Vibe Coding Tracker：用 Rust 实时追踪 Claude Code、Codex 等 AI 编程成本

Track your AI coding costs in real-time. Vibe Coding Tracker is a lightweight, high-performance CLI tool built in Rust that monitors and analyzes your Claude Code, Codex, Copilot, Gemini, OpenCode, Cursor, Hermes, Grok, and DeepSeek Harness usage — with detailed cost breakdowns, token statistics, and code operation insights, all while keeping the memory footprint minimal.

English | 繁體中文 | 简体中文

Note: CLI examples use the short alias vct. If you installed via npm/pip/cargo, the binary might be named vibe_coding_tracker or vct. Create an alias or replace vct with the full name when running commands if needed.

Why Vibe Coding Tracker?

Know Your Costs

Stop wondering how much your AI coding sessions cost. Get real-time cost tracking with automatic pricing updates from LiteLLM.

Ultra-Lightweight

Built with Rust for minimal resource footprint. The interactive TUI dashboard typically sits at around 60 MB of resident memory once the first refresh is done, even with thousands of sessions on disk — no Electron, no bloated runtimes. A persistent per-session ledger under ~/.vct/sessions/ means only new or changed sources are ever parsed again, even across runs.

Beautiful Visualizations

Choose your preferred view:

• Interactive Dashboard: Responsive terminal UI with an immediate loading spinner, background incremental refreshes, a scrollable model list (arrow keys), a live per-process CPU/memory readout, and compact K/M/B number formatting
• Static Reports: Professional tables for documentation
• Script-Friendly: Plain text and JSON for automation
• Full Precision: Export exact costs for accounting

Zero Configuration

Automatically detects and processes logs from Claude Code, Codex, Copilot, Gemini, OpenCode, Cursor, Hermes, Grok, and DeepSeek Harness. No setup required — just run and analyze. A ~/.vct/config.toml is created with sensible defaults on first run if you ever want to tweak behavior (see Configuration).

Rich Insights

• Token usage by model and date
• Cost breakdown by cache types (read / create)
• File operations tracking (edit, read, write lines)
• Tool call history (Bash, Edit, Read, Write, TodoWrite)
• Per-provider totals

Key Features

Quick Start

Installation

Choose the installation method that works best for you:

Method 1: Install from npm

Prerequisites: Node.js v22 or higher

Choose one of the following package names (they are identical):

# Main package
npm install -g vibe-coding-tracker

# Short alias with scope
npm install -g @mai0313/vct

# Full name with scope
npm install -g @mai0313/vibe-coding-tracker

Method 2: Install from PyPI

Prerequisites: Python 3.8 or higher

pip install vibe_coding_tracker
# Or with uv
uv pip install vibe_coding_tracker

# Run without installing, straight from PyPI (uv)
uvx vibe_coding_tracker usage

Method 3: Install from crates.io

Install using Cargo from the official Rust package registry:

cargo install vct-cli

First Run

# View your usage with the interactive dashboard
vct usage

# Or run the binary built by Cargo/pip
vibe_coding_tracker usage

# Analyze code operations across all sessions
vct analysis

Command Guide

Run vct --help to list the commands, and vct <command> --help for a command's flags.

Time range flags (shared by usage and analysis, mutually exclusive, default --all):

Usage Command

Track your spending across all AI coding sessions.

Basic Usage

# Interactive dashboard (recommended)
vct usage

# Static table for reports
vct usage --table

# Plain text for scripts
vct usage --text

# JSON for data processing (every row has cost_usd; matched_model only when priced under another model name)
vct usage --json

# Save enriched JSON with shell redirection
vct usage --json > report.json

# Combine time range with output format
vct usage --weekly
vct usage --table --monthly
vct usage --json --daily

# Merge same model reported under different provider prefixes
# (e.g. openai/gpt-5.5 + azure/gpt-5.5 + gpt-5.5 -> one row)
vct usage --table --merge-providers

Note

Model rows are sorted by cost in ascending order, so the highest-spending model is listed last (right above the TOTAL row in --table). This applies to the interactive dashboard, --table, and --text output; --json lists its rows in no fixed order. The interactive dashboard also hides models with zero usage in the selected range.

Tip

The same model can show up as several rows when it is routed under different provider prefixes (openai/gpt-5.5, azure/gpt-5.5, plain gpt-5.5). --merge-providers collapses rows that share the base name after the first / (versions like gpt-5.5 vs gpt-5.4 stay separate) and sums their already-priced cost. In the interactive dashboard, press m to toggle it live (the choice is saved to ~/.vct/config.toml, so the next launch remembers it); --merge-providers opens the dashboard already merged. --json is left as the raw per-model export.

Preview: Interactive Dashboard (vct usage)

┌ Models · $79.33 · 49.1M tokens · 3 models ───────────────────────────────────────────────────────────────────────────┐
│Model                              Input     Output    Cache Read  Cache Write Total     Cost (USD)   Share           █
│                                                                                                                      █
│gemini-3.1-pro-preview                  129K     10.3K       66.0K       1.70K      207K        $0.40             0.4%█
│claude-haiku-4-5-20251001              5.57K     19.8K       5.20M       55.0K     5.28M        $1.34          █ 10.8%█
│claude-opus-4-8                        25.7K      179K       42.9M        496K     43.6M       $77.59  █████████ 88.8%█
│                                                                                                                      █
│                                                                                                                      █
└──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
                                              Claude $78.93 · Gemini $0.40
┌ Claude ──────────────┐┌ Codex ───────────────┐┌ Copilot ─────────────┐┌ Cursor ────────LIMIT ┐┌ Grok ────────────────┐
│max 20x       just now││plus          just now││individual    just now││free          just now││SuperGrok     just now│
│5h    ▰▱▱▱▱  13% 1h42m││5h    ▰▰▱▱▱  33% 12m  ││prem  ▰▱▱▱▱   3% 24d0h││total ▰▰▰▰▰  94% 16d0h││week  ▰▰▱▱▱  38% 3d4h │
│7d    ▰▰▰▱▱  58% 1d23h││7d    ▰▰▱▱▱  36% 1h54m││reqs  ▰▱▱▱▱ 45/1500   ││auto  ▰▰▰▰▰ 100% 16d0h││ondmd $4.20/$50.00    │
│Fable ▰▰▰▰▱  79% 1d23h││credits 0 · +3 reset  ││                      ││api   ▰▰▰▱▱  44% 16d0h││balance $12.00        │
│bal - · $0.00 used    ││reset expires 17d0h   ││                      ││                      ││                      │
└──────────────────────┘└──────────────────────┘└──────────────────────┘└──────────────────────┘└──────────────────────┘
                Total Cost: $79.33  |  Total Tokens: 49.1M  |  Models: 3  |  Memory: 4.7 MB  |  CPU: 0.0%
                    ↑/↓ scroll  m merge  p providers  Q quota  r refresh  q quit  |  Star on GitHub

Both interactive dashboards draw a centered Loading sessions... spinner as soon as terminal setup finishes. Loading stays responsive to q, Ctrl+C, and resize events. Later scans keep the last successful data visible with a Refreshing... footer. A failed refresh keeps the last-known-good view and retries on the next scheduled or manual refresh.

Preview: Table & JSON (vct usage)

--table prints the same numbers as a static report with a per-provider summary; --json emits one enriched row per model (each with cost_usd) for scripting.

Token Usage Statistics

┌───────────────────────────┬─────────┬─────────┬─────────────┬─────────────┬──────────────┬────────────┐
│ Model                     ┆   Input ┆  Output ┆  Cache Read ┆ Cache Write ┆ Total Tokens ┆ Cost (USD) │
╞═══════════════════════════╪═════════╪═════════╪═════════════╪═════════════╪══════════════╪════════════╡
│ opencode/gemini-3.5-flash ┆  19,421 ┆     254 ┆           0 ┆           0 ┆       19,675 ┆      $0.03 │
│ gpt-5.5                   ┆ 242,227 ┆  16,229 ┆   2,406,912 ┆           0 ┆    2,665,368 ┆      $5.56 │
│ claude-opus-4-8           ┆ 401,937 ┆ 936,186 ┆ 138,099,926 ┆   6,057,836 ┆  145,495,885 ┆    $151.29 │
│ TOTAL                     ┆ 663,585 ┆ 952,669 ┆ 140,506,838 ┆   6,057,836 ┆  148,180,928 ┆    $156.88 │
└───────────────────────────┴─────────┴─────────┴─────────────┴─────────────┴──────────────┴────────────┘

Totals (by Provider)

┌───────────────┬─────────────┬─────────┐
│ Provider      ┆      Tokens ┆    Cost │
╞═══════════════╪═════════════╪═════════╡
│ Claude        ┆ 145,495,885 ┆ $151.29 │
│ Codex         ┆   2,665,368 ┆   $5.56 │
│ OpenCode      ┆      19,675 ┆   $0.03 │
│ All Providers ┆ 148,180,928 ┆ $156.88 │
└───────────────┴─────────────┴─────────┘

// vct usage --json  (one model shown; rows come in no fixed order)
[
  {
    "cost_usd": 151.29,
    "model": "claude-opus-4-8",
    "usage": {
      "cache_creation": {
        "ephemeral_1h_input_tokens": 0,
        "ephemeral_5m_input_tokens": 6057836
      },
      "cache_creation_input_tokens": 6057836,
      "cache_read_input_tokens": 138099926,
      "input_tokens": 401937,
      "output_tokens": 936186,
      "reasoning_output_tokens": 0,
      "total_tokens": 145495885
    }
  }
]

Every row carries the same flat token fields regardless of provider. A few keys appear only on a row that has something to put in them: cache_creation splits the cache writes into their 5-minute and 1-hour TTL shares, server_tool_use counts Claude web searches, above_tier holds the tokens of requests whose prompt crossed a long-context price tier, and a Gemini row that spent tool tokens carries tool_tokens, which LiteLLM has no rate for and which are counted in total_tokens alone.

What It Scans

The tool automatically scans these directories:

• ~/.claude/projects/**/*.jsonl (Claude Code — recursive, includes subagent logs)
• ~/.codex/sessions/**/*.jsonl (Codex — recursive, includes daily subdirectories) plus ~/.codex/archived_sessions/*.jsonl, the flat directory Codex moves a session's log into when you archive it (a session found in both is counted once)
• ~/.copilot/session-state/<sessionId>/events.jsonl (Copilot CLI)
• ~/.gemini/tmp/<project_hash>/chats/*.jsonl (Gemini CLI)
• ~/.local/share/opencode/opencode.db (OpenCode — SQLite database; honors $XDG_DATA_HOME)
• ~/.cursor/chats/*/*/store.db (Cursor — SQLite chat stores, used for analysis and a local usage estimate consistent with the other providers)
• ~/.hermes/state.db (Hermes — SQLite database, honors $HERMES_HOME; usage only)
• $GROK_HOME/sessions/*/*/signals.json (Grok CLI — defaults to ~/.grok; sibling updates.jsonl supplies analysis data)
• $DSH_HOME/sessions/*/*/session.jsonl.zstd (DeepSeek Harness — defaults to ~/.dsh; a root configured with compression = "none" writes session.jsonl instead, and both are read)

Grok usage is one point-in-time local context estimate: vct records signals.json's contextTokensUsed as cache-read tokens and estimates cost at the model's cache-read price. It is not cumulative billed usage. analysis reconstructs completed Read / Write / Edit / Bash / TodoWrite operations from the sibling updates.jsonl. For your actual billed allowance, see the Grok quota panel below.

For noninteractive usage and analysis scans, vct exits with an error when every discovered source fails. If only some sources fail, it keeps the successful results and prints one diagnostic summary to stderr. The TUI stays best-effort and preserves its last successful payload instead.

Session Ledger

Every usage and analysis scan reads and writes a per-session ledger under ~/.vct/sessions/ (one <provider>.json per assistant). A session that has not changed is never parsed twice, so one-shot runs and the dashboard's first paint only spend time on what changed. More importantly, a session stays in the ledger after its source is gone: when an assistant prunes its own history (Claude Code keeps 30 days by default) or a whole session directory such as ~/.claude/projects is deleted, every total still includes it, and noninteractive runs print a one-line note saying how many sessions came from the ledger. A resumed session is re-read when its log changes. The ledger holds only per-day token counts and tool counters in each provider's own shape, never transcripts or prices, and vct never deletes it: an unreadable file is renamed aside and a file written by a newer vct is left alone. Delete the directory to start over from what is on disk. vct analysis --json and vct analysis FILE parse their files in full and do not use it.

Live Quota Panels

vct usage shows live quota usage for Claude Code, Codex, GitHub Copilot, Cursor, and Grok right in the dashboard — with zero setup. No status-line hook, no credentials to enter: vct reads each provider's own credentials, calls its usage API on a background thread, and keeps the panels current while you work. Every gauge is percent used, so a full bar means the window is spent, not untouched. (Prefer a quieter dashboard? Trim panels under [usage.quota] in config.toml, or set it to [] to hide the quota cards entirely. The per-provider cost line under the table is local scan data, not a quota panel, so it stays whatever you set here.)

┌ Claude ──────────────┐┌ Codex ───────────────┐┌ Copilot ─────────────┐┌ Cursor ────────LIMIT ┐┌ Grok ────────────────┐
│max 20x       just now││plus          just now││individual    just now││free          just now││SuperGrok     just now│
│5h    ▰▱▱▱▱  13% 1h42m││5h    ▰▰▱▱▱  33% 12m  ││prem  ▰▱▱▱▱   3% 24d0h││total ▰▰▰▰▰  94% 16d0h││week  ▰▰▱▱▱  38% 3d4h │
│7d    ▰▰▰▱▱  58% 1d23h││7d    ▰▰▱▱▱  36% 1h54m││reqs  ▰▱▱▱▱ 45/1500   ││auto  ▰▰▰▰▰ 100% 16d0h││ondmd $4.20/$50.00    │
│Fable ▰▰▰▰▱  79% 1d23h││credits 0 · +3 reset  ││                      ││api   ▰▰▰▱▱  44% 16d0h││balance $12.00        │
│bal - · $0.00 used    ││reset expires 17d0h   ││                      ││                      ││                      │
└──────────────────────┘└──────────────────────┘└──────────────────────┘└──────────────────────┘└──────────────────────┘

• Claude — plan tier, 5-hour, weekly, and per-model weekly usage from the official OAuth usage API (GET /api/oauth/usage), read from ~/.claude/.credentials.json, plus your credit balance. Polled about once a minute to stay under the endpoint's rate limit; a red LIMIT flag appears in the title when a cap is hit. The per-model weekly row is best-effort and simply hides when that scope is not returned.
• Codex — plan tier, available 5-hour and weekly usage windows, credit balance, and the earliest fetched available earned-reset expiry from the ChatGPT backend (wham/usage + wham/rate-limit-reset-credits) using ~/.codex/auth.json (with approximate remaining messages / spend cap when applicable); falls back to the newest rate_limits in your Codex session logs when the API is unavailable (the title shows Codex vs Codex (session)).
• Copilot — plan tier plus your premium-request quota, shown as two gauges: percent used and the used / total request count (e.g. 45/1500), from GitHub's Copilot API (GET /copilot_internal/user), read from ~/.copilot/config.json. The request impersonates the Copilot CLI. The token is long-lived, so there is no refresh; a 401 / 403 shows a run: copilot login hint.
• Cursor — plan tier, total / auto / API percent used, and on-demand spend from cursor.com (GET /api/usage-summary), using the session token in ~/.config/cursor/auth.json.
• Grok — plan tier plus your included-allowance usage for the current weekly or monthly period, from the Grok CLI's own billing endpoint (GET /v1/billing?format=credits), read from ~/.grok/auth.json. A prepaid balance is shown only once it is non-zero, and pay-as-you-go spend once you have either spent something or set a cap. The request impersonates the Grok CLI; a 401 / 403 shows a run: grok login hint.

Automatic token refresh. For Claude, Codex, and Grok, when a token is near expiry or rejected, vct refreshes it and writes the new token back to the provider's own credential file (in that CLI's exact format). For Grok, every other login in the file is preserved on write. If a refresh cannot proceed, the panel shows a run: <provider> auth login hint instead of breaking. Copilot (long-lived token) and Cursor (kept fresh by its own client) are read-only — vct never writes their credential files.

A panel appears only for a provider whose credentials are present. Panels are placed on a uniform grid that fits as many cards per row as the terminal allows and wraps the rest onto the next row. On a terminal too short to spend rows on the grid, it folds into a one-line digest showing the gauges that still fit and counting the rest; press Q for the full-detail overlay, where every card has room for every line whatever the terminal size, and p for the Provider Usage panel with per-provider tokens and share. Quota panels appear only in the interactive TUI; --table, --text, and --json are unchanged.

Platform note: on macOS, Claude Code stores its OAuth credentials in the system Keychain rather than ~/.claude/.credentials.json, so the Claude panel is not shown on macOS. Cursor's ~/.config/cursor credential path is Linux-oriented.

Analysis Command

Deep dive into code operations — see exactly what your AI assistant did.

See tests/fixtures/sessions/ for sample inputs and matching JSON outputs for the four JSONL providers, plus the Grok session fixture under tests/fixtures/sessions/grok/ and the DeepSeek Harness one under tests/fixtures/sessions/dsh/.

Basic Usage

# Interactive dashboard for all sessions (default)
vct analysis

# Static table output with per-provider totals
vct analysis --table

# Plain text for scripts
vct analysis --text

# Complete parser results for every session
vct analysis --json

# Analyze a single conversation file → stdout JSON
vct analysis ~/.claude/projects/session.jsonl

# Summarize only that conversation
vct analysis ~/.claude/projects/session.jsonl --table

# Save complete JSON with shell redirection
vct analysis --json > report.json
vct analysis ~/.claude/projects/session.jsonl > session-analysis.json

# Combine time range with output format
vct analysis --weekly
vct analysis --table --monthly
vct analysis --json --daily
vct analysis --json --daily > today.json

Preview: Interactive Dashboard (vct analysis)

┌ Models · 135K lines · 1.80K tool calls · 3 models ───────────────────────────────────────────────────────────────────┐
│Model                                   Edit Lines  Read Lines  Write Lines Bash    Edit    Read    TodoWrite  Write  █
│                                                                                                                      █
│claude-opus-4-8                               12.4K       88.1K       9.02K     412     231     604         96      88█
│gemini-3.1-pro-preview                        2.10K       14.8K       1.62K      84      44     112         16      15█
│claude-haiku-4-5-20251001                       840       5.40K         610      31      16      42          6       5█
│                                                                                                                      █
└──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
┌ Providers ───────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│Provider                 Edit Lines  Read Lines  Write Lines  Bash     Edit     Read     TodoWrite   Write    Days    │
│                                                                                                                      │
│Claude                   13.2K       93.5K       9.63K        443      247      646      102         93       12      │
│Gemini                   2.10K       14.8K       1.62K        84       44       112      16          15       3       │
└──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
                 Total Lines: 135K  |  Total Tools: 1.80K  |  Models: 3  |  Memory: 4.8 MB  |  CPU: 0.0%
                                    ↑/↓ scroll  r refresh  q quit  |  Star on GitHub

Preview: Table & JSON (vct analysis)

--table renders the per-model breakdown plus a per-provider summary (with an Active Days column). --text and --table are compact projections of the same normalized parser records. --json keeps the complete records, including per-operation details and token usage. With no <FILE>, the outer array contains one CodeAnalysis object per session; with <FILE>, stdout is that single object and matches the corresponding shape under tests/fixtures/sessions/.

Analysis Statistics

┌─────────────────┬────────────┬────────────┬─────────────┬──────┬──────┬──────┬───────────┬───────┐
│ Model           ┆ Edit Lines ┆ Read Lines ┆ Write Lines ┆ Bash ┆ Edit ┆ Read ┆ TodoWrite ┆ Write │
╞═════════════════╪════════════╪════════════╪═════════════╪══════╪══════╪══════╪═══════════╪═══════╡
│ gpt-5.5         ┆          0 ┆      3,087 ┆           0 ┆   25 ┆    0 ┆   10 ┆         0 ┆     0 │
│ claude-opus-4-8 ┆      1,493 ┆     15,564 ┆         970 ┆  123 ┆  134 ┆  144 ┆         0 ┆    12 │
│ TOTAL           ┆      1,493 ┆     18,651 ┆         970 ┆  148 ┆  134 ┆  154 ┆         0 ┆    12 │
└─────────────────┴────────────┴────────────┴─────────────┴──────┴──────┴──────┴───────────┴───────┘

Warning

Complete analysis JSON can be large and may contain source text, edit bodies, shell commands, absolute paths, repository URLs, user names, machine identifiers, and token metadata. Review it before sharing.

Batch analysis reads live provider stores. If an assistant appends to a session during the scan, later runs can legitimately contain newer data. Unchanged inputs serialize in a deterministic order.

Noninteractive analysis returns an error when every discovered source fails or uses an unrecognized schema. If only some sources fail, successful output is preserved and a warning is written to stderr.

analysis FILE follows the same rule for malformed or unsupported records inside one file: parsed JSON/text/table output is preserved on stdout and a generic skipped-record warning is written to stderr.

Codex code-mode sessions expose a completed JavaScript exec cell but no structured trace for its nested tools. VCT counts that cell as one Bash call and preserves its source in complete JSON, but does not guess nested Read/Edit/Write operations.

Update Command

Keep supported direct installations up to date automatically.

At startup, vct silently checks for a newer release at most once per UTC day. A newer release is downloaded and applied only when the current binary was installed by the official scripts/install.sh or scripts/install.ps1 installer and carries its .vct-managed marker. The check, download, and replacement never print to stdout or stderr. A network failure, rate limit, read-only directory, or another running vct process only leaves the current command running; the reason is recorded in the diagnostic log. VCT_OFFLINE=1 skips the check completely.

After a Unix update, vct re-executes the current command with the new binary. On Windows, the current command finishes and a helper applies the replacement afterwards. The automatic path never modifies installs from cargo, npm, PyPI, a distro package, or an unmarked manual or development build. Update those through their package manager instead. To add automatic updates to an existing direct install, run the current official installer once more so it creates the marker.

Set general.auto_update = false in ~/.vct/config.toml to disable startup checks and automatic updates. The explicit vct update command retains its existing manual behavior. VCT_OFFLINE=1 also applies to it: --check, the interactive prompt, and --force alike return without contacting GitHub.

Basic Usage

# Check for updates
vct update --check

# Interactive update with confirmation
vct update

# Force update — always downloads latest version
vct update --force

Preview (vct update --check)

Current version: v1.3.0
Checking for latest release...
Latest version: v1.3.0 — you are up to date!

Version Command

Report the embedded build metadata (binary version, Rust toolchain, Cargo version):

vct version          # Pretty table
vct version --text   # One-field-per-line, script-friendly
vct version --json   # Machine-readable JSON

┌───────────────┬──────────┐
│ Version       ┆ 1.3.0    │
│ Rust Version  ┆ 1.96.0   │
│ Cargo Version ┆ 1.96.0   │
└───────────────┴──────────┘

Quota Command

Print a provider's raw quota/usage API response — no parsing, no aggregation.

Calls the same quota endpoint the usage dashboard uses (Claude / Codex / Copilot / Cursor / Grok) exactly once and prints the raw body, so you can inspect the exact API shape or sanity-check your credentials. It reads each provider's stored credentials and does not refresh tokens: if a token is expired, re-auth with that provider's own CLI (claude / codex / copilot / cursor-agent / grok).

The previous name vct fetch is kept as a hidden alias, so existing scripts keep working.

Basic Usage

# Raw JSON (default)
vct quota claude
vct quota codex
vct quota copilot
vct quota cursor
vct quota grok

# Flattened plain text
vct quota codex --text

# Flattened key/value table
vct quota copilot --table

Note

The response body is printed to stdout as-is. On an HTTP error the body is still printed and the process exits non-zero; a 401/403 additionally prints a run: <cli> login hint on stderr.

Configuration

vct keeps its user settings in ~/.vct/config.toml. The file is created with defaults on first run, so you never have to write it by hand — edit it only when you want to change a default. It is generated from vct's typed settings and carries a #:schema directive on the first line, so a schema-aware TOML editor (taplo / VS Code "Even Better TOML") gives you autocomplete and validation. Print the schema yourself with vct config schema. A file written by an older vct is upgraded to the current layout in place the next time vct reads it (or on demand with vct config migrate), so an upgrade never leaves you on a stale format. That upgrade also adds a quota panel released after your file was written — once. Remove the name again and it stays removed, and a panel you had already dropped is never brought back.

Run vct config show to print the file; its comments explain every setting.

Note

Cursor usage is a local estimate from the chat stores, so it behaves like Claude Code / Codex / Copilot / Gemini (all computed from local session files) and needs no network. It undercounts Cursor's real spend, because much of it is billed under Cursor-internal model names the local data cannot price — treat Cursor cost as approximate.

Note

vct writes diagnostics to ~/.vct/logs/vct-YYYY-MM-DD.log (plain text, file only — never shown in the dashboard). It stays quiet when healthy (default level warn) and the file is created lazily, so a healthy run leaves nothing behind. When a quota fetch fails or a session is skipped, that is where the reason is recorded — bump logging.level to debug for the full detail.

Managing the file

# Print the config file path
vct config path

# Print the current settings
vct config show

# Open the file in $VISUAL / $EDITOR (falls back to vi / notepad)
vct config edit

# Print the JSON schema (regenerate with: vct config schema > vct.schema.json)
vct config schema

# Upgrade a legacy-format file to the current layout in place
vct config migrate

Smart Pricing System

How It Works

1. Automatic Updates: Fetches pricing from LiteLLM once per UTC day
2. Deterministic Matching: Finds the most specific model match even for versioned or provider-prefixed names
3. Failure Safety: A failed fetch cannot replace a good cache; vct keeps the previous map and backs off for five minutes before another attempt

Model Matching

Priority Order:

1. Exact Match: claude-sonnet-4 → claude-sonnet-4
2. Normalized: claude-sonnet-4-20250514 → claude-sonnet-4
3. Substring: custom-gpt-4 → gpt-4
4. Fuzzy: Uses Jaro-Winkler similarity (70% threshold)
5. Fallback: Shows $0.00 if no match found

Generic placeholder names (e.g. default, what cursor-agent records for auto-mode sessions) and very short names never take a substring/fuzzy match — unpriced is safer than a coincidental neighbor's price.

Cost Details

• Context tiers are per request: LiteLLM's "above Nk tokens" rates (e.g. GPT-5.x above 272k, Gemini above 200k) and its Qwen / doubao volume rows alike apply only to requests whose own prompt context crossed the boundary — never to a whole scan's token sum. Each request bills at whichever of the model's published rows its own context reached, so a model priced in four bands charges its top band for the requests that got there. Providers without per-request granularity — and offline scans — bill at the lowest rate, so tiered-model costs are a lower bound there.
• Beyond tokens: Claude web-search tool calls (server_tool_use.web_search_requests) are billed per query on top of the token cost; every other model's per-query charge is $0.
• OpenCode: a novel model name is priced from its tokens only on an exact LiteLLM match; with no exact match, vct trusts the assistant message's own stored cost instead of guessing from a loosely-similar name.
• Hermes: priced the same way as OpenCode — an exact LiteLLM match prices from tokens, otherwise vct uses Hermes's own stored cost.
• DeepSeek Harness: dsh records no cost of its own, so every model is priced from LiteLLM through the same lookup chain as the other file-based providers — there is no stored-cost fallback like OpenCode's or Hermes's. It routes to whatever the deployment configures, so expect deployment-defined model names, which land at $0 when nothing in LiteLLM matches.
• Grok: contextTokensUsed is priced as cache-read tokens only (falling back to the input rate when the model publishes no cache-read price); this is a point-in-time local context estimate, not cumulative billed usage.

Development

Contributor setup, building from source, tests, code quality, commit conventions, and release and packaging notes live in CONTRIBUTING.md.