convez.ai
AgenteDevOps & InfraestruturaIntermediário
xihuai18

Agentic Metric X

xihuai18/agentic-metric

Monitor de tokens e custos para agentes de IA, estilo 'top' para código

Ver no GitHub
InicialPluginScore · de 0 a 100

Monitor local de uso e custos para agentes de IA como Claude Code e Codex, com dashboard TUI em tempo real e relatórios detalhados. Todos os dados permanecem na sua máquina — sem telemetria externa.

Instalação

$npx convez-ai xihuai18/agentic-metric

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
4
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
3

Casos de uso

  • Rastrear consumo de tokens por sessão e modelo de IA
  • Gerar relatórios de custos por agente, período e projeto
  • Monitorar uso remoto via SSH com cache local
  • Analisar padrões de uso com heatmaps e gráficos

Veredito

Ferramenta prática e privada para quem usa Claude Code ou Codex em produção. Dashboard responsivo e arquitetura extensível compensam a curva de aprendizado inicial. Recomendado para equipes que precisam de visibilidade de custos sem depender de plataformas externas.

README

A tradução para o português deste README ainda está na fila.

Agentic Metric X

PyPI Python Downloads Downloads/month

中文文档

A local and SSH remote monitoring tool for AI coding agents — like top, but for your coding agents. Track token usage and costs across Claude Code and Codex — with a TUI dashboard and CLI.

Supported platforms: Linux, macOS, and Windows.

All data stays under your control. No telemetry is sent anywhere. By default the tool only reads local agent data files (~/.claude/, ~/.codex/). If SSH remotes are configured, it reads those remote agent data files over SSH and caches a local copy for aggregation.

Agentic Metric TUI

Features

  • Cost estimation — Per-model pricing table with CLI management, calculates API-equivalent costs; supports long-context and cache-duration pricing
  • Unified report — One report command for today / week / month / custom date range, with agent × provider × model breakdown, top projects, cost drivers, and hourly/daily/weekly heatmaps
  • TUI dashboard — Terminal UI with auto-refresh, stacked summary cells, heatmap strip, 30-day cost chart, and agent → provider → model breakdown
  • Multi-agent — Plugin architecture; supports Claude Code and Codex today, extensible

Agent Data Coverage

FieldClaude CodeCodex
Session ID✓✓
Project path✓✓
Git branch✓✓
Model✓✓
Input tokens✓✓
Output tokens✓✓
Cache tokens✓✓¹
User turns✓✓
Message count✓✓
First/last prompt✓✓
Cost estimation✓✓

¹ Codex exposes cache-read tokens only; cache-write is not reported. Codex's input_tokens already includes cached tokens, so the collector stores input_tokens − cached_input_tokens to avoid double-charging at both input and cache-read pricing.

Installation

Requires Python 3.10+.

pip install agentic-metric-x

Or with uv:

uvx agentic-metric-x              # Run directly without installing
uv tool install agentic-metric-x   # Or install persistently
uv tool upgrade agentic-metric-x   # Upgrade to latest version

Usage

agentic-metric                       # Launch TUI dashboard (default when no command given)
agentic-metric tui                   # Launch TUI dashboard explicitly
agentic-metric sync                  # Force sync collectors to the local database
agentic-metric sync --rebuild        # Rebuild the derived database from source session logs
agentic-metric report --today        # Today's usage report
agentic-metric report --week         # This week (Mon → today)
agentic-metric report --month        # This month
agentic-metric report --range 2026-04-01:2026-04-23   # Custom date range
agentic-metric today                 # Shortcut for `report --today`
agentic-metric week                  # Shortcut for `report --week`
agentic-metric month                 # Shortcut for `report --month`
agentic-metric history -d 30         # Last N days (default 14)
agentic-metric pricing               # Manage model pricing

Report Options

OptionDescription
--todayToday's usage
--weekThis week (Mon → today)
--monthThis month
--range FROM:TOCustom date range, e.g. 2026-04-01:2026-04-23
--fullShow extra drill-down tables (by host, agent, provider, model, and time)
--limit N / -n NRows in driver tables and per-provider model breakdown (1–25, default 8)
--jsonOutput machine-readable JSON instead of tables (for scripts / pipes)
--watch N / -w NRefresh the report every N seconds (Ctrl-C to stop)
--no-syncSkip syncing collectors before querying

report shows a header with total cost / sessions / turns / tokens / cache-hit rate, a delta vs. the previous equivalent period, a compact provider cost rollup, a heatmap strip (hours for --today, days for --week, weeks for --month), plus a default agent × provider × model breakdown, top projects, top cost drivers, and optional extra drill-down tables.

Pricing Management

Model pricing is used for cost estimation. Builtin pricing is included for common models. You can add new models, override builtins, configure long-context rates, and configure observable cache-duration rates via CLI. Overrides are stored in $DATA/agentic_metric/pricing.json.

Basic model pricing

agentic-metric pricing list                                                # List all model pricing
agentic-metric pricing set deepseek-r2 -i 0.5 -o 2.0                       # Add a new model
agentic-metric pricing set claude-opus-4-7 -i 4.0 -o 20.0 -cr 0.4 -cw 5.0  # Override builtin
agentic-metric pricing reset deepseek-r2                                   # Reset one model to builtin
agentic-metric pricing reset --all                                         # Reset all overrides

Long-context pricing

Some models charge higher rates when a single request exceeds a token threshold. The tool applies these rates per-event when collectors provide event-level usage.

agentic-metric pricing long-context set gpt-5.5 --threshold 272000 -i 10 -o 45 -cr 1 -cw 0
agentic-metric pricing long-context reset gpt-5.5        # Remove user override
agentic-metric pricing long-context disable gpt-5.5      # Disable builtin rule
agentic-metric pricing long-context enable gpt-5.5       # Re-enable builtin rule

Cache-duration pricing

Anthropic charges different cache-write prices depending on cache TTL. By default the tool uses the 5-minute rate; override for 1-hour cache duration when applicable.

agentic-metric pricing cache set claude-sonnet-4 --write-1h 6    # Set 1h cache write price
agentic-metric pricing cache reset claude-sonnet-4                # Remove override

Unknown models are not priced by default. They are displayed as Unknown with cost ? until you add explicit pricing with agentic-metric pricing set. Provider speed/priority modes are not shown or priced separately because the local history files do not expose reliable non-standard markers.

After a pricing change, the command resyncs local history so event-level costs such as long-context requests are recalculated from the original JSONL data.

If cached history ever looks stale after changing roots/providers, run agentic-metric sync --rebuild. This deletes only this tool's derived SQLite database and rebuilds it from the original Claude Code and Codex session files; config.json, pricing.json, and agent data directories are left untouched.

TUI Keybindings

Shown in the footer as: PgUp/PgDn Range · ←→ View · . Now · r Auto-refresh · p Pricing · ? Help · q Quit.

KeyFooterAction
← / →ViewSwitch view (Today / Week / Month)
PageUp / PageDownRangeMove time range earlier / later
.NowJump back to "now" (reset offset)
↑ / ↓—Scroll the breakdown panel
rAuto-refreshToggle fast auto-refresh; pauses the slow periodic sync while active
pPricingOpen the read-only pricing view (flags unknown models in range)
?HelpShow the keybinding cheatsheet
qQuitQuit

Data auto-syncs every 5 min by default; there is no manual one-shot sync key. To copy data out of the dashboard, use the CLI (agentic-metric report / today / week / month). The heatmap panel shows the selected range's provider cost rollup; the trend panel shows provider totals over the longer trend window.

Refresh intervals can be overridden in the config file ($DATA/agentic_metric/config.json):

{ "intervals": { "data_sync": 300, "auto_refresh": 30 } }

Builtin Model Pricing

Prices are USD per 1M tokens. Verified against official pricing docs (2026-04-25; Claude Fable 5 on 2026-06-12; Claude Opus 4.8 on 2026-06-02).

ModelInputOutputCache ReadCache Write

| claude-fable-5 | $10.00 | $50.00 | $1.00 | $12.50 | | claude-opus-4-8 / claude-opus-4-7 / claude-opus-4-6 / claude-opus-4-5 | $5.00 | $25.00 | $0.50 | $6.25 | | claude-opus-4-1 / claude-opus-4 / claude-3-opus | $15.00 | $75.00 | $1.50 | $18.75 | | claude-sonnet-4-6 / claude-sonnet-4-5 / claude-sonnet-4 / claude-sonnet-3-7 / claude-3-7-sonnet / claude-3-5-sonnet | $3.00 | $15.00 | $0.30 | $3.75 | | claude-haiku-4-5 | $1.00 | $5.00 | $0.10 | $1.25 | | claude-haiku-3-5 / claude-3-5-haiku | $0.80 | $4.00 | $0.08 | $1.00 | | claude-3-haiku | $0.25 | $1.25 | $0.03 | $0.30 |

ModelInputOutputCache ReadCache Write

| gpt-5.5 | $5.00 | $30.00 | $0.50 | — | | gpt-5.4-mini | $0.75 | $4.50 | $0.075 | — | | gpt-5.4-nano | $0.20 | $1.25 | $0.02 | — | | gpt-5.4 | $2.50 | $15.00 | $0.25 | — | | gpt-5.2-codex / gpt-5.2-chat-latest / gpt-5.2 / gpt-5.3-codex / gpt-5.3-chat-latest / gpt-5.3 | $1.75 | $14.00 | $0.175 | — | | gpt-5.1-codex-max / gpt-5.1-codex / gpt-5.1-chat-latest / gpt-5.1 / gpt-5-codex / gpt-5-chat-latest / gpt-5 | $1.25 | $10.00 | $0.125 | — |

ModelInputOutputCache ReadCache Write

| gemini-3.5-flash | $1.50 | $9.00 | $0.15 | — | | gemini-3.1-pro / gemini-3-pro | $2.00 | $12.00 | $0.20 | — | | gemini-3.1-flash-lite | $0.25 | $1.50 | $0.025 | — | | gemini-3-flash | $0.50 | $3.00 | $0.05 | — | | gemini-2.5-flash | $0.30 | $2.50 | $0.03 | — | | gemini-2.5-flash-lite | $0.10 | $0.40 | $0.01 | — | | gemini-2.0-flash | $0.10 | $0.40 | $0.025 | — | | gemini-2.0-flash-lite | $0.075 | $0.30 | $0.00 | — |

Run agentic-metric pricing list for the full table including your overrides.

Architecture

src/agentic_metric/
├── cli.py              # Typer CLI commands and Rich report rendering
├── config.py           # Platform paths, collector roots, SSH remote specs
├── pricing.py          # Builtin + user pricing, cost estimation engine
├── formatting.py       # Pure formatting helpers (cost/tokens, source/host labels)
├── collectors/
│   ├── __init__.py     # Collector registry and base class
│   ├── claude_code.py  # Claude Code JSONL history parser
│   ├── codex.py        # Codex JSONL history parser
│   └── remote.py       # SSH-backed wrapper: mirror a remote root, then parse
├── store/
│   ├── __init__.py
│   ├── database.py     # SQLite database (sessions, session_usage buckets)
│   └── aggregator.py   # Query layer: range totals, heatmaps, breakdowns
└── tui/
    ├── __init__.py
    ├── app.py            # Textual TUI application
    ├── widgets.py        # Custom TUI widgets (summary cells, heatmap, trend)
    ├── help_screen.py    # Keybinding cheatsheet modal
    ├── pricing_screen.py # Read-only pricing view (flags unknown models)
    └── styles.tcss       # Textual CSS

Data flow

  1. Collectors read agent data files (~/.claude/, ~/.codex/) and sync session history into the database.
  2. Database stores sessions and per-day session_usage buckets in SQLite.
  3. Aggregator runs SQL queries for reports (range totals, heatmaps, breakdowns by agent/model/project).
  4. CLI renders Rich tables and panels. TUI uses Textual for the dashboard.
  5. Pricing engine calculates costs per-event (long-context aware).

Data Sources

Paths differ by platform. $DATA refers to:

LinuxmacOSWindows
$DATA~/.local/share~/Library/Application Support%LOCALAPPDATA%
AgentPathData
Claude Code~/.claude/projects/JSONL sessions, token usage, model, branch
Claude Code~/.claude/stats-cache.jsonDaily activity stats
Codex~/.codex/sessions/JSONL sessions, token usage, model

By default, Claude Code honors CLAUDE_CONFIG_DIR and Codex honors CODEX_HOME. To scan multiple roots, assign a provider explicitly, or include SSH remotes, create this tool's own config file:

{
  "collectors": {
    "codex": {
      "roots": [
        {"path": "~/.codex", "provider": "openai"},
        {"path": "~/.codex-custom", "provider": "custom"}
      ]
    },
    "claude_code": {
      "roots": [
        {"path": "~/.claude"},
        {"path": "~/.claude-alt"},
        {"path": "~/.claude-provider-b", "provider": "provider-b"}
      ]
    }
  },
  "remotes": [
    {
      "name": "remote-dev",
      "host": "remote-dev",
      "collectors": {
        "codex": {
          "roots": [{"path": "~/.codex", "provider": "openai"}]
        },
        "claude_code": {
          "roots": [{"path": "~/.claude"}]
        }
      }
    }
  ]
}

The default config path is $DATA/agentic_metric/config.json; set AGENTIC_METRIC_CONFIG to point at another JSON file. Claude Code roots without a provider are not guessed; Codex roots without a provider can still fall back to the JSONL model_provider.

Remote entries use your existing ssh config. host may be an SSH alias or a hostname; optional fields are name, user, port, timeout, and ssh_options. If a remote does not configure collector roots, it reuses the local collector root configuration; with no local config that means ~/.claude and ~/.codex on the remote. Remote sync expands ~ on the remote host, reads projects/ and sessions/ over SSH, stores a cache under $DATA/agentic_metric/remote-cache/, and aggregates those rows with local usage. It also keeps a remote file manifest in the cache, so repeated syncs only download changed session/index files instead of transferring full agent directories again. Cached files that disappear from the remote manifest are moved under .stale/ and no longer parsed; usage already recorded from them stays in the database, so historical reports keep it. If a remote path is missing, that collector is skipped rather than reusing old cache contents. CLI/TUI totals stay merged, while breakdowns and top projects retain the host/source dimension.

Reports keep the headline totals merged across local and remote data. Detail tables use a compact Source label: local rows show the root (~/.codex), and remote rows show host:root (remote-dev:~/.codex). Top projects use the same prefix for remote paths (remote-dev:/workspace/project) so matching local and remote project paths are never silently merged. Multiple local roots that share a project path collapse into one row (they would render identically), so the same project worked on through two local agent homes is not double-listed. --full adds a provider rollup by source; the default report keeps the source × agent × provider × model breakdown plus top projects.

All aggregated data is stored locally in $DATA/agentic_metric/data.db (SQLite).

Unsupported Agents

  • Cursor — Cursor stopped writing token usage data (tokenCount) to its local state.vscdb database around January 2026 (approximately version 2.0.63+). All inputTokens/outputTokens values are now zero. Cursor has moved usage tracking to a server-side system. Since this tool is designed to be fully offline with no network requests, monitoring Cursor usage is not supported.
  • OpenCode / Qwen Code / VS Code Copilot Chat — collectors for these agents existed up to v0.1.8 and were removed in v0.2.0 as this fork narrowed its focus to Claude Code and Codex. If you need them, stay on v0.1.8 from upstream.

Privacy

  • Local by default — no network requests unless SSH remotes are configured
  • Read-only — never modifies agent config or data files
  • All stats stored in a local SQLite database
  • Delete the data directory at any time to remove all data (~/.local/share/agentic_metric/ on Linux, ~/Library/Application Support/agentic_metric/ on macOS, %LOCALAPPDATA%\agentic_metric\ on Windows)

Development

git clone https://github.com/xihuai18/agentic-metric
cd agentic-metric
pip install -e ".[dev]"
pytest

License

MIT — see LICENSE.

Forked from MrQianjinsi/agentic-metric (v0.1.8 upstream). See CHANGELOG.md for what changed in this fork.

Ver o original no GitHub

Bom para

Relacionados