AgentX is a terminal chat agent built on the OpenAI Responses API over WebSocket transport.
For most users, the simplest path is: install it globally, run agentx-setup once, then start agentx.
If you installed the package globally, use:
agentxIf you are working from the repository root, use:
node agentx.mjsagentx-setupUse that to save your OpenAI API key and runtime settings in ~/.agentx.
agentx --helporagentx -hshows quick helpagentx --versionoragentx -vprints the package versionagentx --debugprints raw websocket logs and suppresses live status lines- MCP calls and streamed arguments are shown in cyan when configured
Set your API key in the shell environment, or let agentx-setup write it into ~/.agentx for you:
export agentx_api_key="your-key-here"
# or: export AGENTX_API_KEY="your-key-here"AgentX prefers agentx_api_key and falls back to AGENTX_API_KEY.
The launchers load ~/.agentx when present. Startup also displays the active model and runtime settings.
To enable MCP tools, place an .agentx.mcp.json file in your home directory. Copy .agentx.mcp.json.example as a starting point and update its server URL and authorization. AgentX loads ~/.agentx.mcp.json automatically; a missing file is allowed, while invalid JSON prevents the prompt template from loading.
- Start AgentX.
- Type a normal message and press Enter.
- AgentX sends that message to OpenAI.
- The response streams into the terminal as it is generated.
- The prompt shows your user, short hostname, current working directory, and a
#marker. - AgentX waits for your first message before contacting OpenAI.
- Tool calls may stream command arguments and shell summaries live.
- If
.agentx_responseidexists, the session resumes automatically. If it contains pending tool calls, AgentX asks how you want to continue.
Install or update the latest release at any time with:
npm -g install @eliware/agentx-cli@latestSee AGENTS.md behavior for how project instructions are discovered and loaded.
npm -g uninstall @eliware/agentx-cli
rm -f $HOME/.agentx*