What is CodeAF?
CodeAF, from AgentField AI, is a terminal coding agent written in Go and shipped as one binary of about 53 MB. It is tuned for open models such as DeepSeek, GLM, Kimi, MiniMax and Qwen, through OpenRouter, direct keys, Ollama or any OpenAI-compatible endpoint, and can also sign in to a ChatGPT plan for Codex models.
Its pitch is a software factory rather than a pair programmer: one window lists every project, a request can fan out into tasks on their own branches, a planner, worker and checker model are chosen per task, and work that passes its check merges onto your branch. Built-in specialists cover coding (senior-dev), local security audits and GitHub pull request review, and a headless do command serves CI and scripts.
The repo is Apache-2.0, created in August 2026, with about 605 stars. It is labelled an early preview; the latest release is v0.8.2 (23 October 2026).
Is CodeAF worth it?
CodeAF is built on the assumption that cheap open models can do most coding work if the harness around them plans, checks and retries well. Its answer is a crew per task, with a planner, a worker and a checker each picked from your connected providers, plus a $5 per-task spending cap by default and an optional daily cap. Work that passes its check lands on your branch, never on main, dev or a release branch; work nothing could verify waits for you.
The benchmark claims need context. The README reports senior-dev solving 54.9% of 113 DeepSWE issues on DeepSeek V4 Flash, ahead of nine other harnesses on the same model and at the lowest cost per solved issue. That is a single vendor-run comparison on one model, not an independent leaderboard, and the site's "3.9x as accurate as Claude Code" line should be read in that light. The memory figures, 27 MB per added session against 216 MB for claude, are easier to believe and matter if you run many sessions on one box.
The data defaults are opt-out. Anonymous usage counts go to AgentField by default, and a Model Pool shares text-free scores about the models you ran, also on by default; CODEAF_TELEMETRY=off or DO_NOT_TRACK=1 stops both. Neither sends code or prompts, according to the telemetry document, but teams with strict egress rules should set the variable before first run.
At about 605 stars and 298 open issues and pull requests, with one contributor responsible for most commits, this is early. Compared with OpenCode or Kilo Code it is less proven, but its fan-out tasks, remote attach over ssh and headless mode go further than either for someone juggling several repositories.
Where it earns its place
A solo developer paying for DeepSeek through OpenRouter
Fix a batch of small bugs across three repositories without three terminals
Describe the bugs in one message; CodeAF turns each into a task on its own branch and merges those that pass their checks. Spend is capped at $5 per task unless changed.
A platform team running nightly maintenance
Bump dependencies and open changes from CI
Call codeaf do with a brief, a timeout and --json, and read one JSON object and an exit code. Set CODEAF_TELEMETRY=off in the CI environment first.
A tech lead reviewing contractor pull requests
Get a first-pass review and local security audit
Run /review owner/repo#123 and /sec changes since main; both read code without running it and produce Markdown or JSON reports. A clean report does not prove the code is safe.
Before you commit
- Setup
- curl install script puts one binary in ~/.codeaf/bin; first start connects OpenRouter in the browser or takes a key. Built-in specialists run on macOS and Linux.
- Integrations
- OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, Ollama, any OpenAI-compatible endpoint and Codex through a ChatGPT plan; GitHub pull requests for review; ssh for remote attach.
- Data and privacy
- Anonymous usage counts and Model Pool scores are sent by default; no prompts, code, paths or keys according to the telemetry document. CODEAF_TELEMETRY=off or DO_NOT_TRACK=1 stops both. Prompts go to whichever model provider you connect.
- Cost at scale
- No paid edition. Cost is model usage; the default cap is $5 per task with an optional daily cap, and cheap open models are the design target.
- Limits
- Early preview; 298 open issues and PRs; the hosted relay for reaching machines without ssh is not switched on yet; benchmark claims are vendor-run.
What’s good
- Tuned for open models with per-task model selection
- Per-task and daily spend caps
- Parallel tasks on separate branches, never merged to main
- Small Go binary; headless mode for CI
What’s not
- Benchmark results are self-reported on one model
- Telemetry and Model Pool on by default
- Early preview with 298 open issues and PRs
- One maintainer writes most of the code
Best for
Developers who want to run coding agents on cheap open models across several repositories and are happy with early-preview software.
CodeAF alternatives
CodeAF: common questions
- Is CodeAF worth it?
- An ambitious harness that takes open models seriously, with headline benchmarks that are its own and telemetry you have to switch off. We rate it 3.6 out of 5.
- How much does CodeAF cost?
- Free and open source (Apache-2.0). You pay your model provider; default per-task spend cap is $5.
- What are the downsides of CodeAF?
- Benchmark results are self-reported on one model; telemetry and Model Pool on by default; early preview with 298 open issues and PRs; one maintainer writes most of the code.
- Who is CodeAF best for?
- Developers who want to run coding agents on cheap open models across several repositories and are happy with early-preview software.
- How much does CodeAF cost for a team or in production?
- No paid edition. Cost is model usage; the default cap is $5 per task with an optional daily cap, and cheap open models are the design target.
- Is CodeAF safe for company data?
- Anonymous usage counts and Model Pool scores are sent by default; no prompts, code, paths or keys according to the telemetry document. CODEAF_TELEMETRY=off or DO_NOT_TRACK=1 stops both. Prompts go to whichever model provider you connect.
- What are the best CodeAF alternatives?
- The closest alternatives we have reviewed are OpenCode, Kilo Code, Aider and OpenHands.
Answers checked Oct 31, 2026.