Usage

Usage tracking, budget percentages and model ranking for opencode: a CLI, and a /usage slash command for the TUI.

Every connected provider in one table, with each one's remaining budget as a comparable percentage — whether the provider bills in tokens, requests, credits or neurons.

Why

opencode talks to many providers at once, and each one meters differently: a weekly window here, premium requests there, credits, neurons, tokens per day. Nothing tells you which one you are about to exhaust.

This reads opencode's own SQLite and auth store — no configuration, no account — and normalizes every quota into one percentage, so the answer to "which provider still has room" is one glance.

Install

npm install -g @1930dev/opencode-usage   # the opencode-usage CLI

Requires Bun: the CLI reads opencode's SQLite through bun:sqlite.

Usage

# Local usage from opencode's SQLite (all providers, all projects)
opencode-usage usage                    # last 7 days, grouped by provider

# Window
opencode-usage usage --today            # today only
opencode-usage usage --since 30d        # last 30 days

# Grouping
opencode-usage usage --by agent         # by agent (build/plan/etc)
opencode-usage usage --by day           # by day
opencode-usage usage --by model         # by model
opencode-usage usage --by project       # by project directory
opencode-usage usage --by provider      # the default; lists every connected
                                        # provider, idle ones included

# Output
opencode-usage usage --json             # machine-readable
opencode-usage usage --pct              # add the % BUDGET column (--by provider only)

# Connected providers + live quota
opencode-usage providers
opencode-usage providers --no-net       # cached only

# What the providers themselves answer about their budget
opencode-usage probe                    # one free-model call per provider
opencode-usage probe --json

# Model ranking by intelligence per blended dollar
opencode-usage top
opencode-usage top --limit 10 --json

Plugin for opencode

The package also ships a /usage slash command for the opencode TUI. It shows the same table as the CLI, with the % budget column drawn as a progress bar:

 Usage — today                                                                      esc

 PROVIDER               MSGS   TOK IN  TOK OUT      COST BUDGET
 opencode-go               7        0        0     $0.00 ████████████ 100% weekly
 nvidia                   98    10.9M    19.9K     $0.00
 cloudflare-workers-ai     0        0        0     $0.00
 github-copilot            1     4.2K      830    $12.35 █████░░░░░░░  42% 1M tok/day

 TOTAL                   106                      $12.35

The dialog sizes itself to the terminal: a wide frame spells the budget labels out, a narrow one shortens them to the window (wk, mo, d).

opencode plugin @1930dev/opencode-usage

That installs the package and writes it into ~/.config/opencode/tui.json. To do it by hand:

{
  "$schema": "https://opencode.ai/tui.json",
  "plugin": ["@1930dev/opencode-usage"]
}

TUI plugins go in tui.json, not in opencode.json and not in ~/.config/opencode/plugins/. Both of those are loaded as server plugins, and opencode rejects a TUI-only module there with must default export an object with server().

To run it from a clone instead, build first and point tui.json at the bundle:

bun install && bun run build

{ "plugin": ["/absolute/path/to/opencode-usage/dist/tui.js"] }

Restart opencode after a rebuild: the bundle is read once at start.

% BUDGET (--pct)

Normalized percentage of budget consumed per provider, from these sources (in priority):

1. Live quota — provider-reported usage:

   • amd: daily spend ceiling (USD/day)
   • github-copilot: premium requests entitlement (7000/mo)
   • opencode-go (Zen): rolling 5h / weekly / monthly % (binding window)
   • openrouter: credits used / total credits
   • orcarouter: spend since top-up; a per-key credit cap turns into a %
   • zai: coding plan quota when the key has an active GLM plan; otherwise the wallet balance (fallback)
   • snowflake-cortex: tokens in the last 30d from the account usage views (SQL API); the percentage comes from a snowflake-cortex line in budgets.json

2. Documented limits — published quotas, used when the provider reports none. Each one is read against the period it resets on: a daily limit against today, a monthly one against the calendar month. The window you asked for (--since) sizes the table and never the budget.

   • cerebras: 1M tokens/day (free tier)
   • digitalocean: 5M tokens/day (paid)
   • google: 1500 requests/day (free tier)
   • groq: 200k tokens/day (free tier)

   These are documented but not measurable, and show —. opencode's database records tokens and requests; a credit and a neuron are the provider's own unit, derived from the model and the request by a rule it does not record. Give them a line in budgets.json to get a percentage in USD instead.

   • cloudflare-workers-ai: 10k neurons/day (free tier, resets 00:00 UTC)
   • nvidia: credit allowance retired — the trial is rate-limited per model
   • opencode: no spend API for API keys
   • zai: a key without a coding plan shows its wallet balance (live), and the percentage comes from a zai line in budgets.json

3. budgets.json — your monthly USD per provider, read against what the provider cost so far this calendar month:

   {
     "digitalocean": 5,
     "nvidia": 1
   }
   

   Place at ~/.config/opencode-usage/budgets.json.

Providers without any source show —.

Free-model probe (probe)

probe sends one minimal request against a free model per connected provider and reports what the answer reveals about the budget: a rate-limit header, a "no credits" error, or the meter itself (usage.neurons for Cloudflare). It reads the key from opencode's auth store and never prints it. Providers covered:

• cloudflare-workers-ai — the response carries neurons per request (the free tier's unit), so the probe answers whether the account can still run
• nvidia — an HTTP 200 means the free per-model allowance still works
• orcarouter — the free model answers rate_limit_error with a retry-after, and credits raise the cap
• zai — an HTTP 429 with code 1113 means the balance is empty

opencode exposes no chat API for keys and snowflake-cortex is read through its SQL quota, so both are skipped.

Ranking (top)

Models ranked by intelligence index per blended dollar:

MODEL                    NAME                IQ     CODING   $/M       IQ/$
glm-5.3-flash            GLM 5.3-Flash       58     72       $0.11     534.9
...

Intelligence indices from Artificial Analysis, pricing from models.dev.

top needs an AA_API_KEY for the indices. This package ships neither a key nor a copy of the data: the free Data API tier is "internal use only; no redistribution", and attribution is required on every tier — which is why the credit is printed under the table and carried in --json.

Requirements

• Bun >= 1.0 — the CLI reads opencode's SQLite through bun:sqlite
• opencode with existing sessions, at ~/.local/share/opencode/opencode.db
• For live quota: provider credentials already configured in opencode, at ~/.local/share/opencode/auth.json

Configuration

• AA_API_KEY — Artificial Analysis key, used by top for the intelligence indices. Optional: without it top still prints prices, only without IQ and IQ/$. A free key is at artificialanalysis.ai/data-api
• OPENCODE_AUTH_PATH — override the path to opencode's auth.json
• OPENCODE_DB_PATH — override the path to opencode's database
• OPENCODE_USAGE_BUDGETS — override the path to budgets.json
• OPENCODE_USAGE_CACHE — cache directory (default ~/.cache/opencode-usage)

Development

bun install          # install workspace deps
bun run build        # write dist/tui.js and dist/cli.js
bun run preview      # render the /usage dialog headless, at several widths
bun test             # run tests
bun run coverage     # run tests with the coverage gate
bunx tsc --noEmit    # typecheck

Every line of every source file is covered, and bunfig.toml fails the run if that stops being true. packages/cli/src/cli.ts holds no logic for that reason: the commands live in main.ts, which a test drives directly.

The workspace is a Bun monorepo. The packages are listed by dependency order, since cli and plugin both build on core:

• @opencode-usage/core — shared data layer (SQLite, quotas, matching, ranking)
• @opencode-usage/cli — the opencode-usage binary
• @opencode-usage/plugin — the opencode TUI plugin

Only the root package is published; core is inlined into both bundles.

License

MIT