CodeAF is an open-source terminal coding harness from AgentField built to get frontier-grade results from open-weight models.
It works as a "software factory" on your machine: one window for every project, requests split into parallel tasks on their own branches, and finished work merged only after it passes a check.
- 01
Single Go binary installed with one curl script; no other services to run
- 02
One window listing every project and conversation, with per-conversation approval rules, models and spend limits
- 03
Turns natural-language requests into parallel tasks, each in its own repository copy and branch
- 04
Auto-selected model crew per task (worker, planner, checker) with /crew pinning and open-weight-only mode
- 05
Work that passes its check lands on your branch automatically, never on main or release branches
- 06
Subharnesses: specialist workflows such as /senior-dev, plus custom ones you can design and save
- 07
Built-in providers: OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, Codex via ChatGPT plan, Ollama and OpenAI-compatible endpoints
Open Source
CodeAF is Apache-2.0 and free to install as a single Go binary. You pay only for model usage through your own providers (OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen or any OpenAI-compatible endpoint), can sign in Codex with a ChatGPT plan, or run local models via Ollama at no API cost. Per-task spend limits default to $5 with an optional daily cap.
Best for
Developers who want to run several coding tasks across multiple repositories from one terminal window while keeping costs low with open-weight models.
Not ideal for
Teams that need a mature, stable tool today: CodeAF is an early preview with expected rough edges, and its anonymous model-pool telemetry is on by default (it can be turned off).
CodeAF is an ambitious, Apache-2.0 alternative to running one terminal per agent, with automatic task splitting, checking and model selection aimed at cheap open models. It is promising for cost-conscious multi-project work, but still early, so expect change and verify its benchmark claims for your own repos.
Early preview under rapid development, with v0.7.x releases and daily dev builds in October 2026. Its /senior-dev subharness is reported first of ten harnesses on a DeepSWE run using DeepSeek V4 Flash; a native PR-AF code-review subharness is listed as coming soon.
As of
Is CodeAF free?+
Yes - CodeAF is Open Source. CodeAF is Apache-2.0 and free to install as a single Go binary. You pay only for model usage through your own providers (OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen or any OpenAI-compatible endpoint), can sign in Codex with a ChatGPT plan, or run local models via Ollama at no API cost. Per-task spend limits default to $5 with an optional daily cap.
Is CodeAF open source?+
Yes - CodeAF is open source. CodeAF is Apache-2.0 and free to install as a single Go binary. You pay only for model usage through your own providers (OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen or any OpenAI-compatible endpoint), can sign in Codex with a ChatGPT plan, or run local models via Ollama at no API cost. Per-task spend limits default to $5 with an optional daily cap.
Who is CodeAF best for?+
Developers who want to run several coding tasks across multiple repositories from one terminal window while keeping costs low with open-weight models.
Who is CodeAF not ideal for?+
Teams that need a mature, stable tool today: CodeAF is an early preview with expected rough edges, and its anonymous model-pool telemetry is on by default (it can be turned off).
What are the best CodeAF alternatives?+
The closest CodeAF alternatives on ai.dosa.dev are Claude Code, Codex CLI, Amp - all listed under Terminal & CLI Agents.
Who makes CodeAF?+
CodeAF is developed by AgentField. It is listed in the Terminal & CLI Agents category on ai.dosa.dev.
Favorite this tool to revisit it later, or Zap it to contribute to the public vote count.