Token Cost Tracker

Live tokens/sec, full-session token totals, and a per-model cost breakdown for the opencode terminal UI — including DeepSeek peak/off-peak pricing.

Zero config: add it to tui.json, restart opencode, done.

Contents

• Preview
• Features
• Requirements
• Install
• Usage
• How it works
• Limitations
• Troubleshooting
• Development
• License

Preview

Status line (bottom of the chat pane, centered, never overlapping the sidebar):

cache 12.34k  ·  in 45.67k  ·  out 8.90k  ·  reason 1.20k  ·  avg 42.3 t/s  ● off-peak

Cost panel (right sidebar, collapsible):

▾ Cost                        $0.0154
  ▸ deepseek-flash             $0.0070
    In $0.15 · Out $0.60 · C $0.003
    cache         12.3k    $0.0000
    input         45.7k    $0.0069
    output         0.9k    $0.0011

  ▸ deepseek-flash (peak)      $0.0210
    In $0.30 · Out $1.20 · C $0.006
    cache          4.1k    $0.0000
    input         30.0k    $0.0090
    output         6.0k    $0.0144

  ▸ Subagents                  $0.0031
    in 9.4k · out 3.1k · c 6.2k

Expanding Subagents lists each child session with its own in/out/c (reasoning folded into out):

  ▾ Subagents                  $0.0031
    in 9.4k · out 3.1k · c 6.2k
    explore: find the…         $0.0022
      in 5.0k · out 2.0k · c 4.0k
    general: run tests…        $0.0009
      in 4.4k · out 1.1k · c 2.2k

  Total                        $0.0185

Features

• Status line — session-wide token totals (cache reads, input, output, reasoning), token-weighted average speed (t/s), and a blinking DeepSeek peak/off-peak indicator.
• Cost panel — per-model totals, the catalog rates (In / Out / C), and per-category tokens with estimated cost. Click a model name to expand output and reason separately; collapsed, they are combined.
• Full-session accuracy — totals are read from the whole session through the SDK, not just the recent window the TUI keeps in memory, so counts do not shrink when you reopen a session.
• DeepSeek peak/off-peak pricing — peak usage is shown as its own ‹model› (peak) entry priced at 2× the listed off-peak rates (peak = 01:00–04:00 and 06:00–10:00 UTC, Mon–Fri).
• Subagents — a collapsible Subagents section shows the cumulative in/out/c token usage and cost of every subagent (child) session, and expands to a per-subagent breakdown (subagent_type: description), each with its own in/out/c. Reasoning is folded into out.
• Session total — a Total line adds the active session's models plus all subagents, so you can see the real session spend at a glance.
• Zero config, no build step — opencode's runtime transpiles the TSX and provides solid-js / @opentui.

Requirements

• opencode >= 1.15.0
• Terminal wider than 120 columns (or a manually toggled sidebar) to see the Cost panel.

Install

From npm (recommended)

Add the package to your global tui.json (~/.config/opencode/tui.json):

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["opencode-token-cost-tracker"]
}

Restart opencode. The package is installed automatically on first run.

From a local file

Copy tps-monitor.tsx into ~/.config/opencode/tui/ and reference it by path:

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["./tui/tps-monitor.tsx"]
}

Usage

• The status line is centered within the chat column so it never runs under the sidebar.
• The Cost panel renders in the right sidebar. Click the Cost header to collapse or expand the whole panel.
• Click a model name to toggle its output/reason breakdown.
• The Subagents section appears when a session has spawned subagents — it shows their cumulative cost and expands to a per-subagent breakdown.
• Prices come from models.dev — the same catalog opencode uses.

State is kept in opencode's KV store:

Key	Default	Meaning
token-cost-tracker.cost_expanded	true	Cost panel expanded
token-cost-tracker.model.<provider>/<model>	false	Per-model expanded (add #peak for peak entries)
token-cost-tracker.subagents_expanded	false	Subagents section expanded

How it works

• Totals are computed from session.messages (no limit) via the SDK and cached per session. The live TUI store is merged on top so streaming stays real-time.
• Average t/s = Σ(output + reasoning) ÷ Σ(turn span), where a turn's span runs from its first streamed text/reasoning token to completion (the in-progress turn uses "now").
• Cost = tokens × catalog rate ÷ 1,000,000. Reasoning is billed at the output rate when no dedicated rate exists. DeepSeek peak turns are priced at 2×.
• Subagent (child) sessions are discovered from the parent's task tool parts, so their tokens and cost are attributed to the subagent that produced them, recursively.
• No build step: opencode's Bun runtime transpiles the TSX and injects solid-js / @opentui.

Limitations

• Costs are estimates and can differ from your provider invoice (tiered >200k context pricing, cache-write pricing, promos, rounding).
• Peak/off-peak is decided from each turn's start time, so a turn that crosses a boundary is classified by its start.
• Cache write tokens are not included; most providers bill those separately.
• Subagent usage refreshes with the history poll (every 30 s), so it can lag the live parent numbers.

Troubleshooting

• Cost panel is missing — widen the terminal past 120 columns or toggle the sidebar; the panel renders in the sidebar slot.
• Plugin does not load — make sure tui.json is valid JSON and the name matches the npm package, then restart opencode.
• Numbers look stale — history refreshes every 30 seconds; the live window updates immediately.

Development

git clone https://github.com/sulaimanh7877/opencode-token-cost-tracker

The plugin entry point is tps-monitor.tsx (exported as ./tui). There is no build step.

License

MIT © Sulaiman