Run OpenAI Codex as an Agno agent with CodexAgent

Agno's CodexAgent, imported from agno.agents.codex, wraps the official openai-codex Python SDK, which bundles the Codex CLI and starts it locally as an app-server subprocess. Codex keeps running its own loop of shell commands, file edits, web search and MCP calls, and Agno adds sessions, streaming, the AgentOS API and the UI around it. It joins the Claude Agent SDK, LangGraph, DSPy and Antigravity adapters.

Install the SDK and authenticate once.

pip install openai-codex
codex login

Then use it like any Agno agent.

from agno.agents.codex import CodexAgent
from agno.db.sqlite import SqliteDb
 
agent = CodexAgent(
    name="Codex Coder",
    model="gpt-5.6-luna",
    sandbox="workspace-write",
    approval_mode="deny_all",
    reasoning_effort="high",
    cwd=".",
    db=SqliteDb(db_file="tmp/codex.db"),
)
 
agent.print_response("Summarize this repository", stream=True, session_id="repo-tour")

To put the same agent behind an API, hand it to AgentOS. It streams over SSE at /agents/codex-coder/runs and shows up in the AgentOS UI next to your native agents.

from agno.os import AgentOS
 
agent_os = AgentOS(agents=[agent])
app = agent_os.get_app()
 
if __name__ == "__main__":
    agent_os.serve(app="codex_app:app", reload=True)

Every Agno session_id gets one Codex thread. With a db, CodexAgent saves the thread id on the Agno session, so the second message in repo-tour resumes the same thread even after a restart. Without a db, the link lives in memory for as long as the agent object does. When a thread can't be resumed, for example with ephemeral=True or after Codex deleted its session files, CodexAgent starts a fresh thread and puts the stored chat history at the top of the prompt.

The constructor maps straight onto Codex settings. sandbox takes read-only, workspace-write or full-access, and approval_mode takes auto_review or deny_all. Use deny_all on a server, since nobody is there to answer a permission request. output_schema holds Codex's final answer to a JSON Schema and returns it as a JSON string, and config passes Codex config overrides to every thread, such as an mcp_servers table. CodexAgent checks these values when a run opens its thread, and a bad one fails that run.

# real output: CodexAgent(sandbox="yolo").run("hi")
status: RunStatus.error
content: Unknown sandbox 'yolo'. Expected one of: read-only, workspace-write, full-access

See the cookbook, and learn more about running Codex in the documentation.