Last updated · Santosh Kumar Radha

CodeAF vs Claude Code: open-model harness or Anthropic agent?

CodeAF is an Apache 2.0 coding agent for any model; Claude Code is built around Claude. Models, cost, workflow and when to pick each.

CodeAF by AgentField is an open-source (Apache 2.0) alternative to Claude Code: a terminal coding agent that runs on the model you choose, including open-weight and local models, where Claude Code is Anthropic's agent for Claude models. Pick Claude Code if you want Claude and its wider set of surfaces; pick CodeAF if you want model choice, a lower cost per solved issue on open models, or many handed-off tasks running from one window.

CodeAF is an early preview (the README says so). Its first stable release was v0.5.0 on 2026-09-30 and the current one is v0.8.3, released 2026-11-19. Claude Code is older and more mature: its latest release on 2026-11-21 was v2.1.330, published 2026-11-20.

At a glance

CodeAFClaude Code
LicenceApache 2.0 (LICENSE)Proprietary: "All rights reserved", use under Anthropic's terms (LICENSE.md)
ModelsOpenRouter, ai&, DeepSeek, GLM, Kimi, MiniMax, Qwen, Codex through a ChatGPT plan, Ollama, any OpenAI-compatible endpoint; Claude, GPT and Gemini through OpenRouter (FAQ)Claude models, through Anthropic, Amazon Bedrock, Google Cloud or Microsoft Foundry (docs); Anthropic does not support routing it to non-Claude models (docs)
Local modelsYes: Ollama with no key, llama.cpp or LM Studio as OpenAI-compatible endpoints (FAQ)Not documented
CostFree software; you pay your model provider. $500 daily backstop, $5 per task by default (budgets)Needs a Pro, Max, Ultra, Team, Enterprise or Console account (setup); plan usage or API tokens (costs)
InterfaceTerminal, one Go binary for macOS, Linux and Windows; no IDE extension documentedCLI, desktop, VS Code, JetBrains, web, mobile, Slack and Chrome (platforms)
AutonomyHanded-off work becomes tasks in their own copy of the folder; senior-dev runs up to $10 and 3 hours by default (senior-dev)Subagents, background sessions and agent teams (subagents)
ApprovalsYOLO by default, Alt+A changes it; dangerous commands always ask (FAQ)Six permission modes and an OS sandbox (permission modes)
Headless / CIcodeaf do "..." --json, exit 0, 2 or 3 (headless)claude -p and a GitHub Action (headless, Actions)
MCP and extensions/connect to over a hundred services; Claude Code MCP servers are not imported (connections)MCP client, hooks and plugins (MCP, hooks, plugins)
Project memoryReads CLAUDE.md, AGENTS.md and skills folders in place (FAQ)CLAUDE.md, and it can read AGENTS.md (memory)

The model and licence difference

This is the deciding row. Claude Code is built for Claude, and Anthropic's own gateway page says it "doesn't support routing Claude Code to non-Claude models through any gateway". If you want Claude, that is not a limit, it is the product.

CodeAF is model-agnostic. Bring your own API key (BYOK) for any provider in the list above, sign in with a Codex plan through /connect (Codex plan), or run a local model at no cost. There is no Claude plan sign-in today: Claude models are reached through OpenRouter. Every handed-off task runs on a crew of three seats (a worker, a planner and a checker, which you can think of as subagents), and each seat can be a different model. /crew models open limits every seat to open-weight models (README). For the wider picture, see a coding agent that works with any model.

One window for many tasks

Both tools can run more than one agent. Claude Code has subagents that start in the background, background sessions and agent teams. CodeAF is organized around that way of working from the first screen: work you hand off becomes a task in its own copy of your folder, on its own branch, while you keep talking, and home shows every project and conversation on the machine (FAQ). Teams group conversations under a manager, an agent whose job is the other agents (teams).

For a large change, /senior-dev hands the brief to senior-dev, CodeAF's built-in agent for large coding changes. It works alone on a branch named after the task, runs your project's own build and tests on what it hands in, and never moves your branch.

Cost and quality on the same model

The question "is there a cheaper alternative to Claude Code with similar quality?" has two honest answers.

First, CodeAF can run on cheaper and local models, and Claude Code cannot. Second, on our DeepSWE comparison, every harness ran the full 113-task set on one model, deepseek/deepseek-v4-flash-0731 through OpenRouter, with the official verifiers and 3 hours per task. CodeAF in senior-dev mode solved 62 of 113 (54.9%) at 22 cents a task, first of ten harnesses in that comparison. Claude Code 2.1.267 solved 16 (14.2%) at 19 cents a task, so its cost per solved issue was 3.4x CodeAF's 39.9 cents (method page, arms.csv). Task by task, CodeAF solved 51 tasks Claude Code missed, and Claude Code solved 5 that CodeAF missed (head to head).

Read that row carefully. Both ran DeepSeek V4 Flash, not Claude, through an Anthropic-compatible endpoint, which is not a setup Anthropic supports. It measures the harness on that model, not Claude Code with Claude. It is our own comparison, one pass per harness; CodeAF used the provider's default sampling while the others sent temperature 1.0 and top-p 0.95; and Claude Code 2.1.268, which fixed HTTP 400 errors on third-party Anthropic-compatible endpoints, came out before most of that arm's tasks started.

Using them together

You do not have to choose one. Claude Code can lead and hand the implementation to CodeAF through the shell (playbook):

sh
codeaf do "<a precise brief: what to build, which files, what done means>" --json

It prints one JSON object with ok, stop, answer, files and spend_usd, and exits 0 when done, 2 when it did not finish and 3 when a limit stopped it (headless). For a large change, codeaf senior-dev --max-cost 2 --json "<brief>" runs senior-dev as a subharness. In the other direction, CodeAF reads your CLAUDE.md, the skills in .claude/skills and the skills of enabled Claude Code plugins, where they live. Claude Code's permissions, hooks and MCP servers are not imported; you connect services with /connect instead.

When to choose Claude Code

Choose Claude Code if you want Claude models with a first-party tool and a Claude plan you already pay for. It is more mature, and it reaches more places: a desktop app, VS Code and JetBrains extensions, the web, mobile, Slack and Chrome. Its extension surface is richer, with MCP servers, hooks and plugins, and its permission modes and OS sandbox are documented in detail. If your team reviews changes in an editor, or needs hooks to enforce policy, Claude Code fits better today.

When to choose CodeAF

Choose CodeAF if you want an open-source coding agent under Apache 2.0, a model you pick per task (open-weight, local or a Codex plan), spend limits at every level, and many tasks running in isolated copies from one terminal. It also fits remote boxes: codeaf chat --host devbox runs the conversation on a machine you reach over your own ssh (remote machines). If you are also weighing OpenAI's agent or another open one, see CodeAF vs Codex CLI and CodeAF vs OpenCode.

Questions people ask

What is an open-source alternative to Claude Code?

CodeAF is an open-source (Apache 2.0) terminal coding agent by AgentField that works with OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, Ollama, a Codex plan or any OpenAI-compatible endpoint. Other open-source options engines name are OpenCode, Codex CLI and Aider. Claude Code itself is proprietary: its repository's licence reads "All rights reserved".

Is there a cheaper alternative to Claude Code with similar quality?

CodeAF itself is free, and you pay only your model provider, or nothing with a local model. On our DeepSWE comparison, with every harness on DeepSeek V4 Flash, CodeAF in senior-dev mode solved 62 of 113 tasks against Claude Code's 16, at 39.9 cents per solved issue against 3.4x that. That run used DeepSeek, not Claude, so it does not tell you how Claude Code performs with Claude.

Can CodeAF use my Claude subscription?

No, not today. CodeAF has no Claude plan sign-in; Claude models are reached through OpenRouter with an API key. A ChatGPT or Codex plan can sign in through /connect.

Does CodeAF read my CLAUDE.md and Claude Code skills?

Yes. CodeAF loads a project's CLAUDE.md and AGENTS.md into its instructions and reads skills in .claude/skills and the skills of enabled Claude Code plugins where they live, copying nothing. Claude Code's permissions, hooks and MCP servers are not imported.

Can Claude Code hand work to CodeAF?

Yes, through the shell. Claude Code can run codeaf do "<brief>" --json for a general job or codeaf senior-dev --json -- "<brief>" for a large change, then read the JSON and the exit code. The Claude Code plans, CodeAF builds playbook shows the one CLAUDE.md rule that sets this up.

Does CodeAF have a VS Code extension like Claude Code?

No. CodeAF runs in the terminal, locally or on a remote machine over ssh, and no IDE extension is documented. If you want the agent inside VS Code or JetBrains, Claude Code covers that today.

Sources

Last updated 2026-11-21. Checked against CodeAF v0.8.3 and Claude Code v2.1.330; the DeepSWE rows are CodeAF senior-dev mode and Claude Code 2.1.267.