ADF CLI
An npm package that runs ADF agents from a terminal, with no desktop app. It installs one command, adf: a terminal app, one-shot commands, and the ADF daemon that hosts the agents.
On this page
Install
Published npm release
npm install -g @agentdocumentformat/cli@latestAs checked on 12 November 2026, the latest published release is v0.8.9. Its package metadata declares engines.node: "^22.23.4 || ^24.22.1 || ^26.12.1". Use a stable Node.js release: 22.23.4 or newer within 22.x, 24.22.1 or newer within 24.x, or 26.12.1 or newer within 26.x.
v0.8.9 enforces this range at runtime and checks for the fixed Brotli decoder. Node.js 27 and later are intentionally refused until tested; older unsupported releases and prereleases are also refused. For a different package version, check its declared requirements with npm view @agentdocumentformat/cli@VERSION engines.node. Requirements for running or building main are listed separately below.
Native modules (SQLite, the OS keychain) install from prebuilt binaries for Windows, macOS (arm64, x64), and Linux (x64, arm64, glibc), so no compiler is needed.
If npm reports EEXIST for bin/adf, an older package named agent-document-format is installed. Remove it with npm uninstall -g agent-document-format and install again.
Contents
| Part | Command | Description |
|---|---|---|
| Terminal app | adf | The whole fleet in one screen. Chat with any agent or loop, approve tool calls, edit files, watch live events. |
| CLI | adf <command> | One-shot commands for scripts and quick checks. adf help lists them. |
| ADF daemon | adf daemon | The background process that runs the agents and serves the Daemon API. adf starts it for you. |
The terminal app and the one-shot commands are clients of the daemon. Quitting the app, or closing the terminal, leaves every agent running.
Documentation
- Getting started
- From nothing to an agent you can chat with.
- Terminal app
- Screens, keys, and slash commands.
- CLI reference
- Every command, flag, environment variable, and key.
- ADF daemon
- Run, configure, and secure the daemon.
- Daemon operations
- Data directory, ports, token, logs, and shutdown.
- Remote daemon
- Pair with another computer, or reach its daemon through SSH, a TLS proxy, or Tailscale.
- Daemon API
- The HTTP API the daemon serves. OpenAPI 3.
Upgrade
npm install -g @agentdocumentformat/cli@latest
adf daemon restartA running daemon keeps the old code until it restarts. adf --version prints the installed version,adf daemon status the one the running daemon reports.
Quick start
adf # terminal app, starts the daemon
adf identity new # or: adf identity restore
adf auth login grok # or add an API key in the app
adf new my-agent --start
adf agents
adf chat my-agent "hello"
adf daemon status # also stop, restart, logs -fThe first launch shows a checklist: identity, model, first agent, channel. Step by step: Getting started.
Local models
Use Ollama, LM Studio, or another local server with an OpenAI-compatible chat completions API. Set its base URL and the exact model ID in a provider, then select that provider and model for the agent. An API key is needed only if the server requires authentication. See Models and providers.
Both the server and the selected model must support native tool calling (function calling) for the agent to use tools. A server that only returns chat text is insufficient. Vision is optional: choose a server and model that support image input and enable images for the agent when it needs to understand pictures or screenshots.
Typical base URLs are http://localhost:11434/v1 for Ollama and http://localhost:1234/v1 for LM Studio. Start the model server and make a suitable model available first. The ADF daemon connects to this URL, so localhost means the computer running the daemon. For a daemon on another computer, use a model-server address that computer can reach.
Commands
Every command, flag, environment variable, slash command, and key is in the reference. An<agent> argument can be an agent ID, handle, or name. --json prints raw JSON.
| Group | Commands |
|---|---|
| Agents | agents, status, start, stop, loops, chat, interrupt, abort, events |
| Agent resources | config, files, file, inbox, outbox, timers, identities |
| Approvals and questions | tasks, task, approve, deny, asks, answer |
| Owner identity and new agents | identity, identity new | restore | unlock | lock, templates, new |
| Diagnostics | runtime, providers, settings, usage, mcp, adapters |
| Network | network, network mesh | server | tools | lan | runtimes |
| Sign-in | auth, auth login | logout |
| Other computers | pair, remotes, remotes use | forget |
| Daemon | daemon, daemon start | status | stop | restart | logs | token | remote | pair | devices |
Studio and the CLI
ADF Studio and the CLI share the settings file and owner identity on one machine, so providers, tracked folders, and agents set up in one appear in the other. Studio runs its agents in the same daemon, so Studio, the terminal app, and adf commands are clients of one process and show the same agents at the same time. A chat, an approval, or a config change in one appears in the others.
Studio, adf, and the terminal app also keep one list of saved connections to other computers. A computer paired with adf pair appears in Studio, and one added in Studio appears in adf remotes. Details:Connections and runtimes, Remote daemon.
Run from source
Main source: running or building ADF main requires Node.js ^22.23.4 || ^24.22.1 || ^26.12.1, as checked on 12 November 2026 at commit edc712c. These requirements currently match npm v0.8.9, but main and published releases can change independently. Main's runtime guard intentionally refuses Node.js 27 and later until tested, as well as older unsupported releases, prereleases, and builds without the fixed Brotli decoder.
From a clone of the ADF repository, the npm scripts run the TypeScript source directly:
git clone https://github.com/christianbalevski/adf.git
cd adf
npm install
npm run adf # terminal app
npm run adf -- agents # any command
npm run tui:mock # terminal app against a mock daemon
npm run daemon # the daemon in the foreground