An API for interactive fiction and text adventures

What a text game needs, and what you get

What a session costs

The whole history is sent with each turn, so the 60th call is far longer than the first. Caching is why that stays cheap: the part a call repeats from the call before it is read from cache. Without a single cache hit, the same session is $0.30 on DeepSeek. How caching works.

A narrator in twenty lines

The world and the rules go in the system message. Keep it the same from turn to turn, so the start of every call repeats and is read from cache. Stream the reply and print it as it arrives.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://nokeep.ai/v1",
    api_key=os.environ["NOKEEP_API_KEY"],
)

history = [{"role": "system", "content": "You are the narrator of a dark fantasy text adventure. The player is a smuggler in the drowned city of Vell. Describe what they see, hear and smell in second person, in under 150 words. Never act or speak for the player. End each turn with the scene open, not with a list of options."}]

def turn(move):
    history.append({"role": "user", "content": move})
    stream = client.chat.completions.create(
        model="deepseek-v4.1-flash-uncensored",
        messages=history,
        stream=True,
        max_tokens=400,
        temperature=0.9,
        extra_body={"min_p": 0.05, "repetition_penalty": 1.05, "prompt_cache_key": "campaign-42"},
    )
    text = ""
    for chunk in stream:
        if chunk.choices:
            piece = chunk.choices[0].delta.content or ""
            print(piece, end="", flush=True)
            text += piece
    history.append({"role": "assistant", "content": text})

turn("I row the stolen skiff under the old customs bridge.")

prompt_cache_key is any string you keep the same for one campaign. It sends the campaign’s calls where its start is already cached. Streaming covers usage in the last chunk and stopping a reply early.

Keep the game in your code

• State outside the prose. Keep inventory, health and flags in your own data, and put a short summary of them in the system message. A model remembers a story well and a number badly.
• Ask for JSON when you need facts back. A second, short call with response_format can return what changed this turn, for your code to apply. JSON output.
• Long campaigns. A million tokens is a lot of turns, but each one is billed. Past a few hundred turns, replace the oldest with a summary you write once, and keep the summary in the same place so the start still repeats.
• Branching and undo. The API keeps nothing between calls, so a save point is just the message list at that turn. Resend an earlier list to rewind.

Which model

What happens to your players’ stories

We keep one row per call: the time, the key, the model, token counts, credits and the status, for 30 days. Never the messages, never the reply. DeepSeek runs on a private tier that stores no prompts or replies. Players’ stories live wherever your game keeps them, and nowhere of ours. What we keep.