agent-top

htop for local coding agents.

You have three Claude Code sessions, a Codex thread in VS Code, and a Gemini CLI you forgot about. Which one is burning tokens right now? Which one is waiting on you? Which MCP server is still alive after the agent that started it died? agent-top answers that in one terminal view, the way htop answers it for processes and btop answers it for the whole machine.

Recorded from a synthetic snapshot (docs/demo-snapshot.json, replayed with --replay) rather than a live machine, because a recording of real sessions would publish real project names, working directories and session ids. Regenerate with vhs docs/demo.tape.

What it shows

The detail pane shows the process tree (agent, subagent, mcp, shell, tool) and the token breakdown. Orphaned MCP processes, servers with no live agent above them, are listed in red.

That failure is not hypothetical. Codex has a run of reports about leaked MCP process trees, three of them still open:

Nothing about this is specific to Codex. Every harness that spawns helper processes has the same shape of bug available to it, which is why agent-top looks for the symptom rather than for one vendor's bug.

Supported harnesses

Install

brew install kannandreams/tap/agent-top

Every route ends at the same single binary — no Python, no Node, no daemon, nothing to configure:

--locked builds against the dependency versions the release was tested with; drop it if you would rather cargo picked newer ones. Building from source needs Rust 1.85 or newer (edition 2024).

To upgrade: brew upgrade agent-top, or re-run the cargo install command.

Usage

agent-top                    # interactive, refreshes every second
agent-top --once             # print the table once and exit
agent-top --json             # one snapshot as JSON, for scripts and bug reports
agent-top --interval-ms 500  # faster refresh
agent-top --stopped-window-min 120
agent-top --replay snap.json # render someone else's --json, keys and all
agent-top --prices           # the effective price table, and where each row came from

Homebrew installs shell completions for you. Otherwise, generate them with agent-top --completions zsh (or bash, fish, elvish, powershell) and source the output from wherever your shell keeps them.

--replay renders a saved snapshot in the full interactive UI without reading anything on the local machine, so a bug report can be inspected exactly as the reporter saw it.

Keys: j/k move, s cycle sort, r reverse, t toggle the detail pane, Tab switch that pane between the process tree and the tool trace, x hide stopped sessions, p pause, ? help, q quit.

Tool trace

Tab turns the detail pane into a waterfall of the selected agent's recent tool calls, on a shared time axis:

 tool trace   5 of 71 calls · window 1m00s
   in tools 58%  slowest Bash 20.0s  1 in flight  1 failed
 Bash             2.5s  ▉▉▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏
 Read             40ms  ▏▏▉▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏
 ↳Grep           12.0s  ▏▏▉▉▉▉▉▉▉▉▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏
 Edit            300ms! ▏▏▏▏▏▏▏▏▏▏▏▏▉▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏
 Bash            20.0s… ▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▉▉▉▉▉▉▉▉▉▉▉▉▸

Width is the call's share of the window; colour is how long it took, on a log scale from green under a second, through amber, to red approaching a minute. Those are two channels on purpose: at a typical zoom most calls are one cell wide, so width alone would say nothing about a 40 ms read next to a 30 s test run. ↳ and blue mark a subagent's call, … and amber a call still running, ! and red one the harness reported as failed.

in tools is the share of the window covered by at least one call (overlapping calls merged, not summed) — the rest is the model thinking, which is usually the answer to "why has this agent been busy for eight minutes".

No configuration and no telemetry opt-in: the spans are reconstructed from the transcript the harness already writes, by pairing each call with its result (Claude's tool_use / tool_result on tool_use_id, Codex's function_call / function_call_output on call_id) and reading the timestamps that bracket them. Only the call's name, id and timing are read, never its arguments or output. The spans are in --json as well, so they can be fed to a real tracing tool.

Prices

Prices are data, not code. The table shipped in the binary lives in crates/agent-top-core/prices.toml, and a file of your own is merged over it at startup:

# ~/.config/agent-top/prices.toml   (USD per million tokens)

[[model]]
prefix = "gpt-5-codex"
input = 1.25
output = 10.0
cache_read = 0.125

An entry whose prefix matches a built-in one replaces it, so a price that has gone stale can be corrected without waiting for a release. A new prefix is added, which is how the models this project does not ship prices for get costed at all. Cache writes default to Anthropic's multipliers of the input price (1.25x for the 5 minute TTL, 2x for the hour) and can be set explicitly with cache_write_5m and cache_write_1h.

The longest matching prefix wins, so claude-fable-5-1 beats claude-fable-5, and a date-suffixed id like claude-sonnet-4-6-20251114 resolves to its base model. agent-top --prices prints the effective table with the source of every row, which is the quickest way to find out why something is showing n/a. A price file that cannot be parsed is reported on stderr and ignored; the built-in prices still apply.

A model with no entry anywhere is never guessed at. Its tokens are counted and reported as unpriced, and any total containing them is shown as a floor.

Where the numbers come from

The whole point of this tool is that its numbers are right, so it is explicit about which ones are exact and which are inferred.

• Tokens are counted, never estimated. They come from the usage records the harness writes itself, deduplicated per API message so a response split across several transcript lines is counted once.
• Costs come from a table you can read and change. agent-top --prices shows it. A model with no price is reported as unpriced rather than guessed at, which is why a total containing one is shown as a floor (≥, +) instead of a number that looks more precise than it is.
• Attribution says how confident it is. Claude Code publishes a per-pid registry, so a session is matched to its process exactly. Codex has no equivalent, so the match is made on working directory and start time, and the detail pane labels that row a heuristic rather than presenting it as fact.
• Only metadata is read. Token counts, model ids, tool names, timestamps. Never a prompt, a tool input, or a tool result.
• Nothing is written, signalled, or sent anywhere. agent-top never kills or writes to an agent and makes no network calls. Killing an orphaned MCP server is your decision, with your own kill.

docs/architecture.md has the mechanism underneath: the process walk, the incremental transcript tail, and how a snapshot is assembled on each tick.

Roadmap

See docs/roadmap.md. Short version: exact Codex attribution, a logical subagent tree from transcripts, trace export to OTLP, user-supplied price tables, and a hook subcommand for harnesses that support it. docs/releasing.md is the release runbook.

Development

cargo test
cargo run -- --once
cargo clippy --all-targets

Two crates, split by dependency rather than by size: crates/agent-top-core is discovery, transcript parsing, pricing and the process model, with no terminal dependency, so all of it is testable without a TTY and it is exactly what --json prints; crates/agent-top is the ratatui front end and the CLI. Both are published, because a crate on crates.io cannot depend on an unpublished one — agent-top-core exists on the registry so that agent-top can. The internal engineering handbook (PRD, RFCs, ADRs, decisions) lives in the sibling agent-top-internal-docs repository.

License

MIT