nac CLI
An open agent harness from Arcee for long tasks: an orchestrator splits work into threads and collects their reports
- 294 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- Threads run commands and edit files on your machine
- ChatGPT sign-in tokens and provider API keys are stored locally and sent with requests to those services
Install
Manual install
curl -fsSL https://raw.githubusercontent.com/arcee-ai/nac/main/scripts/install.sh | shPlaces nac-web in ~/.local/bin. Add that directory to PATH if it's missing.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Threads run commands and edit files on your machine
- ChatGPT sign-in tokens and provider API keys are stored locally and sent with requests to those services
This is third-party code. Review the repository files before installing.
What it does
nac is an open agent harness from Arcee for long tasks that must stay true to the original intent: experiments, training runs, infrastructure work. A central orchestrator splits the work into parts and launches threads. It never runs commands or edits files itself. Each thread returns an episode, a structured summary of what was done. The install.sh script puts the nac-web command into ~/.local/bin, and `nac-web` starts a local dashboard with threads and episodes. Before the first run you pick a model provider: Arcee with a device code, your own ChatGPT account through OAuth, or an API key from one of the catalog providers. The install-skill.sh script adds the nac-onboarding skill to your agent. It walks you through installing nac, choosing a provider and connecting MCP servers.
Who it is for. For people who run long autonomous tasks: experiments, model training, infrastructure work.
Good fit when
- The task spans many steps and the plan must keep the original intent even while single threads run for hours
- The work splits into independent parts that are easy to run as parallel threads
- You already have a ChatGPT subscription and want to run OpenAI models by signing in to your account with no separate API key
- You want a local web dashboard to watch several parallel threads of one task
- You want to connect nac to MCP servers right from an agent you already run
Not a fit when
- You need a short one-off edit: the threads-and-episodes scheme is built for long multi-step tasks
- You have no access to any model provider: a session cannot start without a sign-in
- You want a plugin inside Claude Code or Cursor: nac runs as a separate harness with its own dashboard
- You need an API that never changes: the threads-and-episodes architecture is still evolving
Example request
Start a long nac task: prototype a training pipeline and keep the plan within the original descriptionLimitations
Before the first run you need a sign-in to at least one model provider: Arcee, ChatGPT through OAuth, or an API key for DeepSeek, Fireworks, Together, OpenAI or Anthropic. All work goes through threads, which adds a layer compared with a terminal agent. The install, skill and uninstall scripts download from raw.githubusercontent.com. Before production use, read the docs sections on model request safety and the sandbox. The dashboard listens only on the local address 127.0.0.1:3210.
How to disable. Run the removal script `curl -fsSL https://raw.githubusercontent.com/arcee-ai/nac/main/scripts/uninstall.sh | sh`. It removes nac-web and its files from ~/.local/bin.
FAQ
How is nac different from a plain terminal agent?
A plain agent plans and edits files itself. In nac only the orchestrator plans, while threads it launches run commands and make edits, and each returns an episode with the result.
Can I use my ChatGPT subscription with nac?
Yes, `nac-web codex-auth login` signs in to your ChatGPT account through OAuth and unlocks OpenAI models.
How do I connect nac to MCP servers from an agent I'm already running?
Install the onboarding skill with the install-skill.sh script and the --target agents flag, then load nac-onboarding in your agent. The skill walks you through installing nac, choosing a sign-in and connecting your MCP client to nac.
Where does nac show task progress?
nac-web starts a local web dashboard at 127.0.0.1:3210 with running threads and their episodes. No cloud account is needed.
How do I upgrade nac to the latest stable release?
`nac-web upgrade` installs the latest stable release over the current one, and the uninstall.sh script removes nac completely.
Discussion
No questions or reviews yet.
Related
CLI toolsTech Leads Club Agent Skills
Agent Skills
A vetted skill registry with a CLI that installs into twenty-plus agents and an MCP server for on-demand skill lookup
A Python terminal coding agent in the manner of Claude Code: reads the repository, edits files, runs commands, with memory, sub-agents and MCP
A configurator that reads your repo and sets up agents, skills, MCP servers and rules for its stack across several assistants
A plugin for Codex and Claude Code that runs a task through a chain of roles from planning to an independent audit
1C AI Assistant MCP proxy
1c-ai-mcp
An MCP server that forwards questions, code review and code explanation requests to the 1C AI Assistant cloud service
Rust coding agent for the terminal and CI built around DeepSeek models that reads code, edits files and runs commands
README in short
The README credits slate, nanocode and pi as sources of ideas. A quick start follows, then provider choice and links to docs on configuration, the HTTP API and safety.