Install

Install the local application, connect sources and configure MCP.

#Prerequisites

AgentGraph expects Python 3.12 or later and uv.

curl -LsSf https://astral.sh/uv/install.sh | sh

#Install AgentGraph

Install AgentGraph with every first-party connector:

uv tool install 'agentgraph-server[all]'

This makes the agentgraph command available in your shell.

Or install only the connector support you need:

uv tool install 'agentgraph-server'
uv tool install 'agentgraph-server[google]'
uv tool install 'agentgraph-server[slack]'
uv tool install 'agentgraph-server[discord]'
uv tool install 'agentgraph-server[rss]'
uv tool install 'agentgraph-server[web]'

#Start the server

Start the server in a terminal session. It keeps that terminal occupied; use another terminal for onboarding and other commands. See below to run it in the background.

agentgraph serve

#Connect sources

Run guided onboarding to set up each installed connector that provides an interactive flow:

agentgraph onboard

#Install the browser extension

Install the AgentGraph Chrome Extension from the Chrome Web Store.

After the extension is installed, start agentgraph serve, then open a supported resource and keep it focused past the default three-second observation threshold.

#Install the agentgraph skill

The following command installs the agentgraph skill in ~/.agents/skills and ~/.claude/skills:

agentgraph install-skill

#Connect an MCP client

MCP is the recommended connection for ChatGPT Desktop Work Mode, Codex, Claude Desktop, and Claude Code.

agentgraph mcp-config

Codex and Claude Code register the server from your terminal:

codex mcp add agentgraph -- "$(which agentgraph)" mcp-serve
claude mcp add --transport stdio --scope user agentgraph -- "$(which agentgraph)" mcp-serve

ChatGPT Desktop Work Mode takes the same command through its MCP configuration screen: enter the printed executable under Command to launch and mcp-serve under Arguments.

Claude Desktop reads a config file — add the printed JSON to ~/Library/Application Support/Claude/claude_desktop_config.json.

How those tools reach the graph follows AGENTGRAPH_QUERY_TRANSPORT: by default the local server when one is reachable, otherwise the database directly.

#Run in the background

agentgraph serve should be a long-running process to host the viewer and provide continuous polling of the connected sources.

#macOS launchd

cat > ~/Library/LaunchAgents/com.agentgraph.serve.plist <<EOF
<?xml version="1.0" encoding="UTF-8"?>
<plist version="1.0">
<dict>
  <key>Label</key>
  <string>com.agentgraph.serve</string>
  <key>ProgramArguments</key>
  <array>
    <string>$(which agentgraph)</string>
    <string>serve</string>
  </array>
  <key>RunAtLoad</key>
  <true/>
  <key>KeepAlive</key>
  <true/>
</dict>
</plist>
EOF

launchctl load ~/Library/LaunchAgents/com.agentgraph.serve.plist

#Linux systemd

mkdir -p ~/.config/systemd/user
cat > ~/.config/systemd/user/agentgraph.service <<EOF
[Unit]
Description=AgentGraph local knowledge graph server
After=network.target

[Service]
ExecStart=$(which agentgraph) serve
Restart=on-failure
RestartSec=5
StandardOutput=journal
StandardError=journal

[Install]
WantedBy=default.target
EOF

systemctl --user daemon-reload
systemctl --user enable --now agentgraph