Providers

providers/get-provider-usage dispatches every selected provider independently and returns one normalized JSON object per provider. Most providers also expose a normalized providers/get-<id>-usage entrypoint, and every provider is covered by providers/get-provider-health. Claude's specialized helper emits KEY=VALUE data for the dispatcher to normalize; pi uses an inline dispatcher envelope plus providers/get-pi-analytics, and hermes computes its dispatcher envelope straight from ~/.hermes/state.db (indexed, milliseconds) with providers/get-hermes-analytics feeding the expanded telemetry card.

This document is the authoritative reference for adapter authors. It records, for every provider:

• the coverage level the plugin can truthfully report,
• the credential / local source it reads,
• the read-only key-validation endpoint (zero token consumption),
• the quota / balance API if one exists,
• the subscription plans and billing model,
• the dashboard URL where usage lives,
• the current flagship model(s) and recent changelog highlights,
• and the official documentation source used.

Reviewed 2026-08-26. Provider APIs change; re-check the linked sources before changing an adapter.

Coverage levels

Every provider maps to exactly one coverage level. The level dictates what the widget can truthfully render.

Level	Meaning	Example providers
Quota	Real usedPercent + reset window from a protocol/API.	codex, copilot, antigravity, openrouter, zai, glm, fireworks (with account ID), commandcode, opencode (Zen mode)
Balance	Remaining prepaid balance / credits in real currency.	kimi, deepseek
Analytics	Consumption counters (requests/tokens/neurons/cost) with no remaining-quota value.	cloudflare (GraphQL), 9router, claude (local), pi (local), hermes (local), opencode (local, default), codex (local, alongside its quota)
Auth / configured	Validates credentials with a read-only endpoint when possible; otherwise reports only that a credential is configured and states the limitation. No usage numbers.	gemini, mistral, nvidia, qwen, byteplus, groq, cohere, replicate, together, minimax, xai, kilo, ai21
Local runtime	Local process / installed models.	ollama, vertexai (gcloud)
Informational	No public read-only API at all; the card just links to the dashboard.	perplexity, cursor, cline, kiro, warp, amp

Provider kinds

Since 1.11.0 every provider also carries one or more kinds (independent from the coverage level above). The kind drives the card's badge, icon, and accent color:

Kind	Providers	Meaning
provider	everything else (default)	Standard API/account provider.
agent	pi	Local coding-agent harness; telemetry only, no account quota.
agent,provider	hermes	Dual nature: agent harness and provider front; the card renders as "Agent · Provider".
gateway	9router	Router that aggregates other providers behind one endpoint.
local	ollama	Local runtime, no cloud account.

Coverage level (what data can be truthfully reported) and kind (what the thing is) are orthogonal: for example hermes is Analytics coverage with kinds agent,provider.

Coverage matrix

The matrix below summarises the authentication/billing surface for every supported provider. Legend: ✅ yes · ❌ no public API · ⚠️ partial / dashboard-only.

Provider	Coverage	Read-only key check	Quota / balance API	Sub. plan	PAYG	Env var	Dashboard	Docs source
codex	Quota	codex app-server	✅ duration-labelled session / weekly windows, including weekly-only responses	✅ Plus / Pro / Team	✅	codex login	developers.openai.com	Codex app-server
claude	Quota + local analytics	local ~/.claude (or $CLAUDE_CONFIG_DIR)	✅ 5h / 7d / weekly per-model (limits[])	✅ Pro / Max	✅	claude OAuth, optional CLAUDE_CONFIG_DIR	~/.claude	local analytics
copilot	Quota snapshot	copilot_internal/user	✅ premium (AI credits since 2026-06) + overage; chat/completions shown only when not unlimited	✅ Free / Pro / Pro+ / Education / Business / Enterprise	—	COPILOT_GITHUB_TOKEN, GH_TOKEN, or GITHUB_TOKEN (the gh CLI token is tried first)	github.com/settings/copilot	GitHub Copilot
antigravity	Quota (Cloud Code Assist)	✅ local Antigravity OAuth (keyring / IDE session, auto-refreshed)	✅ Gemini / Claude & OpenAI / unknown family quota + reset; optional per-model detail; per-account failures	✅ Antigravity plan	—	~/.config/Antigravity IDE	Antigravity IDE	loadCodeAssist + v1internal:fetchAvailableModels on cloudcode-pa.googleapis.com (multi-account)
gemini	Auth	✅ GET /v1beta/models	❌ dashboard-only	⚠️ tiered PAYG (no flat sub)	✅ prepaid credits → tiers	GEMINI_API_KEY	aistudio.google.com	ai.google.dev
cloudflare	Analytics	✅ GET /user/tokens/verify	⚠️ GraphQL analytics (not bill)	✅ Workers Paid \$5/mo	✅ \$0.011 / 1k neurons	CLOUDFLARE_API_TOKEN	dash.cloudflare.com	developers.cloudflare.com
mistral	Auth	✅ GET /v1/models	❌ dashboard-only	✅ Vibe Pro \$14.99 / Team \$24.99	✅ per token	MISTRAL_API_KEY	console.mistral.ai	docs.mistral.ai
glm / zai	Quota	✅ GET /api/monitor/usage/quota/limit	✅ per-window % + reset timestamp	✅ GLM Coding Plan \$18–\$160/mo	✅ per token	ZAI_API_KEY / GLM_API_KEY (ZHIPU_API_KEY also accepted; precedence depends on the adapter — see GLM / Z.ai)	z.ai/manage-apikey	docs.z.ai / open.bigmodel.cn
nvidia	Auth	⚠️ GET /v1/models is public and cannot validate a key	❌ dashboard-only	⚠️ rate-limited developer trial	✅ NVIDIA Cloud Credits	NVIDIA_API_KEY	build.nvidia.com/credits	docs.api.nvidia.com
minimax	Auth	✅ GET /v1/models	❌ dashboard-only	✅ Token Plan \$20/\$50/\$120/mo	✅ per token (M3 50% off)	MINIMAX_API_KEY	platform.minimax.io	platform.minimax.io/docs
commandcode	Quota	✅ GET /provider/v1/models	✅ /alpha/billing/credits (5h + weekly)	✅ Go / GOAT / Pro / Max / Team Pro / Provider API	✅ USD credit balance + per-window USD usage	COMMAND_CODE_API_KEY or CLI ~/.commandcode/auth.json	commandcode.ai/billing	Provider API docs
opencode	Quota	✅ GET /zen/go/v1/models	✅ /zen/go/v1/usage (5h + weekly + monthly)	✅ OpenCode Go \$10/mo	—	OPENCODE_API_KEY or CLI ~/.local/share/opencode/auth.json	opencode.ai/zen	opencode.ai/docs/go
kimi	Balance / Quota	✅ GET /v1/models	✅ /v1/users/me/balance (funds) · /coding/v1/usages (Kimi Code weekly + 5h)	✅ Kimi Code (tempo tiers)	✅ per token (USD/CNY)	MOONSHOT_API_KEY, KIMI_API_KEY, or KIMI_CODING_API_KEY	platform.kimi.ai	platform.kimi.ai/docs
qwen	Auth	⚠️ GET /compatible-mode/v1/models	❌ Alibaba billing console	✅ Coding Plan Pro \$50/mo	✅ per token	DASHSCOPE_API_KEY	modelstudio.console.alibabacloud.com	alibabacloud.com/help
xai	Auth	✅ GET /v1/api-key	❌ dashboard-only (usage.cost_in_usd_ticks per request)	⚠️ prepaid credits only	✅ per token	XAI_API_KEY	console.x.ai/billing	docs.x.ai
kilo	Auth	⚠️ GET /api/gateway/models (no-auth)	❌ dashboard-only (402 signal)	✅ Kilo Pass \$19/\$49/\$199/mo	✅ at-provider-cost credits	KILO_API_KEY	app.kilo.ai	kilo.ai/docs
kiro	Informational	❌ no public API	❌ no public API	✅ Free / Pro / Pro+ / Pro Max / Power	⚠️ overage \$0.04/credit	—	app.kiro.dev	kiro.dev/docs
9router	Local analytics	local SQLite/JSON	✅ local requests/tokens/cost	—	—	~/.9router	—	local store
pi	Local analytics	local JSONL session logs	✅ local cost/tokens — no quota API (pi has no rate limits)	—	—	—	—	local store
hermes	Local analytics	local SQLite state (~/.hermes/state.db)	✅ local sessions/tokens/cost per model and project — the agent side; provider billing (Nous Portal / routed providers) stays on the portal	—	—	—	—	NousResearch/hermes-agent
openrouter	Quota/balance	✅ GET /api/v1/key	✅ limit/daily/monthly/balance	—	✅ per token	OPENROUTER_API_KEY	openrouter.ai/credits	openrouter.ai/docs
deepseek	Balance	balance endpoint	✅ account + granted balance	—	✅ per token	DEEPSEEK_API_KEY	platform.deepseek.com	api-docs.deepseek.com
ollama	Local runtime	GET /api/tags	✅ installed + running models	—	—	OLLAMA_HOST	localhost:11434	docs.ollama.com
vertexai	Local runtime	gcloud auth	⚠️ GetUsage exists with separate IAM Signature V4 credentials	—	✅	gcloud	console.cloud.google.com	cloud.google.com/vertex-ai
byteplus	Auth	✅ GET /api/v3/models	❌ dashboard-only	—	✅	BYTEPLUS_API_KEY	console.byteplus.com	docs.byteplus.com
together	Auth	✅ GET /v1/models	❌ no documented read-only credits endpoint	—	✅	TOGETHER_API_KEY	api.together.ai	docs.together.ai
groq	Auth	✅ GET /v1/models	❌ dashboard-only	—	✅	GROQ_API_KEY	console.groq.com	console.groq.com/docs
cohere	Auth	✅ models API	❌ dashboard-only	—	✅	COHERE_API_KEY	dashboard.cohere.com	docs.cohere.com
replicate	Auth	✅ account endpoint	❌ dashboard-only	—	✅ per second	REPLICATE_API_TOKEN	replicate.com/account	replicate.com/docs
fireworks	Quota (with account ID)	✅ inference models API	✅ GET /v1/accounts/{account_id}/quotas	—	✅	FIREWORKS_API_KEY; optional FIREWORKS_ACCOUNT_ID	app.fireworks.ai	docs.fireworks.ai
ai21	Auth	✅ GET /studio/v1/models	❌ dashboard-only	—	✅	AI21_API_KEY	studio.ai21.com	docs.ai21.com
perplexity	Informational	❌	❌	✅ Pro	✅	—	perplexity.ai/settings	docs.perplexity.ai
cursor	Informational	❌	❌	✅ Hobby / Pro / Business	—	—	cursor.com/settings	cursor.com
cline	Informational	❌	❌	—	✅	—	app.cline.bot	cline.bot
warp	Informational	❌	❌	✅ Warp Pro	—	—	app.warp.dev	warp.dev
amp	Informational	❌	❌	✅ Amp Pro	—	—	ampcode.com	ampcode.com

Local fallbacks and aliases

OpenRouter → 9Router local fallback. When OPENROUTER_API_KEY is unset, the dispatcher serves the card from the local 9Router store instead (fetch_9router_native openrouter openrouter soft); the account label reads "openrouter via 9router (local)" so locally routed data is never mistaken for upstream OpenRouter API quota. With the variable set, the real /api/v1/key snapshot is used.

Aliases. The dispatcher (get-provider-usage) and health checker (get-provider-health) accept these case-insensitive aliases:

Canonical	Aliases
commandcode	cmd, cmdcode
antigravity	agy
kimi	moonshot
glm	zhipu
zai	z.ai
nvidia	nim
vertexai	vertex
byteplus	ark, modelark
qwen	dashscope, alibaba
xai	grok

Provider reference

Detailed adapter notes for the focus providers (Gemini, Cloudflare, Mistral, GLM/Z.ai, NVIDIA, MiniMax, Kimi, Qwen, xAI, Kilo, Kiro). All HTTP probes below are read-only and consume no tokens.

Gemini (Google)

API base	https://generativelanguage.googleapis.com/v1beta (OpenAI-compat: /v1beta/openai/). No China mirror; enterprise path is Vertex AI.
Env var	GEMINI_API_KEY (fallbacks GOOGLE_API_KEY, GOOGLE_GENERATIVE_AI_API_KEY); also detected via local Gemini CLI (~/.gemini/oauth_creds.json).
Auth	Header x-goog-api-key: <key> (preferred) or query ?key=<key>.
Key check	GET /v1beta/models → 200 lists models; 403 missing key; 400 malformed. Zero tokens.
Quota / balance	❌ None. Usage, spend, and rate limits are dashboard-only.
Plans	Free + tiered PAYG (Tier 1 $250 cap → Tier 3 $20k+ cap, auto-upgrade by spend/age). No flat monthly API subscription. Enterprise = Gemini Enterprise Agent Platform.
Billing	Prepaid credits → pay-as-you-go per token. Batch API 50% off; cached/prompt discounts.
Flagship pricing (/M tok)	gemini-3.5-flash $1.50/$9.00 · gemini-3.1-pro-preview $2.00/$12.00 (≤200k) · gemini-3.1-flash-lite $0.25/$1.50 · gemini-2.5-pro $1.25/$10.00.
Dashboard	aistudio.google.com (rate limits), console.cloud.google.com/billing.
Changelog	Gemini 3.5 Flash GA; Gemini 3.1 Pro preview (vibe-coding/customtools). Gemini 2.0 Flash/Flash-Lite shut down 2026-06-01. Imagen 4 deprecated (→ Gemini 2.5 Flash Image). Veo 3/2 → Veo 3.1. Gemini Embedding 2 (multimodal).
Adapter	fetch_gemini_native — API-key probe; falls back to local Gemini CLI OAuth detection.

Cloudflare (Workers AI)

API base	https://api.cloudflare.com/client/v4 (account-scoped AI: /accounts/{account_id}/ai/...; OpenAI-compat: /accounts/{account_id}/ai/v1).
Env var	CLOUDFLARE_AI_TOKEN (preferred) or CLOUDFLARE_API_TOKEN; CLOUDFLARE_ACCOUNT_ID unlocks GraphQL analytics.
Auth	Authorization: Bearer <token>. Token needs Workers AI - Read/Edit.
Key check	GET /user/tokens/verify → 200 {"success":true,"result":{"status":"active"}}. Zero neurons.
Quota / balance	⚠️ GraphQL analytics only (not the bill). POST /graphql with aiInferenceAdaptiveGroups dataset scoped by accountTag. Returns sum { requests neurons } and dimensions { date modelName taskType }. The dedicated tutorial page was removed in 2025 — verify node/fields via introspection before relying on it. No REST "remaining neurons" endpoint.
Plans	Workers Free: 10,000 neurons/day. Workers Paid ($5/mo): same free allotment then $0.011/1k neurons. Enterprise: custom.
Billing	Per neuron ($0.011/1k). Per-model rates also expressed as /M tokens (e.g. `@cf/zai-org/glm-5.2` \1.40/$4.40).
Dashboard	dash.cloudflare.com → account → Workers AI.
Changelog	2026-06-16 GLM-5.2; 2026-06-12 Kimi K2.7 Code; 2026-04-20 Kimi K2.6 (reasoning field, chat_template_kwargs.thinking); 2026-03-19 prompt caching (x-session-affinity); 2026-02-13 GLM-4.7-Flash; 2025-08-05 gpt-oss-120b/20b; 2024-05-17 OpenAI-compat added; unit-based neuron pricing since 2024-09. Major deprecation wave 2026-05-30 (Llama-2/3/3.1 base+AWQ, Mistral 7B, Gemma 7B/3-12B, phi-2).
Adapter	fetch_cloudflare_native — token verify; if CLOUDFLARE_ACCOUNT_ID is set, queries aiInferenceAdaptiveGroups for 7-day totals + latest day (graceful fallback to the token-verified note card).

Mistral AI

API base	https://api.mistral.ai/v1. No regional mirror.
Env var	MISTRAL_API_KEY.
Auth	Authorization: Bearer <key>.
Key check	GET /v1/models → 200; 401 on bad key. Zero tokens. (/v1/billing and /v1/usage return 404.)
Quota / balance	❌ None. Dashboard-only ("Mistral Studio dashboard offers detailed tracking of API usage").
Plans	Vibe consumer/coding subs: Free / Pro $14.99 / Team $24.99 / Enterprise / Education $5.99. API = PAYG per token. Batch 50% off; cached input 10% of input price.
Billing	Per token (per 1M): mistral-medium-3.5 $1.5/$7.5 (coding flagship) · mistral-large-3 $0.5/$1.5 · mistral-small-4 $0.1/$0.3 · devstral-2 $0.4/$2.0 (agentic coding) · codestral $0.3/$0.9 (FIM) · magistral-medium $2.0/$5.0.
Dashboard	console.mistral.ai (keys, usage, billing). Pricing: mistral.ai/pricing.
Changelog	2026-04 Mistral Medium 3.5; 2026-03 Mistral Small 4 (Apache 2.0); product restructure into Vibe / Studio / Admin (La Plateforme rebranded to Studio). New Workflows, Connectors, Observability. Fine-tuning + old Agents endpoints deprecated.
Adapter	fetch_mistral_native — /v1/models validation only.

GLM / Z.ai (Zhipu AI)

API base	Global https://api.z.ai/api/paas/v4; Coding-Plan https://api.z.ai/api/coding/paas/v4; China https://open.bigmodel.cn/api/paas/v4. Fully OpenAI-compatible.
Env var	Per adapter: fetch_zai_native (global card) uses ZAI_API_KEY → GLM_API_KEY → ZHIPU_API_KEY; fetch_glm_native (China console card) uses GLM_API_KEY → ZHIPU_API_KEY → ZAI_API_KEY. Any one key drives both cards.
Auth	Authorization: Bearer <key>.
Key check	GET /api/monitor/usage/quota/limit → 200 with success: true and data.limits[]; 401/403 on bad key. Zero tokens. Falls back to GET /paas/v4/models when quota endpoint is unavailable.
Quota / balance	✅ /api/monitor/usage/quota/limit returns data.limits[] — each limit has type (TIME_LIMIT or TOKENS_LIMIT), percentage (0–100), nextResetTime (epoch ms), remaining, unit, and number. Live cross-check against the Z.ai Usage page confirms the period unit table used by the adapter: unit=4 → 5-hour session window, unit=6 → weekly quota, unit=5 → monthly web/search/reader quota, unit=3 → total token allotment (no reset window). Also returns data.level (plan tier, e.g. lite).
Plans	PAYG per token or GLM Coding Plan (Lite $18, Pro $72, Max $160/mo; quarterly/annual discounts). Subscription usage exposes a 5-hour session quota, weekly token quota, and monthly Web Search / Reader / Zread quota; supported in Claude Code, Cline, OpenCode, Roo, Kilo, Crush, Goose, OpenClaw. GLM-5.2 & GLM-5-Turbo consume 3× quota at peak (14:00–18:00 UTC+8), 2× off-peak (1× off-peak promo through end of September).
Billing	Per 1M tokens: glm-5.2 $1.40/$4.40 (cached $0.26) · glm-5/glm-5-turbo $1.0–$1.2/$3.2–$4.0 · glm-4.7/4.6/4.5 $0.60/$2.20 · glm-4.7-flash & glm-4.5-flash Free · glm-5v-turbo (vision) $1.2/$4.0. Web Search $0.01/use.
Dashboard	API keys, subscription (Coding Plan), and billing (finance). China: open.bigmodel.cn.
Changelog	GLM-5.2 live (1M lossless context, 128K max output, MCP, structured output). Lineage: 4.5 → 4.6 → 4.7 → 5 → 5-Turbo → 5.1 → 5.2. New GLM-5V-Turbo (vision coding), GLM-Image, CogVideoX-3, GLM-ASR-2512, GLM-OCR. Coding Plan restructured (legacy plans migrated by 2026-04-30).
Adapter	fetch_glm_native (China console, open.bigmodel.cn) and fetch_zai_native (global, api.z.ai). Both call GET /api/monitor/usage/quota/limit for real quota data; fall back to GET /models auth-only check if the quota endpoint is unavailable. Sorts limits by urgency (highest % first among timed windows), then maps to primary / secondary / tertiary.

NVIDIA (NIM / build.nvidia.com)

API base	https://integrate.api.nvidia.com/v1 (OpenAI-compatible). No regional mirror; backed by DGX Cloud.
Env var	NVIDIA_API_KEY (NGC personal key, nvapi- prefix).
Auth	Authorization: Bearer <key>.
Key check	GET /v1/models is a public catalog and cannot validate a key; the card truthfully reports only that a key is set.
Quota / balance	❌ None documented. NVCF/Cloud Functions billing doc is auth-gated (401). Balance is dashboard-only. Practical exhaustion signal: inference returns 402/429.
Plans	Rate-limited Developer trial, NVIDIA Cloud Credits (USD blocks), and NVIDIA AI Enterprise for self-hosted NIM.
Billing	Credit-based per request (1 credit ≈ 1 standard request; heavy models cost more). No posted $/M-token table for the hosted catalog.
Dashboard	build.nvidia.com/credits. NGC keys at org.ngc.nvidia.com.
Changelog	Nemotron-3 family (Ultra-550B, Super-120B-A12B, Nano-30B-A3B, Nano-Omni-30B). Heavy expansion into third-party frontier models (DeepSeek V4 Pro/Flash, Mistral Large 3 675B + Medium 3.5 + Small 4, MiniMax M2.7/M3, Stepfun, Z.ai GLM 5.1/4.7, Qwen3-Coder-480B). New async 202+polling pattern for heavy endpoints. Healthcare microservices (AlphaFold2, Boltz2, OpenFold3).
Adapter	fetch_nvidia_native — public /v1/models catalogue probe; it does not claim that the key is validated.

MiniMax

API base	https://api.minimax.io/v1 (OpenAI-compat); Anthropic-compat https://api.minimax.io/anthropic. Legacy TTS host api.minimax.chat. No .cn host.
Env var	MINIMAX_API_KEY. Two key types: pay-as-you-go API Key vs Token-Plan Subscription Key (not interchangeable).
Auth	Authorization: Bearer <key>.
Key check	GET /v1/models → 200 lists MiniMax-M3, MiniMax-M2.7, MiniMax-M2.5… Zero tokens.
Quota / balance	❌ None. Dashboard-only. Token-Plan usage shown as a console bar (5-hour rolling + weekly). Errors: 1004 auth failed, 1008 insufficient balance, 1002 rate limit, 1039 token limit exceeded.
Plans	Token Plan (replaces old "Coding Plan"): Plus $20 / Max $50 / Ultra $120/mo — full-spectrum multimodal, 5h+weekly windows, no rollover. Credits: $5/$25/$100 packs (1000cr = $1, 365-day). PAYG also available.
Billing	Per 1M tokens: MiniMax-M3 (≤512K, 50% off) $0.30/$1.20 (cache $0.06) · >512K $0.60/$2.40 · Priority tier 1.5× · MiniMax-M2.7 $0.30/$1.20 · MiniMax-M2.7-highspeed $0.60/$2.40. Audio speech-2.8-hd $100/M chars; Hailuo video $0.19–$0.56/clip.
Dashboard	platform.minimax.io: keys /user-center/basic-information/interface-key, balance /user-center/payment/balance, Token Plan /user-center/payment/token-plan.
Changelog	2026-06-01 MiniMax-M3 (1M ctx, adaptive thinking, coding SOTA). 2026-03-18 M2.7/M2.7-highspeed. 2026-02 M2.5. 2025-12-22 M2.1. 2025-10-27 M2 + Hailuo-2.3. Token Plan replaced Coding Plan (broader coverage, separate Subscription Key).
Adapter	fetch_minimax_native — /v1/models validation.

Command Code

API base	https://api.commandcode.ai/provider/v1 (OpenAI/Anthropic-compatible); quota https://api.commandcode.ai/alpha/billing/credits (Bearer, no cookies).
Credentials	COMMAND_CODE_API_KEY (preferred when available to DMS), or the apiKey in CLI-owned ~/.commandcode/auth.json after cmd login. Both are the same Provider API key used for /provider/v1/models, /alpha/whoami, /alpha/billing/credits.
Auth	Authorization: Bearer ***
Quota / balance	/alpha/billing/credits returns windowLimits.fiveHour (5h USD used/cap, Unix-ms resetAt), windowLimits.weekly (7-day USD used/cap, Unix-ms resetAt), and credits.monthlyCredits (remaining USD this billing cycle). /alpha/billing/subscriptions exposes planId (e.g. individual-goat) and currentPeriodEnd.
Plans	Go $1 / $10 credit · GOAT $10 / $70 credit, 5h $14 / weekly $35 · Pro $20 / $80 credit, 5h $16 / weekly $40 · Max 10× $100 / $150 credit, 5h $45 / weekly $90 · Max 20× $200 / $300 credit, 5h $90 / weekly $180 · Team Pro $40 / $40 credit, 5h $12 / weekly $24 · Provider API $15 + pay-as-you-go.
Billing	Per-token, no markup on the Provider plan. Credits roll over and never expire. Window caps are USD ceilings — model deals (e.g. minimax-m3 2×, google/gemini-3.7-flash 50% off, mimo-v2.5-pro 99% off) apply automatically.
Stability	The /alpha/ namespace is experimental — no documented versioned contract, can change without notice. Adapter degrades to the documented /provider/v1/models endpoint on alpha failure and emits a clearly-labeled "quota endpoint unavailable" note; it never fabricates a percentage.
Dashboard	commandcode.ai/billing — billing, plan, and credit top-ups.
Adapter	fetch_commandcode_native — /alpha/billing/credits + /alpha/whoami + /alpha/billing/subscriptions, with documented-endpoint fallback.

OpenCode

Two independent modes. The local mode is the default whenever an OpenCode database exists, because OpenCode is commonly driven against third-party and free-tier providers with no Zen subscription — the Zen quota endpoint would only report an entitlement error there. Set OPENCODE_USAGE_SOURCE=api to force the Zen path.

Local source	${OPENCODE_DATA_DIR:-${XDG_DATA_HOME:-$HOME/.local/share}/opencode}/opencode.db (SQLite, read-only). Assistant messages carry providerID, modelID, tokens and cost; only usage metadata is read, never message content. Costs come from whatever OpenCode recorded — a missing cost makes the window unknown (—) rather than $0.
Local adapter	fetch_opencode_native (local branch) + providers/get-local-analytics opencode, cached 120s.

OpenCode Go (Zen quota mode)

API base	https://opencode.ai/zen/go/v1 — OpenCode Zen's Go-plan quota surface, alongside the OpenAI-compatible /zen/v1 inference API.
Credentials	OPENCODE_API_KEY (preferred when available to DMS), or the key in CLI-owned ${XDG_DATA_HOME:-$HOME/.local/share}/opencode/auth.json (saved by opencode auth login). The adapter follows this XDG path on Linux, the plugin's supported platform. The same key created in OpenCode Studio/Zen drives both inference and quota introspection.
Auth	Authorization: Bearer ***
Quota / balance	/zen/go/v1/usage returns rollingUsage (5h window), weeklyUsage, and monthlyUsage, each {status, resetInSec, usagePercent} — percentages are computed server-side, so the adapter never derives them from raw byte/token counts. The adapter accepts a quota response only when all three windows have valid percentage/reset values; otherwise it falls back to auth-only status. resetInSec is seconds-from-now, converted to an absolute ISO 8601 timestamp. useBalance: true is rendered as balance fallback enabled; the endpoint does not return the balance amount, so none is invented.
Plans	OpenCode Go $10/mo — $12 of usage per 5-hour window, $30 weekly, $60 monthly.
Billing	Flat monthly subscription; no PAYG surfaced through this endpoint.
Stability	The /zen/go/ namespace shipped alongside the Go plan (opencode PR #16513) and has no documented versioned contract. Adapter degrades to /zen/go/v1/models (auth-only) on failure and emits a clearly-labeled "quota endpoint unavailable" note; it never fabricates a percentage. A missing/invalid key returns 401 AuthError; a valid key without a Go subscription returns 403 EntitlementError — both surface as hard errors, not soft notes.
Dashboard	opencode.ai/zen — Zen usage and Go plan management.
Adapter	fetch_opencode_native (API branch) — /zen/go/v1/usage, with /zen/go/v1/models fallback.

Kimi (Moonshot AI)

Two systems	Kimi exposes two independent quota surfaces with non-interchangeable keys. ① Open Platform (prepaid balance): sk-xxx key on api.moonshot.ai/.cn. ② Kimi Code / Coding Plan (subscription quota): sk-kimi-xxx key on api.kimi.com/coding/v1 — the same data the Kimi CLI /usage command reads. The adapter routes by key (see Env var / Adapter).
API base	Open Platform — Global https://api.moonshot.ai/v1 (USD); China https://api.moonshot.cn/v1 (CNY). Coding Plan — https://api.kimi.com/coding/v1 (override with KIMI_BASE_URL). Platform rebranded: platform.moonshot.ai → platform.kimi.ai (global) / platform.moonshot.cn → platform.kimi.com; API hosts unchanged.
Env var	Balance: MOONSHOT_API_KEY (fallback KIMI_API_KEY); optional MOONSHOT_API_BASE overrides host probing. Coding Plan: KIMI_CODING_API_KEY, or a KIMI_API_KEY/MOONSHOT_API_KEY carrying the sk-kimi- prefix; optional KIMI_BASE_URL. An explicit coding key wins over the balance path.
Auth	Authorization: Bearer <key> (both systems). Coding Plan requests also send User-Agent: KimiCLI/1.6.
Key check	Balance: GET /v1/models → 200; 401 on bad key. Coding Plan: the /usages probe doubles as the key check (401/403 = wrong key type, 404 = wrong base URL).
Quota / balance	Balance: ✅ GET /v1/users/me/balance → {data:{available_balance, voucher_balance, cash_balance}} (USD on .ai, CNY on .cn; available_balance = cash + voucher). Coding Plan: ✅ GET /coding/v1/usages (older deployments answer on /usage) → weekly + 5-hour windows with used/limit/remaining and a reset time; mapped weekly → primary, 5h → secondary.
Plans	Open Platform: no subscription — PAYG + prepaid top-up vouchers; rate-limit tiers scale with cumulative recharge ($1 Tier0 → $3,000 Tier5). Kimi Code subscription tiers (named after musical tempos): Adagio $0 (no Kimi Code), Moderato $19, Allegretto $39, Allegro $99, Vivace $199/mo (annual ≈ −20%). Quota refreshes on a 7-day cycle plus a 5-hour burst limiter (≈300–1,200 requests/5h, up to 30 concurrent by tier); entry paid tier ≈ 2,048 Kimi Code requests/week. 403 "reached your usage limit for this billing cycle" = weekly quota exhausted.
Billing	Per 1M tokens (256K ctx): kimi-k2.7-code $0.95/$4.00 (cache hit $0.19) · kimi-k2.7-code-highspeed $1.90/$8.00 · kimi-k2.6 $0.95/$4.00 · kimi-k2.5 $0.60/$3.00. Automatic context caching.
Dashboard	Balance: platform.kimi.ai/console (global) / platform.kimi.com/console (China). Coding Plan: Kimi Code console + membership page; in-CLI /usage.
Changelog	kimi-k3 (max quality) and kimi-k2.7-code / -highspeed (routine coding, ~180 tok/s). kimi-k2.6 (multimodal flagship). kimi-k2.5 + moonshot-v1 retire 2026-08-31 → migrate to kimi-k2.7-code or kimi-k3. kimi-k2 series deprecated 2026-05-25; kimi-latest removed 2026-01-28; kimi-thinking-preview removed 2025-11-11. Full OpenAPI spec at platform.kimi.ai/docs/openapi.json.
Adapter	fetch_kimi_native routes by key: sk-kimi-/KIMI_CODING_API_KEY → fetch_kimi_code_native (Coding Plan GET /usages → weekly primary + 5h secondary, source: kimi-code); otherwise the balance API with global→China host probing (source: kimi-api, primary = available balance, secondary = voucher/cash split).

Qwen / DashScope (Alibaba Model Studio)

API base	5 regional modes: International (Singapore) https://dashscope-intl.aliyuncs.com/compatible-mode/v1; US (Virginia) https://dashscope-us.aliyuncs.com/compatible-mode/v1; China (Beijing) https://dashscope.aliyuncs.com/compatible-mode/v1; HK https://cn-hongkong.dashscope.aliyuncs.com/compatible-mode/v1; EU (Frankfurt) https://{workspace}.eu-central-1.maas.aliyuncs.com/compatible-mode/v1. Coding Plan uses https://coding-intl.dashscope.aliyuncs.com/v1 with a dedicated sk-sp-… key.
Env var	DASHSCOPE_API_KEY (fallback QWEN_API_KEY); optional DASHSCOPE_WORKSPACE_ID header. Coding Plan uses sk-sp-….
Auth	Authorization: Bearer <key> + optional X-DashScope-WorkSpace.
Key check	⚠️ GET /compatible-mode/v1/models is used by the adapter but is not officially documented in the OpenAI-compat surface — treat as best-effort. 401 confirms a bad key.
Quota / balance	❌ None. Billing flows through Alibaba Cloud User Center (per-minute inference, per-hour batch; monthly settle).
Plans	Coding Plan Pro $50/mo: 6,000 req/5h, 45,000/week, 90,000/month — qwen3.5-plus, kimi-k2.5, glm-5, MiniMax-M2.5, qwen3-coder-next/+. Lite plan closed to new subs since 2026-03-20. Free quota (1M in+out tokens × 90 days) on International region only.
Billing	Per 1M tokens (intl, USD): qwen3-max $1.20/$6.00 (≤32K; doubles >128K) · qwen3.5-plus $0.40/$2.40 (1M ctx) · qwen3.5-flash $0.10/$0.40 · qwq-plus $0.80/$2.40. Batch 50% off; Context Cache (input-only). CNY on China-mainland.
Dashboard	modelstudio.console.alibabacloud.com (intl); bailian.console.alibabacloud.com (China). Billing: Alibaba Cloud User Center.
Changelog	2026-02 qwen3.5-plus-2026-02-15, qwen3.5-flash-2026-02-23; 2026-01-23 qwen3-max-2026-01-23 (web-search/code-interpreter in thinking mode). New open-source qwen3.5-{397b-a17b,120b-a10b,27b,35b-a3b}. Qwen-Turbo deprecated (→ Qwen-Flash). Coding Plan Lite closed to new subs.
Adapter	fetch_qwen_native — /compatible-mode/v1/models validation.

xAI (Grok)

API base	https://api.x.ai/v1 (inference, OpenAI-compat). Key management: https://management-api.x.ai (separate, June 2025).
Env var	XAI_API_KEY.
Auth	Authorization: Bearer <key>.
Key check	✅ GET /v1/api-key → 200 returns key metadata: {redacted_api_key, name, user_id, team_id, api_key_id, api_key_blocked, api_key_disabled, team_blocked, acls, create_time, modify_time}. 401 without key. This is the best validator — it also surfaces blocked/disabled state. (GET /v1/models also works and returns per-token prices.)
Quota / balance	❌ No remaining-credits field on /v1/api-key. Account balance is dashboard-only. Per-request cost is in every response's usage.cost_in_usd_ticks (1 USD = 1e10 ticks) and usage.cost_in_nano_usd (Responses API); enable via stream_options.include_usage:true.
Plans	API = prepaid credits (no named tiers). service_tier:"default" vs "priority" (2× billing, higher priority — June 2026). Consumer Grok subs (Free/SuperGrok/SuperGrok Heavy) are separate from api.x.ai.
Billing	Per 1M tokens: grok-4.3 (flagship, 1M ctx) $1.25/$2.50 · grok-4.20-0309-{reasoning,non-reasoning,multi-agent} $1.25/$2.50 · grok-build-0.1 (coding, aliases grok-code-fast-1, grok-code-fast, 256K ctx) $1.00/$2.00 (cache $0.20). Batch 20–50% off. Tools billed per 1k calls (web/x search $5, code exec $5, attachment $10). Image $0.02–$0.05/img; video $0.05–$0.08/sec.
Dashboard	console.x.ai: keys /team/default/api-keys, billing /billing, models /team/default/models. Status: status.x.ai.
Changelog	2026-06 Priority Processing + Files public URLs. 2026-05 grok-build-0.1 coding model + Grok Build CLI + Context Compaction API + WebSocket Responses. 2026-04 cost_in_usd_ticks on every response. 2026-03 Grok 4.20 + Multi-agent. 2025-12 Voice Agent API GA. 2025-11 Grok 4.1 Fast + Files API GA + Remote MCP. 2025-10 agentic server-side tools GA. 2025-09 Responses API GA. 2025-07 Grok 4. Anthropic-compat /v1/messages & /v1/complete deprecated; max_tokens → max_completion_tokens.
Adapter	fetch_xai_native — GET /v1/api-key; surfaces key name + blocked/disabled flags in the note.

Kilo (kilo.ai)

API base	Gateway https://api.kilo.ai/api/gateway (OpenAI-compat; model string is provider/model, e.g. anthropic/claude-sonnet-4.5).
Env var	KILO_API_KEY (JWT tied to the Kilo account). Optional headers X-KiloCode-OrganizationId, X-KiloCode-TaskId, X-KiloCode-Version.
Auth	Authorization: Bearer <key>. Anonymous access allowed only for :free models (rate-limited per IP).
Key check	⚠️ GET /api/gateway/models is documented as no-auth — a 200 is inconclusive; a 401 indicates a malformed key. No dedicated /api-key validate endpoint. The reliable bad-key signal is a 402 on a paid call ({error:{code:402, metadata:{buyCreditsUrl}}}).
Quota / balance	❌ None. Per-request usage.cost_microdollars (1 USD = 1e6 microdollars), input_tokens, output_tokens, cache_*_tokens, is_byok. Balance is dashboard-only.
Plans	Kilo Code agent: Free/OSS, Teams $15/user/mo, Enterprise. Inference: Auto Free / BYOK / local ($0); Kilo Gateway ($0 + PAYG at exact provider rates, zero markup); Kilo Pass — Starter $19 (≈$26.60), Pro $49 (≈$68.60), Expert $199 (≈$278.60) with up to 50% bonus credits. KiloClaw managed OpenClaw $55/mo.
Billing	Credit-based, exact upstream provider rates, zero markup (microdollar precision). Free models (:free) cost nothing (200 req/hour/IP → 429). BYOK = $0 on Kilo's side.
Dashboard	app.kilo.ai: /credits, /subscriptions, /claw. Docs: kilo.ai/docs. Status: status.kilo.ai.
Changelog	Agent rebuilt on Kilo CLI. Auto Model router (Frontier/Balanced/Free). Kilo Pass subscription with bonus credits. KiloClaw managed hosting. Kilo Marketplace (spend credits on partner plans; MiniMax first). New BYOK: Inceptron, Kimi Code, Z.ai Coding Plan, Xiaomi Token Plans. Routes 500+ models across 60+ providers (no first-party model).
Adapter	fetch_kilo_native — best-effort /api/gateway/models probe; a 401 rejects the key, otherwise falls back to the configured-status note.

Kiro (kiro.dev — AWS)

API base	❌ None. Kiro is a subscription-only agentic IDE/CLI/Web with no public REST inference API and no API-key model. Login is interactive SSO (GitHub / Google / AWS Builder ID / AWS IAM Identity Center). CLI install: curl -fsSL https://cli.kiro.dev/install | bash.
Env var	— (no KIRO_API_KEY).
Auth	SSO login (not an API key). Enterprise uses AWS IAM Identity Center / SAML/SCIM.
Key check	❌ None.
Quota / balance	❌ None. Credits surfaced in-product only (IDE/CLI/Web subscription dashboard), refreshed at least every 5 minutes.
Plans	Free $0 (50 cr) · Pro $20 (1,000 cr) · Pro+ $40 (2,000 cr) · Pro Max $100 (5,000 cr, new 2026-06-10) · Power $200 (10,000 cr). $20 sign-up bonus on first upgrade. No daily/weekly rate limits. Credits metered to 0.01, no rollover, processed on the 1st. Overage $0.04/credit (opt-in). GovCloud ~+20%, no Free. HIPAA eligible (IDE+CLI) since 2026-05-26.
Billing	Monthly sub + credit multiplier per model: Auto 1×, Claude Sonnet 4.6 1.3×, Claude Opus 4.8 2.2×, Haiku 4.5 0.4×, DeepSeek v3.2 0.25×, MiniMax M2.5 0.25×. Taxes not included.
Dashboard	app.kiro.dev/settings/account. Docs: kiro.dev/docs/billing. Pricing: kiro.dev/pricing.
Changelog	2026-06-17 CLI v2.8 (CLI v3 early access). 2026-06-12 CLI v2.7 (/goal loops, queue steering). 2026-06-10 Pro Max tier. 2026-05-29 Claude Opus 4.8 (2.2×, 1M ctx). 2026-05-26 HIPAA eligible. CLI v2.6/v2.7 transcript export, persistent prefs.
Adapter	Informational card only (json_note_usage kiro-local). Links to app.kiro.dev. No scriptable surface.

Codex local telemetry

Alongside the app-server quota windows, the Codex card shows a local telemetry panel built from ~/.codex/sessions/**/*.jsonl (and archived_sessions/), honouring CODEX_HOME.

Source	event_msg / token_count events. Only usage counters, cwd, and the model name are read — never prompt or response content.
Cumulative counters	total_token_usage is cumulative for the session and is re-emitted unchanged on turns that add no tokens. The adapter sums positive deltas between consecutive snapshots, so repeated values are not double-counted and a counter reset (a new total lower than the previous one) restarts from zero instead of producing a negative row. last_token_usage is deliberately unused: it is also repeated.
Cost	Codex sessions record no cost, so every window reports cost as unknown (—). Consumption against a Plus/Pro subscription is not a dollar charge, and none is invented.
Bucketing	Rows are bucketed by the event's own local calendar day (via TZ), unlike pi/Hermes which bucket a whole session to its start day.
Adapter	providers/get-local-analytics codex, cached 120s.

Codex backend launch discipline

Every codex app-server startup runs a marketplace refresh round; when an upgrade keeps being interrupted, each round leaves git temp directories under ~/.codex/.tmp/ that Codex itself does not clean up (upstream issues: openai/codex#30620, #36093). A quota poll must therefore launch as few backends as possible:

• Daemon/proxy mode. When the Codex CLI supports it (standalone-installer installs), get-codex-usage runs codex app-server daemon start (idempotent) and speaks each poll through codex app-server proxy, so steady-state usage never starts a backend at all. npm/brew/distro installs without the daemon subcommand fall back to spawning a backend per refresh.
• CODEX_APP_SERVER_MODE=spawn forces the spawn path and skips the daemon probe.
• Single-flight. A flock on ~/.cache/AiOverviewControl/codex-usage.lock serializes launches; an invocation arriving while another refresh runs serves the cached snapshot, or waits up to CODEX_LOCK_WAIT seconds (default 8) before returning a structured error.
• Freshness gate. A cached snapshot younger than CODEX_FRESH_TTL seconds (default 60) is answered directly — widget reload bursts no longer spawn anything.
• Graceful teardown. The backend exits on its own when stdin closes; the adapter waits briefly for that self-exit before falling back to SIGTERM/SIGKILL, so startup work in flight is not orphaned mid-clone.

To reclaim disk from already-leaked staging clones (safe when no Codex session is running):

find ~/.codex/.tmp/marketplaces/.staging -mindepth 1 -maxdepth 1 \
  -type d -name 'marketplace-upgrade-*' -exec rm -rf {} +
find ~/.codex/.tmp -mindepth 1 -maxdepth 1 -type d -name 'git-*' -exec rm -rf {} +

Utility scripts

Beyond the per-provider adapters, providers/ ships these utilities:

• get-local-analytics <codex|opencode> — local harness telemetry (7-day chart, top models, top projects, per-window token breakdown), cached 120s. Shares its windowing with scripts/local-analytics.jq.
• local-cost-common — sourced helper building the Hermes cost SQL, so the summary and analytics adapters cannot drift on which ledger rows count as a known cost.
• get-usage-history — prints the local usage history written by the dispatcher (~/.cache/AiOverviewControl/usage-history.jsonl), trimmed by AIOC_HISTORY_MAX.
• export-usage-history — copies that history to CSV or JSONL (see configuration).
• get-provider-wrapper — shared single-provider wrapper behind the get-<id>-usage entrypoints.
• send-quota-alert — deduplicated desktop notification sender used by quota alerts (flock + notify-send).

Direct tests

Every provider is covered by providers/get-provider-health. Most providers also have a normalized providers/get-<id>-usage entrypoint; use the dispatcher for Claude and providers/get-pi-analytics for pi. Examples:

./providers/get-codex-usage | jq .
./providers/get-openrouter-usage | jq .
./providers/get-ollama-usage | jq .
./providers/get-xai-usage | jq .
./providers/get-kimi-usage | jq .
./providers/get-provider-health "codex,openrouter,ollama,xai,kimi" | jq .
./providers/get-provider-usage "gemini,cloudflare,mistral,glm,nvidia,minimax,kimi,qwen,xai,kilo,kiro" | jq .

Output schema

{
  "provider": "codex",
  "source": "codex-app-server",
  "usage": {
    "identity": {
      "providerID": "codex",
      "accountEmail": "account@example.com",
      "loginMethod": "plus"
    },
    "primary": {
      "usedPercent": 25,
      "windowMinutes": 300,
      "resetsAt": "2026-06-11T01:33:15Z",
      "resetDescription": "Session"
    },
    "secondary": null,
    "tertiary": null,
    "updatedAt": "2026-06-10T20:44:55Z"
  },
  "credits": { "remaining": "0" }
}

Errors use {provider, source, error:{code, kind, message}}. A provider without a quota API uses a normal informational usage object (via json_note_usage) rather than an error — primary.resetDescription carries the human label and primary.displayValue (optional) carries a string to render.