Multi-Agent Observability Solution: Unified Monitoring Platform for Claude Code and Codex

The multi-agents-claude-codex project provides a locally deployed real-time observability dashboard that unifies tracking of agent events from Claude Code and OpenAI Codex, helping developers monitor the full lifecycle of multi-agent workflows.

When AI agent work expands from single sessions to multi-agent collaboration, observability becomes a key challenge. Developers face complex questions:

• Which agent is performing what task?
• What are the input and output of tool calls?
• How are calls related between sessions?
• What are the model choices and performance?
• At which stage did a failure occur?

Traditional logging methods struggle with this complexity. Agent events are scattered across outputs from different processes, lacking a unified timeline and correlation analysis. When issues arise, developers often have to switch between multiple terminal windows and log files, leading to low efficiency and easy omission of key information.

The multi-agents-claude-codex project builds a locally running observability platform that captures runtime events from Claude Code and Codex via Hook mechanisms, providing unified monitoring, analysis, and auditing capabilities.

The core innovation of this project lies in supporting two mainstream AI programming agents—Anthropic's Claude Code and OpenAI's Codex simultaneously, and enabling cross-platform event correlation analysis through a standardized data model.

The project captures agent events through two mechanisms:

Claude Code Hooks: Using Claude Code's settings.json configuration, Python scripts are triggered at key lifecycle points (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, etc.) to send event data to a local server.

Codex Hooks: Through the .codex/hooks.json configuration, events from Codex sessions are similarly captured. Notably, the project is designed so that Codex retains ownership of OAuth credentials; Hook scripts only receive event JSON and do not access sensitive authentication information.

Event formats vary across different agent platforms. The project uses the normalize.ts module to convert raw payloads from Claude Code and Codex into a unified standardized event model, while retaining raw data for in-depth analysis.

Standardized events include the following key fields:

• Runtime identifier (claude-code / codex)
• Event category (session, turn, tool, agent, etc.)
• Status information (started, completed, failed, blocked)
• Model name and tool information
• Duration, working directory, session correlation

Event data is stored in a local SQLite database and simultaneously pushed to the front-end dashboard in real time via WebSocket. This design ensures both data persistence and support for real-time monitoring scenarios.

The Vue 3-based web interface provides:

• Timeline view: Displays all agent events in chronological order
• Filters: Filter by source application, session, event type, model, tool
• Chat transcript: View complete conversation history
• Chart analysis: Statistical views showing agent activity patterns

This is the project's most prominent feature. Developers can observe activities from both Claude Code and Codex in the same dashboard, compare their behavior patterns, and even analyze cross-platform collaboration scenarios.

The standardized event model makes this unification possible. Regardless of the underlying agent differences, upper-layer analysis tools use a consistent interface.