Public beta · v0.1 A Novpix LLC product

One subscription.
Every coding agent.

powapi is a subscription-first LLM gateway with drop-in OpenAI and Anthropic compatibility — DeepSeek-first cascade with Sonnet escalation, 24-hour sliding quota windows, and per-tool endpoint presets.

// drop-in

Keep your client. Change one URL.

powapi speaks each tool's wire format directly. Point Claude Code, the OpenAI SDK, or the Anthropic SDK at a powapi base URL and the rest works unchanged — including streaming, tool use, and prompt caching.

~/.config/claude-code/.env drop-in
# Point Claude Code at powapi — keep your existing client
export ANTHROPIC_BASE_URL="https://api.powapi.io/claudecode/v1"
export ANTHROPIC_API_KEY="pwk_•••••••••••••••••"

# That's it. powapi handles routing, quota, and billing.
// how it works

Cascade routing, not a single-model lock-in.

Most agentic turns are routine — token-cheap models handle them faster and cheaper. powapi sends every request through DeepSeek first and escalates to Anthropic Sonnet only when the task or context window genuinely demands it. You get Claude-grade quality on the calls that need it without paying Claude prices on calls that don't.

DeepSeek-first

The default cascade target. Hits DeepSeek's prompt cache on repeat tool-loop turns, slashing input cost on the dominant request shape.

Sonnet escalation

An intent classifier and context-window guard promote a turn to Claude Sonnet 4.6 only when needed — long context, complex reasoning, or explicit model pin.

Cost guardrails

A live margin tracker keeps blended provider cost under target. If a route drifts above target, it falls back before billing the customer for it.

// quota

24-hour sliding windows, not monthly cliffs.

Subscription tokens refresh continuously across a rolling 24-hour window — not at midnight, not at month-end. Run a long agentic session without watching a counter creep toward zero. There is no concurrent-session cap either; coding agents fan out into sub-agents, and the gateway does not punish that pattern.

// endpoints

One gateway, native protocols.

No translation-layer leaks. powapi implements each tool's expected wire format end-to-end — streaming, tool use, prompt caching, the lot.

/claudecode/v1live /anthropic/v1live /openai/v1live /cursorsoon /clinesoon /aidersoon
// pricing

Four subscription tiers.
Pay-per-token overflow.

Effective $/M is held flat across tiers — you pay for headroom, not for a discount per token. Final figures and any active coupons are visible in the panel before checkout.

Starter
$9 / mo
1.7M tokens / 24h
~$0.18 / 1M effective
Max
$99 / mo
19M tokens / 24h
~$0.17 / 1M effective
Ultra
$450 / mo
100M tokens / 24h
~$0.15 / 1M · 10% loyalty

Pay-per-token API keys bypass the tier system entirely — same routing, transparent provider-priced metering with a flat 25% gateway markup. Issue separate keys per project or machine; track tokens, models, and cost in real time from the panel.

// foundation

Built on infrastructure you'd choose anyway.

No bespoke runtime, no proprietary client. Standard Python, standard databases, standard billing — plus the LiteLLM SDK for provider adapters so swapping or adding upstreams is a config change, not a rewrite.

Python 3.12 FastAPI httpx LiteLLM SDK Postgres 16 SQLAlchemy 2.0 Redis 7 Stripe structlog Next.js 15 panel
// team

Built and operated in-house.

powapi is engineered, shipped, and run by Novpix's own team — not an outsourced project. Direct access to the author when you hit a quota edge case or want a new preset for your tool of choice.

M
Mustafa Berber
Author & engineer · powapi
A Novpix LLC product

Try the cascade.

The panel handles signup, key issuance, and billing. Subscription gates an active paid plan — there is no free tier.

Open Panel