A game NPC dialogue API

What NPC dialogue needs, and what you get

What a thousand talks cost

On GLM the same thousand talks are about $1.34 before its brief think, but that think comes before every line, which costs time an NPC can’t spare. Short persona prompts are cheap either way; a long lore sheet in every call is what adds up.

A line of dialogue, streamed

Run this on your game server, not in the client: a key shipped inside a game can be pulled out of it. Show each piece as it arrives, so the speech bubble fills as the player watches.

import os
from openai import OpenAI

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

persona = "You are Brannoc, the blacksmith of Hollowmere: gruff, superstitious, owes the thieves guild money. The player has the quest \"Missing ore\" and has not yet found the mine. Speak only as Brannoc, in one to three short sentences. Never mention game mechanics."

stream = client.chat.completions.create(
    model="deepseek-v4.1-flash-uncensored",
    messages=[
        {"role": "system", "content": persona},
        {"role": "user", "content": "Have you seen any ore carts come through town?"},
    ],
    stream=True,
    max_tokens=120,
    temperature=0.8,
)
for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="", flush=True)

A line your game can act on

When the NPC’s answer should change the game, ask for JSON instead of streaming prose: the line to show, a mood for the animation, and an action for your code. Describe the shape in the prompt as well.

import json

reply = client.chat.completions.create(
    model="deepseek-v4.1-flash-uncensored",
    max_tokens=200,
    response_format={"type": "json_object"},
    messages=[
        {"role": "system", "content": persona + ' Answer as JSON: {"line": string, "mood": "calm" | "angry" | "afraid", "action": "none" | "give_map" | "attack"}.'},
        {"role": "user", "content": "Tell me where the mine is or I tell the guild you skipped town."},
    ],
)
npc = json.loads(reply.choices[0].message.content)
print(npc["line"], npc["mood"], npc["action"])

Many players at once

• Each key takes 120 requests a minute, and an account can hold 10 keys. Past a 429, the Retry-After header says how long to wait. Rate limits.
• Give each environment its own key with a daily cap: a bug that loops a dialogue stops at the cap, not at your balance. Keys and caps.
• Keep personas short and fixed. A persona of about 1,000 tokens or more that repeats from call to call is read from cache at a fraction of the price.
• Write a fallback line for when a call fails, so a busy moment never freezes a scene. A 503 busy costs nothing.

Which model

Use deepseek-v4.1-flash-uncensored for spoken lines: it answers without thinking first, takes samplers and returns JSON. glm-5.3-flash-uncensored is cheaper per token ($0.20 in, $0.65 out per million, against $0.45 and $1.75) but thinks briefly before every answer, so it suits work the player isn’t waiting on, like writing tomorrow’s town gossip. Compare them.

What we keep of players’ words

One row per call: the time, the key, the model, token counts, credits and the status, kept for 30 days. Never what the player typed, never the NPC’s line. Nothing is added to your persona prompt. What we keep.