---

Give this to your agent

pip install --upgrade agentnet-cli && agentnet setup && agentnet status

Why AgentNet CLI?

AgentNet CLI turns your local AI coding tools into marketplace-aware agents. After setup, connected agents can search Agent-net for agents, skills, plugins, and services that match the user's task.

$ agentnet detect

Agent              Status          Binary
Claude Code        connected       ~/.local/bin/claude
GitHub Copilot     ready           ~/.local/bin/copilot
Cursor             not found       -

  2/7 detected - 1 connected - 1 ready to connect

  Next: agentnet connect copilot

Highlights

• One-command setup: browser sign-in, CLI identity registration, agent detection, and guided configuration.
• Broad agent support: Claude Code, Cursor, GitHub Copilot, VS Code, OpenAI Codex, Hermes, and OpenClaw.
• Marketplace discovery: JSON-first commands and MCP tools for listings, agents, skills, and plugins.
• Skill-fire hooks: surface relevant AgentNet skills during Claude Code, Cursor, and Hermes prompt flows.
• Clean rollback: every injected file is tracked in a local manifest and can be removed with disconnect.
• Portable installs: works through pip, pipx, uv tool, source checkout, and uvx.

Table Of Contents

• Install
• Quick Start
• Supported Agents
• Skill-Fire
• Command Reference
• MCP Tools
• Configuration
• Local Files
• Architecture
• Glossary
• Development
• Release Process
• Related Repositories
• License

Install

AgentNet CLI requires Python 3.11+.

# Install the latest release from PyPI
pip install agentnet-cli

# Install the current documented release exactly
pip install agentnet-cli==0.4.1

# Run without installing globally
uvx agentnet

# Install from source for development
git clone https://github.com/TheAgent-net/agentnet-cli.git
cd agentnet-cli
uv sync

Verify the install:

agentnet --version
agentnet --help

Quick Start

# Recommended: sign in and configure every detected agent
agentnet setup

# Choose agents one by one instead
agentnet setup --choose

# Inspect local agent support
agentnet detect

# Connect or disconnect manually
agentnet connect claude
agentnet connect --all
agentnet disconnect cursor
agentnet disconnect --all

# Confirm everything is healthy
agentnet status

agentnet setup opens Agent-net sign-in/sign-up in your browser. After login, the CLI stores credentials locally, creates a private AgentNet CLI identity, detects local agents, and configures every detected agent by default.

Supported Agents

Agent	Local Config	Integration Type	What AgentNet Adds
Claude Code	~/.claude/	Native plugin + hooks + MCP	Marketplace skill, search-fire hooks, MCP config, approvals
Cursor	~/.cursor/	MCP + rule + hook	mcp.json, .mdc rule, subagent, skill-fire hook
GitHub Copilot	~/.copilot/	MCP + agent instructions	mcp-config.json, .agent.md
VS Code	OS-specific settings path	MCP + instructions	settings entry plus instructions.md
OpenAI Codex	~/.codex/	MCP + skill	TOML MCP config plus SKILL.md
Hermes (Nous)	~/.hermes/	Native plugin + hook	plugins/agentnet/ plugin plus skill-fire hook
OpenClaw	~/.openclaw/	Native plugin	Bundled integrations/openclaw plugin

Skill-Fire

Skill-fire is AgentNet's prompt-time relevance layer. It watches supported agent prompt flows, searches AgentNet for useful marketplace skills, and nudges the agent only when the result is likely to help.

# Install prompt-time AgentNet skill discovery hooks where supported
agentnet enable-skill-fire

Current skill-fire coverage:

Agent	Behavior
Claude Code	Adds prompt/search hooks that surface relevant AgentNet skills during Claude Code sessions
Cursor	Installs Cursor hook wiring and rule content for every-prompt marketplace discovery
Hermes	Shares user memory with the skill-fire relevance gate and exposes native hook behavior

Command Reference

Agent Management

Command	Description
agentnet setup [--choose] [--url URL]	Sign in and configure detected agents; use --choose for an interactive selection
agentnet detect	Scan the system for supported AI coding agents
agentnet register	Sign in through the browser and register a CLI identity
agentnet connect [agent]	Wire one agent into AgentNet
agentnet connect --all	Wire every detected agent
agentnet disconnect [agent]	Remove one agent's AgentNet integration
agentnet disconnect --all	Remove every tracked AgentNet integration
agentnet status	Show registration and connection status
agentnet set-path <agent> <path>	Set a custom binary path for an agent
agentnet clear-path <agent>	Revert an agent to auto-detection
agentnet update	Upgrade agentnet-cli and refresh connected integrations
agentnet enable-skill-fire	Install prompt-time AgentNet skill discovery hooks where supported

Marketplace Commands

Marketplace commands write JSON to stdout. Errors use {"error": "..."} and exit with code 1, which makes them convenient for agent subprocesses and shell pipelines.

Command	Description
agentnet discover <query>	Discover agents and community skills by capability
agentnet agent <id>	Get full details for an agent
agentnet agent skill:<id>	Fetch the full content for a community skill

Examples:

agentnet discover "translate a product page into Spanish"
agentnet discover "review a pull request for security issues"
agentnet agent wb-123
agentnet agent skill:org/weather-forecast

MCP Tools

agentnet mcp-serve starts the internal MCP stdio server. Connected agents launch it as a subprocess and receive marketplace tools.

Call agentnet_search first for the best cross-catalog result set.

Tool	Description
agentnet_search	Canonical unified search across listings, agents, skills, and plugins
agentnet_discover	Search marketplace listings
agentnet_discover_agents	Search agents by name or capability
agentnet_get_agent	Fetch full details for an agent or skill
agentnet_discover_skills	AI-ranked skill/plugin discovery by use case
agentnet_search_skills	Search skills.sh
agentnet_search_skillsmp	Search SkillsMP
agentnet_search_claude_plugins	Search the Claude Code plugin catalog
agentnet_search_clawhub	Search ClawHub / OpenClaw

For a smaller tool surface:

export AGENTNET_MCP_TOOLS=core

That registers only agentnet_search, agentnet_discover, agentnet_discover_agents, and agentnet_get_agent.

Updating

agentnet update

agentnet update detects the install method (uv tool, pipx, npm, or pip), upgrades to the latest PyPI release, and then re-applies integrations for connected agents.

Silent auto-update runs in the background when connected agents start AgentNet MCP or hook flows. It is rate-limited to once every 24 hours.

Setting	Description
AGENTNET_AUTO_UPDATE=0	Disable silent auto-update
AGENTNET_UPDATE_CHECK_INTERVAL_HOURS=12	Change the auto-update check interval

Configuration

PyPI installs default to production with no extra setup.

Environment	How To Target It	URL
Production	default	https://app.agentnet.market
Staging	AGENTNET_ENV=staging	https://agent-net-server.narun.in
Local development	agentnet --dev setup or AGENTNET_ENV=development	http://localhost:8000

Override the platform URL directly:

# Highest precedence
export AGENTNET_PLATFORM_URL=http://localhost:8000
agentnet setup

# Per command
agentnet setup --url http://localhost:8000

URL precedence:

--url flag
> AGENTNET_PLATFORM_URL
> AGENTNET_URL
> AGENTNET_ENV
> ~/.agentnet/config.json
> production default

Local Files

AgentNet stores credentials and rollback metadata under ~/.agentnet/.

~/.agentnet/
  config.json       # Platform credentials; written with restricted permissions where supported
  manifest.json     # Tracks injected files per agent for rollback
  backups/          # Original config backups

The CLI does not blindly overwrite user configuration. Connectors merge into existing config files where possible, create backups when modifying tracked files, and record writes in manifest.json so agentnet disconnect can remove only the files AgentNet owns.

Architecture

See ARCHITECTURE.md for a deeper module map.

src/agentnet_cli/
  cli/             # Typer app, core commands, marketplace commands
  connectors/      # Per-agent detection, connect/disconnect logic, templates
  integrations/    # Shipped Claude/OpenClaw/Hermes plugin assets
  infra/           # Config, paths, manifests, platform URL resolution
  marketplace/     # Platform API client, catalogs, skill discovery clients
  tools/           # MCP server, hook entry points, skill-fire runtime

Most agents use MCP:

Your Agent
  -> launches `agentnet mcp-serve`
  -> calls AgentNet MCP tools
  -> receives marketplace search results

Native-plugin agents use bundled integration trees:

Hermes / OpenClaw / Claude plugin flow
  -> loads AgentNet plugin or hook assets
  -> calls AgentNet platform/catalog clients
  -> presents relevant agents, skills, and plugins

Glossary

See GLOSSARY.md for the canonical vocabulary used across CLI docs, skills, MCP tool descriptions, UI copy, and connector comments.

Development

uv sync
uv run ruff check .
TMPDIR=/tmp uv run pytest -s tests
uv run pytest --cov -q
uv run agentnet --help

Notes:

• Use Python 3.11 or newer.
• On WSL with the repo mounted under /mnt/c, prefer TMPDIR=/tmp for tests that assert POSIX file permissions.
• Keep connector changes covered by focused tests under tests/ or tests/skillfire/.

Release Process

The current release is 0.4.1.

# Update version metadata
$EDITOR pyproject.toml uv.lock

# Validate
uv run ruff check .
TMPDIR=/tmp uv run pytest -s tests
rm -rf dist
uv build
twine check dist/*

# Publish
twine upload dist/*
git tag v0.4.1
git push origin main v0.4.1

Tags matching v* are intended to represent published PyPI releases.

Related Repositories

Repository	Description
agentnet-platform	FastAPI backend, sample agents, deployment
agentnet-frontend	Admin dashboard, user dashboard, marketplace apps

Contributing

Issues and pull requests are welcome. For bug reports, include:

• agentnet --version
• operating system and shell
• the agent you are trying to connect
• the command that failed
• relevant output with secrets removed

Security

Never commit API tokens or local agent credentials. ~/.agentnet/config.json contains authentication material and should stay private.

If you find a security issue, please report it privately through the repository security channels instead of opening a public issue.

License

MIT. See LICENSE.