UseAgents CLI

Package-manager-style CLI for installing, managing, and running local AI agents and agentic tools.

UseAgents (agent) manages two kinds of registry entries:

• Direct agents install runnable agent code into ~/.useagents/runtimes/ and can be executed with agent run. Example: hello-world.
• Managed integrations run registry-provided wrapper logic that orchestrates a third-party installer, updater, uninstaller, or status check. Example: claude-code or openclaw.

Managed integrations are not treated as packaged agent tarballs. They are tracked in UseAgents state, but the upstream tool may live in locations owned by the third-party installer, such as a native binary directory or global package manager.

Installation

npm install -g @thesethrose/useagents

Requirements:

• Node.js >=20
• git for git-based installs
• tar for registry direct-agent installs
• Docker only when using Docker-backed sandbox execution

Common Workflows

Search the registry

agent search claude-code
agent search hello
agent search --type agent
agent search --type integration --page 2 --limit 5
agent get hello-world
agent get claude-code

agent search lists registry matches with their type, description, and install command. Direct agents show their latest agent version; managed integrations intentionally do not show wrapper versions.

Install

# Managed integrations orchestrate third-party setup flows
agent install claude-code
agent install openclaw

# Direct agents install runnable code from registry tarballs
agent install hello-world
agent install @your-scope/hello-world

# Direct agents can also be installed from local folders or git
agent install ./path/to/agent
agent install github:your-org/your-agent
agent install https://github.com/your-org/your-agent

Use --force to overwrite an existing direct-agent version or force a managed integration reinstall when the wrapper supports it.

List and inspect

agent list
agent list hello-world
agent info hello-world
agent info claude-code
agent config
agent doctor

agent list separates direct agents from managed integrations. Hidden filesystem entries under ~/.useagents/active/ are ignored.

Run direct agents

agent run hello-world --input '{"name":"Developer"}'
agent run hello-world --sandbox --input '{"name":"Developer"}'

Only direct agents are runnable through agent run. Managed integrations expose their own upstream commands after installation.

Sandbox mode is enabled with --sandbox or by the manifest sandbox policy. Docker-backed sandbox execution is used only when Docker is available and the agent does not require tools, secrets, or a configured model; otherwise the CLI uses local permission enforcement.

Update, upgrade, remove, and uninstall

agent update                 # Check registry connectivity
agent update hello-world     # Reinstall a specific direct agent from its recorded source
agent update claude-code     # Run the managed integration update flow

agent upgrade                # Upgrade recorded direct-agent installs
agent upgrade hello-world
agent upgrade claude-code

agent remove claude-code     # Remove UseAgents tracking, leave upstream tool intact
agent uninstall claude-code  # Run managed uninstall, then remove UseAgents tracking

agent remove hello-world     # Remove a direct agent from UseAgents
agent uninstall hello-world  # Same direct-agent cleanup path

For direct agents, remove and uninstall remove the active symlink, versioned runtime directory, and install record. For managed integrations, remove only removes UseAgents tracking; uninstall asks the wrapper to remove upstream install artifacts such as binaries while preserving credentials or application state unless the wrapper explicitly documents otherwise.

Logs and secrets

agent logs hello-world
agent secret set OPENROUTER_API_KEY
agent secret list

Secrets are stored locally in ~/.useagents/secrets/secrets.json with 0600 permissions when written by the CLI.

Registry auth

agent login
agent logout

agent login prompts for a registry session token from https://useagents.io/settings, validates it against the registry auth endpoint, and stores it in ~/.useagents/state/auth.json with user-only file permissions. Auth is used for registry publish/update APIs that require a logged-in session and verified email.

Registry installs use canonical registry artifact routes rather than publisher-supplied artifact URLs. If registry metadata includes an artifact SHA-256 checksum, the CLI verifies the downloaded tarball or managed-integration wrapper before installing. Deprecated packages install with a warning; yanked, quarantined, archived, or deleted packages do not install.

Command Reference

Direct Agent Manifests

Direct agents are folders with an agent.yaml manifest and a JavaScript entrypoint.

name: hello-world
version: 1.0.0
description: A simple hello world agent
runtime:
  type: javascript
  entrypoint: ./dist/index.js
permissions:
  network: false
  filesystem:
    read: []
    write: []
  secrets: []
tools:
  - echo.text

export async function run(input, ctx) {
  const name = input?.name || "World";
  const echo = await ctx.tools["echo.text"]({ text: `Hello, ${name}!` });
  return { message: echo.text };
}

Manifest Reference

name: hello-world                 # kebab-case identifier, or @scope/name for registry packages
version: 1.0.0                    # x.y.z semver
description: "A simple hello world agent"

runtime:
  type: javascript                # currently the only supported runtime
  entrypoint: ./dist/index.js

model:
  provider: openrouter
  model: anthropic/claude-3.7-sonnet

inputs: {}
outputs: {}

permissions:
  network: false
  # or:
  # network:
  #   enabled: true
  #   domains: ["api.example.com", "*.openai.com"]
  filesystem:
    read: []
    write: []
  secrets:
    - OPENROUTER_API_KEY
  sandbox:
    enabled: false
    tools:
      - echo.text

tools:
  - echo.text
  - http.fetch
  - fs.readText
  - fs.writeText

The manifest schema is defined in src/types.ts and enforced by agent validate, agent install, and agent run.

Runtime Contract

An agent entrypoint must export:

export async function run(input, ctx) {
  return {};
}

The context object includes:

Storage Layout

~/.useagents/
|-- runtimes/             # Versioned direct-agent installs
|   `-- hello-world/
|       `-- 1.0.0/
|-- active/               # Symlinks to active direct-agent versions
|   `-- hello-world -> ../runtimes/hello-world/1.0.0
|-- state/
|   |-- installs.json     # Direct-agent install records
|   |-- integrations.json # Managed integration records
|   |-- logs.jsonl        # Direct-agent execution logs
|   |-- audit.jsonl       # File/network/tool audit log
|   |-- permissions.json  # Remembered permission grants
|   `-- auth.json         # Registry auth token, when logged in
|-- secrets/
|   `-- secrets.json      # Local secret values
`-- cache/
    |-- registry/         # Downloaded direct-agent artifacts
    `-- integrations/     # Downloaded managed integration wrappers

Environment

Development

cd cli
npm install
npm run typecheck
npm run lint
npm run test
npm run check
npm run build
node dist/index.js --help

The source entrypoint guard rejects tsx src/index.ts because index.ts is not a declared CLI binary. Use the built CLI (node dist/index.js) or the installed agent binary for manual testing.

Security Model

1. Direct agents must declare all permissions in agent.yaml.
2. First run requires explicit permission approval, persisted in ~/.useagents/state/permissions.json.
3. Secrets are stored separately and only exposed when declared by the manifest.
4. Filesystem and network access go through permission-aware tools.
5. File, network, and tool operations are written to ~/.useagents/state/audit.jsonl.
6. Optional sandbox mode can use Docker for simple JavaScript agents when the agent has no tools, secrets, or configured model.

Current Limitations

• Direct-agent runtime support is JavaScript only.
• agent upgrade without arguments upgrades recorded direct-agent installs only; named managed integrations can be updated with agent update <name> or agent upgrade <name>.
• Registry cache TTL and offline mode are exposed in configuration output but are not implemented yet.