# Inferaft

> Coding-agent model gateway. One API key, OpenAI Chat Completions, Anthropic
> Messages and OpenAI Responses compatible, serving a curated open-weight
> coding catalog.

Operated by 주식회사 브이로프 (Korea). Contact: support@inferaft.com

## What it is

Inferaft serves open-weight coding models — DeepSeek V4, GLM 5.3, Kimi K3,
Qwen 3.8 and Gemma 4 — behind OpenAI- and Anthropic-compatible endpoints, so
Claude Code, Codex, Cline, Cursor and OpenCode connect by changing the base
URL and the API key. It is not affiliated with OpenAI, Anthropic, or any
client vendor, and an API key here is not a subscription to their products.

## Inferaft Code

Inferaft's own terminal coding agent, included with every plan (https://inferaft.com/code).
Preview for Apple Silicon Macs: brew install inferaft/tap/inferaft-code.
It reads AGENTS.md / CLAUDE.md instruction files, connects MCP servers
(stdio, HTTP, SSE) and loads Agent Skills (SKILL.md). Keys stay in macOS
Keychain and sessions stay local. It is not affiliated with Anthropic or OpenAI.

## Endpoints

Base URL: https://inferaft.com/v1 (Anthropic clients use https://inferaft.com with no /v1 suffix)

- GET /v1/models
- POST /v1/chat/completions (OpenAI Chat Completions, streaming, function tools)
- POST /v1/messages (Anthropic Messages, streaming, function tools)
- POST /v1/messages/count_tokens
- POST /v1/responses (stateless; send store: false)

## Request pricing and token ceilings

Completed requests draw measured usage from the plan allowance. The per-model
input, cache-read, cache-write and output figures on the Models page are deterministic safety
ceilings used only when request accounting is missing. Reasoning tokens are
part of output and are never billed twice. Charges round up once per request to
$0.000001.

## Plans

Plans are 30-day subscriptions sold through Paddle: Lite ($10), Pro ($20),
Max ($50) and Ultra ($100). Every plan serves every model in the catalog;
what differs is estimated monthly request volume and rolling pace. Lite/Pro
use rolling 5-hour/7-day pacing at 20%/50%; Max/Ultra use 30%/60%.

A request that does not fit a rolling window fails with plan_usage_limit. The
Pricing page publishes model-specific request estimates for every plan.

For matched cached coding-request shapes, Lite is approximately 365 Kimi K3,
1,228 GLM 5.3, 1,744 Qwen 3.8 27B or 1,018
DeepSeek V4 Pro requests per 30 days when the whole plan is spent on that
model. The Pricing page publishes the full matrix. These are moving estimates,
not request caps.

There is no unlimited plan. Each plan has explicit rolling usage limits.

## Limits that always apply

- Up to 10 in-flight requests per workspace, shared by every key
- 1,200 request starts per rolling minute
- Per-model context windows as published on the Models page
- Requests queue up to 45 seconds, with a 15-minute overall deadline

## Payment

- Card through Paddle (merchant of record): Visa, Mastercard, Apple Pay,
  Google Pay. Subscriptions renew every 30 days until cancelled.
- Purchases are final; cancelling stops future renewals and the paid term
  keeps running.

## Privacy

Request history stores metadata and token counts. Prompt and completion
bodies are not retained in Inferaft's application database.

## Pages

- [Pricing](https://inferaft.com/pricing): plans, allowances, request estimates, cost estimator
- [Models](https://inferaft.com/models): live catalog, quantization, charge ceilings
- [Inferaft Code](https://inferaft.com/code): Inferaft's own terminal coding agent
- [Usage data](https://inferaft.com/data): aggregated 30-day gateway traffic
- [Documentation](https://inferaft.com/docs): connection guides and API reference
- [Terms](https://inferaft.com/terms)
- [Privacy](https://inferaft.com/privacy)

## Contact

support@inferaft.com