nac: Hand Ambitious Long-Running Tasks to AI Agents Without Losing Focus

Who It's For

• Users who want to hand large, long-running tasks to an AI agent without losing control: nac is an open-source agent runtime framework from Arcee built on a thread-and-episode architecture — a central orchestrator handles planning and breaking down work, but it cannot execute commands or edit files itself; it can only spawn threads. When each thread finishes, it returns an "episode," a structured work summary telling you exactly what it did.
• Users already working with other agent harnesses who want to hook in nac: the official description says nac provides a portable onboarding skill and MCP integration — paste the official onboarding snippet into your agent tool of choice, and it installs nac, completes configuration, and connects MCP.
• Not suitable as a direct command-line executor: the orchestrator cannot run commands or edit files for you; all execution happens inside spawned threads.
• The evidence does not state which specific agent clients are supported, platform compatibility, or resource requirements — these remain unknown.

Your First Attempt in 3 Minutes

Recommended main path: official installer → device-code login → local dashboard.

1. Prerequisites: the evidence does not state the required dependencies or a copy-paste install command; follow the installation instructions in the official README. After installation, the installer places nac-web in $HOME/.local/bin; if needed, add that directory to your PATH.
2. Choose model authentication — this must be completed before starting a session (you can also do it later in the dashboard). The officially recommended path is Arcee device-code login, which grants access to the Latest Open and Arcee Trinity models:

nac-web arcee-auth login

1. Launch the dashboard from your project directory. Success signal: the dashboard is accessible, authentication is complete, and you can select a model.

Later updates: nac ships with a native upgrade feature that installs the latest stable version:

nac-web upgrade

Alternative authentication methods (secondary options only): nac-web codex-auth login obtains OpenAI models via OAuth with a ChatGPT account; or export a catalog provider's API key (e.g., OPENAI_API_KEY). Note: logging in involves your account credentials; a model authentication method must be selected before starting a session. For your first attempt, launch only a single low-risk task with observable results, and prefer read-only actions. Work in an episode is performed by threads, and real tasks may include write operations — there should be a confirmation step before any significant side effects occur.

Use Cases

1. Delegate a long-running task and get a structured result back: in the local dashboard, start a session with an authenticated model (the Latest Open or Arcee Trinity models obtained via the recommended Arcee device-code login) and issue a task. The orchestrator breaks the work down and spawns threads; when a thread finishes, it returns an episode — a structured summary of what it accomplished. Expected outcome: you don't have to watch the whole process; the episode tells you what each stretch of work achieved.

2. Hook nac into your existing agent tool: the official portable onboarding skill and MCP integration let you paste the official onboarding snippet into your agent tool of choice, which installs nac, completes configuration, and connects MCP. The evidence does not provide step-by-step details for specific agent clients, so how your particular agent behaves must be judged by the actual integration results; what can be confirmed is this officially provided onboarding capability.

Limits and Boundaries

• Read/write nature: the orchestrator itself only plans and spawns threads — it does not execute commands or edit files. Actual execution happens inside threads, so real tasks may include write operations and command execution. For your first attempt, start with a task that has observable results, and there should be a confirmation step before any significant side effects occur.
• Authentication requirement: model authentication must be completed before starting a session (recommended: nac-web arcee-auth login device-code login; alternatively nac-web codex-auth login; or export a catalog provider API key such as OPENAI_API_KEY). The evidence does not state whether there is a free tier or any billing — unknown.
• Local components: the installer places nac-web in $HOME/.local/bin, and the dashboard runs locally, which counts as local state writes; the evidence provides no uninstall instructions.
• Explicit official restriction: the orchestrator cannot execute commands or edit files; it can only obtain structured results in episode form via threads.
• Key information not yet confirmed: which specific agent clients are supported, the cost and quotas of model calls, the environment isolation and permission boundaries of thread execution, and the exact install command and name of the portable onboarding skill — none of these are stated in the official evidence. Before taking on high-risk tasks, verify these points in the official documentation or repository; do not infer "no risk" from "no statement found."