TokenMon

A fast, zero-dependency command-line monitor for local AI coding agents.

It measures how fast models actually generate tokens (Tokens Per Second, TPS) by tracking pure generation time—separating thinking and output from tool runs, file edits, and idle waiting.

Screenshots

Expand a screenshot below. Click the image to view it at full size.

Features

• Generation Speed: Measures TPS from recorded output boundaries and labels timing fallbacks, including Claude prompt-to-response estimates.
• Zero Extra Dependencies: Runs on standard Python 3.10+ without installing third-party packages.
• Strictly Read-Only: Safely opens local files and SQLite databases in read-only mode (?mode=ro). Never locks or changes your logs.
• Meaningful Averages: Calculates true weighted speed ($\frac{\text{total tokens}}{\text{total time}}$), median speed, and min/max ranges over rolling time windows (30m, 1d, 7d, 30d, all).
• Supports Popular Agents: Auto-detects Codex (~/.codex), Claude Code (~/.claude/projects/), Antigravity (~/.gemini/antigravity-cli), Oh My Pi / OMP (~/.omp/agent/sessions/), Pi (~/.pi/agent/sessions/), and OpenCode (~/.local/share/opencode; v1 JSON/SQLite and v2 SQLite).
• Activity Timelines: Compare prompts, reasoning, model output, tool execution, and waiting across sessions on one time axis.
• Event Logs: Inspect the step-by-step history of user prompts, thinking, assistant responses, and tool calls.
• JSON Ready: Add --json to pipe clean data into jq or external dashboards.

Installation

Requires Python 3.10+.

Try it immediately with uv:

uvx tokenmon
uvx tokenmon ps
uvx tokenmon logs

For regular use, install it as an isolated tool:

uv tool install tokenmon
tokenmon

If tokenmon is not on your PATH, run uv tool update-shell and restart your shell. Upgrade with uv tool upgrade tokenmon.

Or install with pip in your Python environment:

pip install tokenmon
tokenmon

Usage

TokenMon uses simple Docker-style subcommands: stats (default), ps (sessions), timeline (visual activity), logs (events), and interactive (shell).

Running tokenmon by itself defaults directly to stats. The examples below use the installed command; with uvx, use uvx tokenmon in its place.

Output Formats

Stats, sessions, visual timelines, and historical logs support terminal, JSON, and standalone HTML output. --json remains a shortcut for --output json:

tokenmon stats --json | jq .
tokenmon ps --json | jq .
tokenmon logs 01a10275 --json | jq .
tokenmon timeline --window 1d --json | jq '.sessions'

tokenmon stats --output html > stats.html
tokenmon ps --output html > sessions.html
tokenmon timeline --output html > timeline.html
tokenmon logs 01a10275 --output html > logs.html

HTML reports contain their CSS, SVG charts, and small table-sorting script inline; they load without network access or extra dependencies. Timeline HTML uses separate model, reasoning, tool, user, and turn tracks so simultaneous events remain visible. Empty timeline tracks are hidden. Reports open on the latest activity cluster. Drag a range in the full-history overview, click to move the detail view, or use the time zoom and earlier/later controls. Labels remain fixed and time ticks adapt to the visible range; hover for exact timestamps and durations. Blank time gaps mean no recorded activity and may include waiting or unobserved work. Live --watch and --follow modes require terminal output and reject JSON or HTML.

Generation Speed & Metrics (stats, top, default)

tokenmon                                # auto-detect agents
tokenmon stats codex --window 1d         # select agent and window
tokenmon stats -w --interval 5           # live dashboard; Ctrl+C stops
tokenmon omp,pi -w                      # watch selected agents together
tokenmon omp,pi,codex -w                # combine three agents
tokenmon stats --all --wide              # full history and detailed layout

Windows: 30m, 1d, 7d, 30d, all; the default history cutoff is 30 days. -w means watch, not a window. Use --compact for narrow panes. Watch refreshes metrics, recent streams, and session cards; redirected output appends snapshots. --tasks limits inspected sessions, including with --all.

• Weighted TPS = valid tokens ÷ valid generation seconds. Median and range use individual valid streams.
• Human tables separate each source agent and model pair; JSON keeps its stable model-keyed aggregation.
• Total generation time sums overlapping streams separately; human output uses hours/minutes/seconds, while JSON keeps numeric seconds.
• Streams under one second, above 400 TPS, or with unconfirmed timing are excluded. Claude's single-record turn-span includes latency; OMP's omp-ttft measures first-output-item to completion, not guaranteed first-text-token decoding.
• Pi and OpenCode show sessions and token totals, but no confirmed TPS. OpenCode v2's time.streamed is a stream end, not TTFT.
• Effort and speed come only from recorded metadata, never model names or TPS. Missing JSON fields are null; session settings are current and event settings historical. Stats configurations describe valid streams; singular fields are null when configurations disagree.

Agent Data

--home selects the data root and overrides native environment settings. OMP/Pi roots contain sessions, including nested persisted workers. OMP also honors PI_CONFIG_DIR, OMP_PROFILE/PI_PROFILE, and existing migrated XDG directories. OpenCode's database override may be absolute or data-relative; in-memory stores cannot be monitored externally. Existing/copied logs suffice without running an agent.

Select multiple agents with a comma-separated list wherever an agent selector is accepted: tokenmon omp,pi -w, tokenmon ps omp,pi, or tokenmon logs latest --agent omp,pi,codex. tokenmon logs omp,pi selects the latest session across those agents; follow still pins just one session. Names are case-insensitive, surrounding spaces are ignored, and repeated names/aliases count only once. Unknown names and empty list entries are errors. Use all alone to select detected agents, or omit selection for auto-detection. Each selected adapter uses its own native root unless --home is supplied; that override applies the same root to every selected adapter.

OpenCode formats were checked at 1.0.0/1.1.65 (split JSON), 1.2.0 (SQLite), and 2.0.22 (SQLite projections). Detection is schema-based; SQLite takes precedence over leftover migrated JSON. Assistant/step usage, inherited fork history, and utility aggregates are not counted twice. Variant names are not inferred effort; recorded output-token semantics may differ by provider/version.

All adapters include separately persisted worker sessions within their data roots. Claude worker IDs are agent-<id>@<project>/<parent>; Codex uses native thread IDs and recorded ownership ordinals to exclude copied worker history. Legacy Codex worker forks without that boundary are excluded rather than double-counted. Antigravity includes unindexed worker databases; opaque embedded workers and copied-fork ownership cannot be decoded from the available metadata.

OMP uses recorded request starts to exclude inherited assistant output, recovering the start from completion and duration when necessary. Parented output without request-boundary evidence is excluded. Native OMP session_init workers remain distinct from Pi sessions even when both share a data root.

Active & Recent Sessions (sessions, ps, ls)

List sessions from the last 7 days by default, with message counts, tool runs, and idle status. Override with --window, or use --all for full history:

# List recent sessions
tokenmon ps

# Filter sessions within a time window
tokenmon ps --window 1d

# Filter to a specific agent
tokenmon ps codex

The Status column shows time since the last event using hours, minutes, and seconds, for example Idle (2m 05s) or Inactive (idle 3h 04m 05s). Durations of 24 hours or more include days, for example Inactive (idle 7d 00h 14m 19s). Stats total time, session durations, and recent session cards use the same duration format.

Visual Activity Timeline (timeline)

Compare recent sessions on a shared time axis. Colors identify recorded user, reasoning, assistant/model-output, tool, and turn-boundary events. Events with recorded durations become blocks; instantaneous events remain visible as markers. Blank regions are waiting or unavailable boundaries, never invented activity.

tokenmon timeline                         # compare the 10 most recent sessions
tokenmon timeline --window 1d --tasks 20 # choose history and lane count
tokenmon timeline --agent codex          # select one or more source agents
tokenmon timeline 01a10275                # visualize one session
tokenmon timeline --json | jq '.sessions'

Sessions are displayed longest first. --compact and --wide control label density; redirected output uses distinct block characters without ANSI color. The timeline JSON contains session lanes and ordered recorded-activity intervals.

Chronological Event Logs (logs, log)

See the chronological step-by-step history of prompts, model thoughts, responses, and tool calls:

# View timeline of the latest session
tokenmon logs

# View timeline of a specific session ID or prefix
tokenmon logs 01a10275

# Follow only new events from the session that is latest at startup
tokenmon logs -f

# Follow a specific session ID/prefix, or an agent's latest session
tokenmon logs 01a10275 --follow
tokenmon logs codex -f

# Customize the polling interval (default: 2 seconds)
tokenmon logs -f --interval 1

# Export all session event logs from today in JSON format
tokenmon logs --window 1d --json

Follow mode starts at the current end: it prints a short session header, then only newly observed timeline events. It keeps the initially selected session even when another session becomes newer, and appends output without clearing the screen or replaying history. Updates to an existing event's usage or timestamp do not print that event again. Press Ctrl+C to stop.

-f and --follow work with logs and log. Follow selects one session, so it cannot be combined with the batch export option --window; use --all to select a session older than the default 30-day cutoff. The interval must be a positive, finite number. Polling reads the selected session again rather than following raw file bytes, which also supports SQLite-backed Antigravity sessions.

Interactive Shell (interactive, repl, shell, -i)

Open an interactive terminal shell with live auto-refresh and tab completion:

tokenmon interactive
# or
tokenmon -i

Available commands inside the shell:

(tokenmon) summary 30m        # view 30m throughput table
(tokenmon) sessions           # list active & inactive sessions
(tokenmon) timeline           # compare session activity lanes
(tokenmon) logs latest        # view step-by-step session events
(tokenmon) recent 15          # view latest generation speeds
(tokenmon) watch 2.0 1d       # metrics, recent streams, and session cards (Ctrl+C to stop)
(tokenmon) help               # list all commands

Shell timelines and logs prefer exact session IDs over prefix matches and can open exact IDs outside the recent-session list. CLI exact-ID lookup also bypasses the --tasks discovery limit; history cutoffs still apply, so use --all for older sessions. Prefix matching searches only the discovered recent sessions.

Adding New Adapters

TokenMon uses a base class in src/tokenmon/adapters/base.py:

class BaseAdapter(ABC):
    @property
    @abstractmethod
    def name(self) -> str: ...

    @abstractmethod
    def detect(self) -> bool: ...

    @abstractmethod
    def collect(self, max_sessions: int = 64, min_timestamp: float | None = None) -> list[GenerationSpan]: ...

    @abstractmethod
    def collect_sessions(self, max_sessions: int = 32, min_timestamp: float | None = None) -> list[SessionTimeline]: ...

To add support for a new agent (e.g. OpenCode):

1. Create src/tokenmon/adapters/opencode.py subclassing BaseAdapter.
2. Implement discovery (detect()), stream parsing (collect()), and timelines (collect_sessions()).
3. Register it in src/tokenmon/adapters/__init__.py.

Adapters translate telemetry into shared GenerationSpan and SessionTimeline/TimelineEvent models; the analyzer and CLI own aggregation, formatting, and JSON serialization. Build spans with create_span() so invalid measurements remain visible in exclusion counts. Honor collection limits and completion-time cutoffs without discarding earlier metadata context.

For live follow, override read_session(session_id) to reopen a retained exact source, and give events stable internal event_id values that do not change with usage or timestamps. OMPAdapter and PiAdapter share src/tokenmon/adapters/_pi_jsonl.py for nested journals and pinned reads; their separate entry points preserve native roots and timing semantics.

License

MIT

Copied History and TPS

Workers and forked sessions may inherit earlier assistant outputs. Ownership means counting those outputs only under the session that generated them—not again under a child that copied them.

For example, 1,000 main-session tokens in 10s plus 200 worker tokens in 4s gives 1,200 tokens and 85.7 weighted TPS. Counting the copied main output again, including its duration, gives 2,200 tokens and 91.7 TPS: the main response gets too much weight. Duplicating every response equally leaves weighted TPS unchanged, but totals are still wrong. TokenMon uses recorded ownership boundaries where available; see Agent Data above for format limitations.