agent-top-core

The engine behind agent-top, htop for local coding agents. If you want the tool, install agent-top instead — this crate is its library half, published because a crate on crates.io cannot depend on an unpublished one.

It answers one question, with no terminal dependency at all: which coding agents are on this machine right now, what are they doing, and what have they spent?

use agent_top_core::{Collector, CollectorOptions};

let mut collector = Collector::new(CollectorOptions::default());
let snapshot = collector.collect();

for agent in &snapshot.agents {
    println!("{} {} {} tokens ${:.2}", agent.name, agent.state.label(), agent.usage.total(), agent.cost_usd);
}

A Snapshot is exactly what agent-top --json prints, and it serialises and deserialises, so it can be stored, shipped in a bug report and replayed.

What it does

• Discovery. Walks the process table with sysinfo, marks Claude Code, Codex, Gemini CLI, OpenCode, Aider, Copilot CLI and cursor-agent roots, folds their children into a tree labelled subagent / mcp / shell / tool, and flags MCP servers whose agent has died as orphans.
• Attribution. Joins each process to its transcript — exactly where the harness publishes a registry, by heuristic otherwise, and the result says which it was so a caller never has to guess how much to trust a row.
• Accounting. Tails transcripts incrementally, counting tokens, cost, turns and tool calls from the harness's own usage records. Tokens are counted, never estimated; a model with no known price is reported as unpriced rather than guessed at.
• Tracing. Pairs each tool call with its result to produce ToolSpans with real durations, from harnesses that log no telemetry of their own.

Everything is read-only and local: no network calls, no signalling of agents, no reading of prompt or tool content — metadata fields only.

Stability

Pre-1.0, and the version moves in lockstep with agent-top. The harness module in particular will change as the adapter trait lands. Pin an exact version if you build on it.

MIT. Source, roadmap and the tool itself: https://github.com/kannandreams/agent-top