CodexHud Flow

See what your Codex agents are doing while a task runs. CodexHud is a local VS Code extension that turns session logs into a visual flow of the main agent, subagents, tools, skills, and reasoning. It also shows token usage, estimated API-equivalent cost, context pressure, elapsed time, and completion alerts.

Get started

1. Install CodexHud Flow (PareschiApps.codex-hud) from the VS Code Marketplace.
2. Open a trusted local folder and start a Codex task.
3. Select CodexHud in the status bar or run CodexHud: Open HUD.

The extension requires VS Code 1.132.0 or later and supports local Windows, macOS, and Linux workspaces.

Explore a session

• Switch between Classic robots, an automatically arranged map, and Timeline, a step-by-step view of observed calls and parallel branches.
• Select a figure to inspect its recorded timing and available Input and Output. Calls and results remain attached to the agent that owns them.
• Follow the live task or inspect earlier sessions. Export the full timeline as a PNG.
• Open Project to review accumulated tokens and API-equivalent cost across associated chats and worktrees, with daily, model, and session breakdowns.

Cost figures are estimates based on captured model usage, not subscription charges. Partial or unpriced data is identified in the report.

Alerts and privacy

CodexHud signals confirmed completion through VS Code and local sound. Windows toast and Ntfy phone notifications are optional; configure Ntfy in CodexHud → Settings. Push messages do not include tool payloads, prompts, token counts, or costs.

CodexHud reads local Codex logs without editing them. Its HUD runs inside VS Code, with no web server. Captured project accounting is stored separately from the rebuildable flow history; deleting accounting requires a separate confirmation.

Remote SSH, browser, Dev Containers, and WSL workspaces are not currently supported. For help, use the Marketplace Q&A or the bundled support guide.

CodexHud is an independent community project and is not affiliated with OpenAI or Microsoft.

Carlos Pareschi