Agent-Field/CodeAF

CodeAF is a coding harness aimed at people running open models (DeepSeek, Qwen, GLM, Kimi, MiniMax, Ollama, or any OpenAI-compatible endpoint) who want many agent tasks moving at once. Instead of one agent per terminal, it presents one window across every project on the machine, splits a request into tasks in separate repository copies and branches, checks results, and merges only what passed. It is in early preview, according to the maintainers.

How work flows

You describe problems conversationally, for example a retry test failing on CI, a mobile layout bug and an expiring deploy key; each item can become its own task and branch. Work that passes its existing check lands on your branch automatically, never on main, dev or a release branches. Work nothing could verify waits for 1 accept or 2 not right. It surfaces only blocking questions under needs you.

Three built-in specialists run as tasks: /senior-dev for changes plus build and test results, /sec for local security audits (Markdown, JSON, SARIF), and /review for GitHub pull requests. The senior-dev specialist runs on macOS and Linux without extra installation. The README notes that review and sec read source without running your tests, build, app or exploits and leave code unchanged; their reports should be read for coverage and incomplete analysis, and no findings cannot guarantee absence of security problems. The DeepSWE benchmark measures senior-dev only; the review specialist's imported PR-AF benchmark results belong to the original project, not to CodeAF's version.

Models, automation and headless use

Tasks use an auto-selected crew of worker, planner and checker seats drawn from connected providers, with a per-task spend limit of $5 unless changed and an optional daily cap. /crew lets you pin a seat or restrict every seat to open-weight models. Reminders, schedules and watches run while CodeAF is open; anything missed while it was closed runs once on the next start.

For CI, cron and scripts, codeaf do "..." --json takes your brief, plans, runs and checks it, then prints one JSON object and an exit code, resolving questions with recorded assumptions instead of asking. Related commands are exec (one worker, no plan) and run (execute an edited plan). The headless contract is documented separately.

Remote use and requirements

codeaf chat --host devbox keeps the conversation on the machine that owns the work while the interface runs locally, with your existing ssh configuration; only CodeAF needs to exist there. Closing the link does not stop work, and a dropped connection redials for five minutes. A phone can join the same conversation through a mobile terminal over ssh; the README says no app is involved. Reaching a machine without ssh, via a relay and pairing code, is built but switches on when the hosted relay becomes available.

Installation is a single Go binary for macOS, Linux and Windows, placed at ~/.codeaf/bin by the install script, or built from source. First start connects OpenRouter in a browser or accepts a key; Codex signs in with a ChatGPT plan, several providers take keys, and Ollama needs none.

Privacy and status

Anonymous usage counts are on by default: version, OS, mode, session counts, errors and total token counts under a random install id, with prompts, code, paths and model names explicitly excluded. The Model Pool is also on by default and sends text-free numeric measurements about models; it can be switched off with settings, CODEAF_MODEL_POOL=off, or CODEAF_TELEMETRY=off. The telemetry disclosure points to a test that holds the document to the code.

The project is Apache-2.0 licensed and the maintainers call it an early preview with rough edges; it has 596 stars as of the repository listing.

Project source and documentation