arcee-ai/nac

arcee-ai/nac

Language: Rust

Stars: 13

Forks: 2

Open issues: 0

Created: 2026-03-28T04:02:01Z

Pushed: 2026-06-18T21:59:52Z

Default branch: main

Fork: no

Archived: no

README:

nac

Small coding agent. Heavily inspired by slate. Also takes inspiration from nanocode and pi.

Install the latest edge build:

curl -fsSL https://raw.githubusercontent.com/sapiosaturn/nac/main/scripts/install.sh | sh

Pinned version installs are not supported yet.

The installer places two binaries in $HOME/.local/bin by default:

• nac-web: the web dashboard for managing multiple sessions.
• nac: the terminal UI and utility commands such as codex-auth and upgrade.

Set OPENAI_API_KEY, then run nac-web -C /path/to/project and open the printed local URL.

To use ChatGPT Codex auth instead of an OpenAI API key, run nac codex-auth login and complete the device-code flow in a browser. In nac-web, choose chatgpt-codex-responses in the launch modal, or configure backend = "chatgpt-codex-responses" under [model]. For the TUI, launch with nac --backend chatgpt-codex-responses.

Optional:

• OPENAI_BASE_URL
• OPENAI_MODEL

Linux installs use the portable static build.

Upgrade to the latest edge build:

nac upgrade

nac upgrade reinstalls both nac and nac-web.

Run the web dashboard:

nac-web -C /path/to/project --bind 127.0.0.1:3210

Open http://127.0.0.1:3210/ for the dense session dashboard. nac-web exposes a central session manager for web clients. It resolves one server store at startup, then can create, resume, inspect, submit prompts to, and stream events from multiple sessions at once. Useful endpoints:

• GET /health
• GET /store
• GET /sessions
• POST /sessions
• GET /sessions/{session_id}
• POST /sessions/{session_id}/runs
• GET /sessions/{session_id}/events?after_sequence_id=0
• GET /sessions/{session_id}/events/stream?after_sequence_id=0
• POST /sessions/{session_id}/cancel-active-run

AGENTS.md is loaded hierarchically from the project and globally from NAC_HOME / ~/.config/nac. Skills are discovered from project and user skill directories; the orchestrator sees compact skill metadata and preloads selected skills for worker threads, while workers do not activate skills themselves. nac ignores disable-model-invocation; avoid interactive skills because nac is intended to run rather autonomously. Sessions are stored in the project store (.nac/store.db by default): use nac resume for the picker, nac resume --last for the newest session, or nac resume SESSION_ID for a specific session. Thread history does not auto-compact right now.

Uninstall:

curl -fsSL https://raw.githubusercontent.com/sapiosaturn/nac/main/scripts/uninstall.sh | sh

nac can run tools inside a Podman sandbox (requires Podman to be installed):

nac --sandbox

By default this mounts the current directory into the sandbox at /workspace.

For a custom setup:

• --no-mount-cwd disables the default current-directory mount
• --mount HOST:GUEST adds a read-write mount
• --mount-ro HOST:GUEST adds a read-only mount
• --sandbox-image IMAGE overrides the default image (python:3.13-bookworm)

On macOS, start Podman first:

podman machine init
podman machine start

Recommended config

Optional config lives at ~/.config/nac/config.toml, or at $NAC_HOME/config.toml when NAC_HOME is set. Explicit CLI args and environment variables override TOML defaults. Resumed sessions continue using the model and sandbox settings stored in their session snapshot.

The api_key_env setting names the environment variable to read when OPENAI_API_KEY is not set. Store paths remain relative to the launch working directory.

[agents_md]
fallback_filenames = []
max_bytes = 4194304

[storage]
store_path = ".nac/store.db"

[model]
backend = "openai-responses"
model = "gpt-5.5"
base_url = "https://api.openai.com/v1"
reasoning_effort = "xhigh"
api_key_env = "OPENAI_API_KEY"

[sandbox]
image = "python:3.13-bookworm"

[worker]
thread_timeout_secs = 3600

[mcp_servers.exa_web_search]
enabled = true
transport = "streamable_http"
url = "https://mcp.exa.ai/mcp"

[mcp_servers.context7]
enabled = true
transport = "streamable_http"
url = "https://mcp.context7.com/mcp"

[mcp_servers.grep_app]
enabled = true
transport = "streamable_http"
url = "https://mcp.grep.app"

Supported MCP transports right now are stdio and streamable_http. Stdio servers can provide command, args, and env; streamable HTTP servers provide url and optional headers. MCP string values support ${ENV_VAR} expansion.

For ChatGPT Codex auth, the default base URL is https://chatgpt.com/backend-api; NAC sends non-streaming Responses requests to /codex/responses. Use nac codex-auth status to inspect the saved account and nac codex-auth logout to remove local tokens.