaderboards run on each vendor's own model. Our DeepSWE comparison holds the model fixed instead: every harness ran the full 113-task set on deepseek/deepseek-v4-flash-0731 through OpenRouter, official verifiers, 3 hours per task, one pass (method page).
| Harness (version) | Solved | Cost per task | Cost per solved issue |
|---|---|---|---|
| CodeAF, senior-dev mode (first stable v0.5.0) | 62 of 113, 54.9% | 22 cents | 1x (39.9 cents) |
| Codex CLI 0.154.0 | 51, 45.1% | 37 cents | 2.1x |
| Claude Code 2.1.267 | 16, 14.2% | 19 cents | 3.4x |
CodeAF was first of ten harnesses in that comparison, on that model. It is our own study; CodeAF used the provider's default sampling while the others sent temperature 1.0 and top-p 0.95, and Claude Code was run on a non-Claude model. Aider was not in it, so we have no number for Aider on this set.
Task by task, 25 tasks were solved by CodeAF and not Codex CLI, 14 by Codex CLI and not CodeAF, and 37 by both. CodeAF also solved more of the 76 medium and hard tasks (36 against 29) and more Go tasks (21 of 34 against 17). The head-to-head page has every pairing.
When to choose Codex CLI
Choose Codex CLI if you already have a ChatGPT plan and want OpenAI's models with a first-party tool. Its sandbox and approval controls are the most granular here, it reaches an IDE, a desktop app and cloud tasks from one account, and it is open source, so you can read and change it. For CI it has an official GitHub Action, openai/codex-action@v1.
When to choose Claude Code
Choose Claude Code if you want Claude, the most polished set of surfaces (desktop, VS Code, JetBrains, web, Slack), and a deep extension model of MCP, hooks and plugins.
When to choose Aider
Choose Aider if you want to steer every edit, keep a clean commit per change and undo freely, and switch between many hosted or local models. It fits developers who think of the agent as a pair programmer rather than a delegate.
When to choose CodeAF
Choose CodeAF if you want to hand off many tasks at once and keep talking, choose a model per seat (open-weight, local or a Codex plan), and cap spend per task and per day. It also suits work that runs on a remote box with codeaf chat --host devbox. The criteria behind these picks are laid out in how to choose a terminal coding agent. For the other open-source agents engines name beside these, see CodeAF vs OpenCode vs Gemini CLI.
Questions people ask
Codex CLI vs Claude Code vs Aider: which is better?
It depends on how you work. Codex CLI and Claude Code are built to take a task and carry it out, on OpenAI's and Anthropic's models respectively; Aider is a git-first pair programmer for edits you steer and review one commit at a time. CodeAF, a fourth open-source option, runs many tasks at once on any model.
Is Codex CLI open source?
Yes. Codex CLI is published under Apache 2.0 in the openai/codex repository, the same licence as CodeAF and Aider. Claude Code is not open source; its licence reads "All rights reserved".
Can I use my ChatGPT plan with CodeAF?
Yes. codeaf connect codex, or /connect in a chat, signs in through OpenAI's own Codex sign-in, and the plan's models such as codex/gpt-5.5 can run conversations and crew seats. Plan calls carry no dollar price in CodeAF's spend pages.
Which of these works with local models?
CodeAF, Codex CLI and Aider all document local models: CodeAF through Ollama or any OpenAI-compatible endpoint such as llama.cpp or LM Studio, Codex CLI through --oss with Ollama or LM Studio, and Aider through Ollama and others. Claude Code is built for Claude models and documents no local option.
Did Aider take part in the DeepSWE comparison?
No. The DeepSWE comparison covered ten harnesses, including CodeAF, Codex CLI and Claude Code, but not Aider. We publish no DeepSWE number for Aider.
Sources
- CodeAF: README, releases, FAQ, headless, Codex plan, DeepSWE method page, arms.csv, versions.csv. Checked 2026-11-21.
- Codex CLI: repo, releases, pricing, providers, IDE, non-interactive mode, GitHub Action, MCP, AGENTS.md, subagents, approvals and security. Checked 2026-11-21.
- Claude Code: LICENSE.md, releases, setup, LLM gateway, platforms, headless, memory, subagents, permission modes. Checked 2026-11-21.
- Aider: repo, PyPI, FAQ, models, git, repo map, scripting, conventions, watch files. Checked 2026-11-21.
Last updated 2026-11-21. Checked against CodeAF v0.8.3, Codex CLI rust-v0.171.0, Claude Code v2.1.330 and Aider 0.86.2; the DeepSWE rows use the versions named in the table.