FAQ

A coding agent in your terminal, built to run a software team of agents from one window. You talk to it, and work you hand off becomes tasks, teams with managers, specialists such as senior-dev, and standing orders, across every project on the machine. The Tour shows the ten ideas in a few seconds each.

No. Ask CodeAF in plain words: it looks the answer up in its own manual, which ships inside it, and it can set things up for you. We typed "set up a watch that runs go test every hour": it read the manual, put up a card titled hourly go test with every hour, and after 1 it replied that go test ./... runs every hour and only speaks up if a run fails. Questions such as "what keys open the wall?" work the same way.

Yes. CodeAF is open source under the Apache 2.0 licence, and CodeAF itself is free. You pay your model provider for the tokens you use, or use a plan you already have, or a local model at no cost.

macOS, Linux and Windows, on x86-64 and ARM, as one small binary with nothing else to install. The install line runs in a Unix shell. senior-dev needs a Unix shell and is left out of Windows builds.

Those tools are built around one agent working with you in one terminal, and they do that well. CodeAF is built for a different way of working: a team of agents run from one window, with managers that run other agents, specialists built for one job, Pareto crewing that picks the model for each seat, and standing orders that keep running after you stop typing.

The start feels familiar: you type to an agent in your terminal, / opens the commands, and ? in an empty box lists every key. The difference is what happens to work you hand off: it 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.

Yes. A project's CLAUDE.md and AGENTS.md are loaded into the conversation's instructions. Skills in .claude/skills, .codex/skills, .agents/skills, .cursor/skills and .gemini/skills, and the skills of enabled Claude Code plugins, are read where they live, and nothing is copied.

Only the skills. Claude Code's permissions, hooks and MCP servers are not imported; connect services with /connect instead. There is no Claude plan sign-in today: Claude models are reached through OpenRouter, and a ChatGPT plan can sign in through Codex.

Yes, through the shell. Claude Code can run codeaf do "<brief>" --json for a general job, or codeaf senior-dev --json -- "<brief>" to hand a large change to senior-dev as a subharness, then read the JSON it prints. We had Claude Code do exactly that: senior-dev finished on its own branch with "status": "pass" for about two cents.

OpenRouter, DeepSeek, GLM, Kimi, MiniMax, Qwen, a ChatGPT plan through Codex, Ollama, and any OpenAI-compatible endpoint. Claude, GPT and Gemini models are available through OpenRouter.

Yes. On first start CodeAF connects OpenRouter in your browser, which makes a key for you, or takes a key you paste; the key stays on your machine. codeaf connect lists every model service and whether it is connected.

Every task gets a crew of three seats: a worker that does the work, a planner and a checker. Each seat gets the model that fits the task, at the point where more money stops buying quality, and you can pin any seat yourself.

Every call is priced before it runs. There is a daily limit ($500 by default, as a backstop; /budget default 40 sets your own), a per-task limit ($5 unless you set one), and caps for each team; senior-dev stops at $10 and 3 hours by default.

Yes. Ollama needs no key, and any OpenAI-compatible endpoint works, such as llama.cpp or LM Studio, so model calls stay on your machine. A local model costs nothing.

A team is a named, coloured group of conversations around one outcome. A manager is an agent whose job is the other agents: it hands out work, settles what it can, and asks you only what it cannot decide.

Yes. A manager can split a team under a second manager, and one global manager can sit over every team you run, so a question climbs the chart and reaches you only when no manager can settle it.

Closing the terminal stops nothing: the conversation and its tasks run in the folder's engine, and reminders, schedules and watches fire from a 5-minute timer with no window open. Closing the lid puts the machine to sleep, and nothing runs while it sleeps. For work that must go on then, run it on a machine you can ssh into: codeaf chat --host devbox keeps the conversation on that machine, and you can come back from your laptop or your phone.

A harness built for one job, with its own engine, its own process and its own checks: it delivers what it promised or is marked incomplete. senior-dev, the coding specialist, is here today. A code-review specialist based on PR-AF is coming, and so is creating your own.

For a large, well-specified change: a fix whose cause spans files, a feature with its tests, a migration. Type /senior-dev and the brief, and it works alone on its own branch, then runs your project's own build and tests on what it handed in. In our study on DeepSWE it solved the most issues of ten harnesses on the same model.

Handed-off work runs in its own copy of your folder, so your checkout does not move while it runs. At the end it is merged into your folder, or kept on its own branch when you are on a branch such as main that tasks never merge into. senior-dev always works on a branch of its own.

A new conversation starts in YOLO unless you saved another choice: ordinary commands run without asking, and Alt+A changes the approvals. A short list of dangerous commands, and anything sent in your name such as email, Slack or calendar events, always asks first.

Your prompts and code go to the model provider you connect. CodeAF also sends anonymous usage counts, never your code, prompts, paths or keys (CODEAF_TELEMETRY=off stops them), and Model Pool numbers about how models did (model_pool = off stops them).

The docs can run slightly ahead of a release. Try the dev build, which gets new features first: curl -fsSL https://agentfield.ai/get/devaf | bash installs devaf beside codeaf.

Run codeaf update --dev, or type /update dev in a chat. codeaf update --stable goes back.