agent-top is a terminal dashboard for monitoring locally running coding agents. It is described as "htop for coding agents," designed to solve the problem of multiple coding agent sessions being scattered across different windows, each incurring costs and leaking resources, with no unified view.
Purpose
Coding agents have evolved into long-running processes, and users often run multiple sessions at the same time, each with its own cost consumption and potential process leaks. No single tool shows all of these sessions at once. agent-top fills this gap, providing a unified monitoring view for coding agents much like htop shows processes and btop shows overall system status.
Key Capabilities
- Centralized view: Displays all Claude Code, Codex, Gemini CLI, and OpenCode sessions on the machine in one terminal interface, as well as any other coding tool sessions that can be detected.
- Status monitoring: The STATE panel shows which agents are working and which are waiting for user input.
- Cost tracking: The COST panel shows each session's cost so far, calculated at list price, with data sourced from each tool's own usage records.
- Token statistics: Displays the number of tokens consumed by each session.
- MCP leak detection: Red rows in the detail pane indicate MCP servers whose owning agent has disappeared, i.e. leaks; agent-top detects such cases on every tick.
- Process correlation: Displays helper processes left behind by each session.
- Cost reports: Can aggregate the total cost of all sessions by tool, model, project, or date.
- Snapshots and replay: The entire screen can be saved as a JSON file for bug reports and screenshots.
Usage
agent-top is written in Rust and released as a single static binary, supporting macOS and Linux. No runtime environment, no daemon, no configuration, and no features need to be enabled in the agents. It reads the transcript records already written by each tool and the operating system's existing process table. It never writes, sends signals, or terminates any agent, and data never leaves the local machine.
For installation, first run, and keyboard shortcut instructions, see the "Getting started" document. For each panel of the live view, the questions it answers, and data sources, see "The live view" document.
Use Cases
- Running multiple coding agent sessions at the same time and needing unified monitoring of status and cost.
- Needing to track token consumption and costs for each session in order to control budgets.
- Needing to detect leftover process issues such as MCP server leaks.
- Needing to generate cost reports and analyze spending by tool, model, project, or date.
- Needing to save a snapshot of the current monitoring state for bug reports or screenshots.
Limitations
- Currently only macOS and Linux versions are available; no information is provided for other platforms.
- The specific coding tools supported depend on actual detection capabilities; the official text lists Claude Code, Codex, Gemini CLI, and OpenCode, and states that other coding tools can be detected.
- Costs are calculated at list price and may not reflect actual discounts or subscription prices.
- It does not provide control functions for agents (such as termination or sending signals); it is for monitoring only.
- Other feature details, pricing model, prices, and similar information are not yet provided.
