OpenCode Agent Token Tracker

Local TanStack dashboard for tracking per-agent token consumption from the local OpenCode SQLite database. The core question it answers: which agent used how many tokens, on which model, and when?

Stack

• Node.js server, no Python runtime required
• React + Vite + TypeScript frontend
• TanStack Query for auto-refresh and manual refresh
• TanStack-oriented dashboard tables and charts
• System sqlite3 CLI in -readonly -json mode for large OpenCode databases

What It Tracks

Each view is organized around agent identity — the specific agent that performed work:

• Per-agent token totals — input/output/cache tokens broken down by agent, with model dimension
• Daily agent comparison — day-over-day token usage using message.time_created (not session.time_updated)
• Agent/model matrix — which agents ran on which models and their token footprint
• Task delegation graph — task() calls between agents, including skill loads (load_skills=[...]) and direct skill() invocations
• Top agent sessions — sessions ranked by token consumption with agent attribution
• Attribution recovery — restores agent/model identity for sessions where session.agent/session.model is empty, using assistant message JSON

Data Source

Default database path:

~/.local/share/opencode/opencode.db

The backend invokes sqlite3 with -readonly -json. It does not write to or modify the OpenCode database.

Run

npm install
npm run build
NODE_ENV=production npm start

The server binds to 0.0.0.0:12580 by default. Open locally:

http://127.0.0.1:12580

For development:

npm run dev        # API server on 0.0.0.0:12580
npm run dev:client # Vite frontend on 127.0.0.1:5173, proxies /api

Configuration

OPENCODE_DB=/path/to/opencode.db                      # optional, token usage DB
OPENCODE_CONFIG=/path/to/opencode.json                 # optional, user model cost overrides
OPENCODE_MODELS=/path/to/models.json                   # optional, built-in model catalog
HOST=0.0.0.0                                           # optional
PORT=12580                                             # optional

Cost / Pricing

Model costs are read from OpenCode's own config — no separate price table to maintain:

1. Built-in catalog (~/.cache/opencode/models.json by default) — OpenCode ships hundreds of models with cost: { input, output, cache_read, cache_write } (USD per 1M tokens).
2. User config (~/.config/opencode/opencode.json by default) — provider.<name>.models.<id>.cost overrides or adds to the built-in catalog.

To add or change a model's price, edit ~/.config/opencode/opencode.json and restart the service.

Currency: OpenCode's cost schema has no currency field (implicitly USD). CNY billing for horologium's glm-5-2 / qwen3-7-max is preserved via an internal map in server/model-costs.ts (CNY_MODELS). Add new CNY models there.

Legacy field: older configs may use cached_input instead of cache_read — both are accepted.

Missing cost: models without a cost block contribute token counts but no cost (not an error).

API

GET /api/health
GET /api/summary?days=7

days is clamped to 1..90.

Notes

• Requires sqlite3 with -json support. Verified with SQLite 3.53.3.
• Recent windows are calendar-day windows based on message timestamps.
• Missing session-level agent/model attribution is recovered from assistant message JSON when available.
• Skill token attribution is not exact: OpenCode records total message/session tokens, not token counts per injected skill fragment. The dashboard reports skill load counts and direct skill calls.