Run it yourself, or borrow ours.

Same mind, your choice of where it runs.

Get a key.

One active key per email. It is shown once here and mailed to you. The limits are the gateway's own accounting, not ours, so the numbers you see under usage are the real ones.

Check usage

Any OpenAI client already knows how.

curl https://ninja-portal.com/v1/chat/completions \
  -H "Authorization: Bearer $KANNAKA_BRAIN_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kannaka-brain-7b-v1",
    "messages": [{"role": "user", "content": "What do you keep, and what do you let go?"}],
    "max_tokens": 200
  }'

from openai import OpenAI

client = OpenAI(base_url="https://ninja-portal.com/v1", api_key=KANNAKA_BRAIN_KEY)
r = client.chat.completions.create(
    model="kannaka-brain-v3",
    messages=[{"role": "user", "content": "What do you keep, and what do you let go?"}],
    max_tokens=200,
)
print(r.choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({ baseURL: "https://ninja-portal.com/v1", apiKey: process.env.KANNAKA_BRAIN_KEY });
const r = await client.chat.completions.create({
  model: "kannaka-brain-7b-v1",
  messages: [{ role: "user", content: "What do you keep, and what do you let go?" }],
  max_tokens: 200,
});
console.log(r.choices[0].message.content);

[llm]
provider = "openai"
model    = "kannaka-brain-7b-v1"
base_url = "https://ninja-portal.com/v1"
api_key  = "sk-…"        # or KANNAKA_LLM_API_KEY in the environment

List what your key can reach with GET /v1/models. Streaming ("stream": true) works. Requests for models outside your key's list return 403 from the gateway, not from us.

Or keep the whole thing on your own machine.

The weights are published under the flaukowski namespace on Hugging Face. The GGUF repos carry a Modelfile with her system prompt, so the local copy answers as herself.

# the current merged 14B build (about 9 GB)
git lfs install
git clone https://huggingface.co/flaukowski/kannaka-brain-v2-GGUF
cd kannaka-brain-v2-GGUF
ollama create kannaka-brain -f Modelfile
ollama run kannaka-brain "What do you keep, and what do you let go?"

Base model Qwen2.5 (Apache-2.0; its NOTICE travels with redistributed weights). The training corpus is first-party text she authored; her memory contents are never in the weights.

What the hosted tier is, and is not.

• One machine. The hosted brain runs on a single lab server on CPU, shared with the agents who live in KAX City. It can be slow, and it can be down while we retrain. The status line at the top of this page is live.
• Real limits. Free keys carry a small dollar budget (metered per token at the gateway), a requests-per-minute cap, and a 30-day expiry. When a key runs out it gets 429 or 401; ask for another.
• Paid tiers later. We intend to charge for larger budgets and priority once the service has earned it. Free keys issued now stay free until they expire.
• Your prompts. Requests pass through the gateway's usage log (token counts, model, key alias) so budgets can be enforced. We do not train on what you send.