agentic-metric 0.1.1

Agentic coding metric monitor — multi-agent, TUI, CLI

Agentic Metric

中文文档

A local-only monitoring tool for AI coding agents. Track token usage and costs across Claude Code, Codex, Cursor, OpenCode, VS Code (Copilot Chat), and more — with a TUI dashboard and CLI.

All data stays on your machine. No network requests, no telemetry, no data leaves your computer. The tool only reads local agent data files (e.g. ~/.claude/) and process info.

Features

• Live monitoring — Detect running agent processes, incremental JSONL session parsing
• Cost estimation — Per-model pricing table, calculates API-equivalent costs
• Today overview — Sessions, token usage, and cost summary for the current day
• Historical trends — 30-day daily token/cost trends
• TUI dashboard — Terminal UI with 1-second live refresh, stacked token charts, and trend lines
• Multi-agent — Plugin architecture, supports Claude Code and Cursor, extensible

Data Sources

Agent	Path	Data
Claude Code	~/.claude/projects/	JSONL sessions, token usage, model, branch
Claude Code	~/.claude/stats-cache.json	Daily activity stats
Cursor	~/.config/Cursor/User/globalStorage/state.vscdb	Composer sessions, token usage, model
Cursor	Process detection	Running status, working directory
Codex	~/.codex/sessions/	JSONL sessions, token usage, model
VS Code	~/.config/Code/User/workspaceStorage/*/chatSessions/	Chat sessions (JSON + JSONL), token usage (JSONL only), model
VS Code	~/.config/Code/User/globalStorage/emptyWindowChatSessions/	Empty-window chat sessions
VS Code	Process detection	Running status, working directory
OpenCode	~/.local/share/opencode/opencode.db	SQLite sessions, messages, token usage, model
OpenCode	Process detection	Running status, active session matching

All aggregated data is stored locally in ~/.local/share/agentic_metric/data.db (SQLite).

Installation

pip install agentic-metric

Usage

agentic-metric status          # Show currently active agents
agentic-metric today           # Today's usage overview
agentic-metric history         # Historical trends (default 30 days)
agentic-metric history -d 7    # Last 7 days
agentic-metric sync            # Force sync data to local database
agentic-metric tui             # Launch TUI dashboard
agentic-metric bar             # One-line summary for status bars

Status Bar Integration

agentic-metric bar outputs a compact one-line summary (e.g. AM: $1.23 | 4.5M) for embedding into status bars like i3blocks, waybar, tmux, vim statusline, etc.

i3blocks / waybar:

[agentic-metric]
command=agentic-metric bar
interval=60

tmux:

set -g status-right '#(agentic-metric bar | head -1)'
set -g status-interval 60    # refresh every 60 seconds (default 15)

vim / neovim statusline:

set statusline+=%{system('agentic-metric\ bar\ \|\ head\ -1')}
" statusline refreshes on cursor move, mode change, etc.
" to force a periodic refresh, add a timer:
autocmd CursorHold * redrawstatus
set updatetime=60000          " trigger CursorHold after 60s idle

TUI Keybindings

Key	Action
q	Quit
r	Refresh data
Tab	Switch Dashboard / History tab

Agent Data Coverage

Different agents expose different levels of local data. Here's what's available for each:

Field	Claude Code	Codex	Cursor	VS Code (Copilot)	OpenCode
Session ID	✓ JSONL	✓ JSONL	✓ composerId	✓ sessionId	✓ session table
Project path	✓ JSONL	✓ JSONL	◐ partial (from bubble or conversationState)	✓ workspace.json URI	✓ session.directory (launch cwd)
Git branch	✓ JSONL	✓ JSONL	✗ not stored	✗ not stored	✗ not stored
Model	✓ JSONL	✓ JSONL	✓ modelConfig / bubble modelInfo	✓ result.details (e.g. "Claude Haiku 4.5 • 1x")	✓ message.modelID
Input tokens	✓ per-message	✓ cumulative	◐ older versions only	◐ JSONL format only	✓ per-message
Output tokens	✓ per-message	✓ cumulative	◐ older versions only	◐ JSONL format only	✓ per-message (includes reasoning)
Cache tokens	✓ read + write	✓ read only	✗ not exposed	✗ not exposed	◐ read only (write always 0)
User turns	✓	✓	✓	✓	✓
Message count	✓ user + assistant (excl. tool_result)	✓ user + assistant	✓ conversation headers	✓ turns × 2	✓ user + assistant
First/last prompt	✓	✓	✓ from bubble text	✓ message.text	✓ from part table
Cost estimation	✓	✓	◐ only when tokens available	◐ only when tokens available	◐ estimated only (reported cost always 0)
Live active status	✓ PID + session file match	✓ PID + session file match	◐ process-level only (latest session marked active)	◐ process-level only	✓ PID + DB session match

Key differences:

• Claude Code & Codex — Each running process maps to a JSONL session file with a unique session ID. This allows precise matching between live processes and DB sessions for accurate active status.
• Cursor — Live detection only sees the process PID, while historical sessions use composer UUIDs from state.vscdb. There is no way to link a running Cursor process to a specific composer session, so the most recent session is marked active when the process is running.
• Token coverage — Cursor stored per-bubble tokenCount (input/output) in state.vscdb in earlier versions (~2025-05 to ~2025-12), but newer versions no longer populate this field — all values are zero. Cursor appears to have moved usage tracking to a server-side system, so local token data is unavailable for recent sessions. Cache token breakdown (read/write) has never been available.
• Model name — Cursor's "default" model setting doesn't record which model was actually used on the backend. These sessions show default in the model column.
• VS Code (Copilot Chat) — Has two storage formats: legacy JSON (older sessions, no token data) and newer incremental JSONL (with result.usage containing promptTokens/completionTokens). Token usage is only available for sessions stored in JSONL format. Model names are extracted from Copilot's display strings (e.g. "GPT-4o • 1x") and normalized to pricing keys. Workspace paths support local (file://), SSH remote (vscode-remote://ssh-remote+host), and container (attached-container+...) URIs.
• OpenCode — Stores all data in a local SQLite database (opencode.db). Token data is per-message with input, output, reasoning, and cache.read/cache.write fields. Reasoning tokens are counted as output tokens (billed at output rate). The cost field in messages is always 0, so all costs are estimated using the pricing table. cache.write is also always 0.

Privacy

• Fully offline — no network requests, no data sent anywhere
• Read-only — never modifies agent config or data files
• All stats stored in a local SQLite database
• Delete ~/.local/share/agentic_metric/ at any time to remove all data