# CodeAF

> Open-source terminal software factory that plans, runs and checks coding tasks in parallel on open models, with built-in build, security audit and PR review specialists.

CodeAF is an open-source coding harness from AgentField AI, written in Go and shipped as a single binary. It gives developers one terminal window to hand off coding work across projects, see what is running and answer only the questions that need them. The GitHub README labels it an early preview.

## What It Is

CodeAF is a coding agent harness built for open models. A request is split into tasks, each run in its own copy of the repository on its own branch, and work that passes an independent check lands on the user's branch (never main, dev or a release branch). A home view lists every project and conversation on the machine, with questions waiting on the user, running work and spend.

## Built-in Specialists

Three specialists run as tasks beside the conversation. /senior-dev implements a coding change and reports build and test results. /sec audits local code and produces security findings with Markdown, JSON and SARIF reports. /review reads a GitHub pull request and reports findings with file locations and suggested fixes. Security and review read source without running tests or exploits and leave code unchanged. The README says review and sec are imported from PR-AF and SEC-AF.

## Models, Crews and Automations

CodeAF works with OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, Codex through a ChatGPT plan, Ollama and any OpenAI-compatible endpoint. Each task runs on a crew of worker, planner and checker seats chosen per task, which can be pinned or limited to open-weight models. Automations let users set reminders, scheduled work and watches in plain language. A headless mode (codeaf do ... --json) returns one JSON result and an exit code for CI, cron and scripts.

## Remote Work and Safety

The conversation can run on a remote dev box over the user's own SSH while the screen stays local. Only reads, searches and git status run without asking; destructive commands still ask even with --yolo. Furrow, which ships with CodeAF, can rewind files git does not track. Per-task spend limits apply, and opt-out anonymous telemetry and a Model Pool that sends no code are documented. The vendor reports that senior-dev solved 54.9% of 113 DeepSWE tasks at the lowest cost per solve among ten harnesses on DeepSeek V4 Flash.

## Features
- One window for every project with a home view of what needs you, what is running and spend
- Talk-to-tasks planning with parallel tasks on separate branches and independent checks before merge
- Built-in /senior-dev, /sec and /review specialists
- Security audit reports in Markdown, JSON and SARIF
- GitHub pull request review with findings and file locations
- Per-task model crews (worker, planner, checker) with pinning and open-model-only limits
- Automations for reminders, scheduled work and watches in plain English
- Headless mode with JSON output and exit codes
- Remote execution over SSH with reattachable conversations
- Rewind of files via furrow, including untracked files
- Loads AGENTS.md, CLAUDE.md and existing skills folders
- Spend limits and approval prompts for non-read actions

## Integrations
OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, Ollama, Codex, OpenAI-compatible endpoints, GitHub, Slack, Gmail, Notion, Linear, Jira, GitLab, Sentry, Datadog, Supabase, Cloudflare, Stripe, Airtable

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, CLI

## Pricing
Open Source

## Version
v0.8.2

## Links
- Website: https://agentfield.ai/codeaf
- Documentation: https://agentfield.ai/docs/codeaf
- Repository: https://github.com/Agent-Field/CodeAF
- EveryDev.ai: https://www.everydev.ai/tools/codeaf
