Make it
your own.

PUBLIC DEVELOPER ACCESS

Access is
in preparation.

GH0ST has an existing controlled Robinhood → ICP → Robinhood inference route. General developer requests are not yet publicly admitted.

The selected public integration uses GH0ST-paid on-chain requests. Its reviewed deployment is not active yet. This advanced page describes an optional independent local runtime.

The download contains developer clients, source and examples. The advanced instructions below describe an optional self-hosted runtime; they do not connect to the existing GH0ST canister deployment.

01 / ADVANCED SELF-HOSTING OPTION

A local runtime.
A real response.

You need Node.js 22 or later, Python 3.10 or later with venv/pip, and a C++ build toolchain. The native CPU runtime is built on your machine. Model setup downloads about 1.11 GB from the pinned upstream source and verifies its complete SHA-256.

curl --fail --location --output gh0st-agent-kit-0.3.0-alpha.1.tgz \
  https://gh0st.fyi/Agentkit/downloads/gh0st-agent-kit-0.3.0-alpha.1.tgz

mkdir gh0st-agent-kit
tar -xzf gh0st-agent-kit-0.3.0-alpha.1.tgz -C gh0st-agent-kit --strip-components=1
cd gh0st-agent-kit
npm install --ignore-scripts --no-audit --no-fund
npm run runtime:setup
npm run model:setup
npm run api

Wait for Qwen is ready. The API listens on http://127.0.0.1:8790/v1. Keep that terminal running while your project makes requests.

If your current python3 is older than 3.10, select a supported executable explicitly:

GH0ST_QWEN_SETUP_PYTHON=/absolute/path/to/python3.12 npm run runtime:setup

The archive contains source, clients and examples. It does not contain model weights or a prebuilt native environment. Model and runtime setup are required on each inference machine.

02 / JAVASCRIPT & TYPESCRIPT

A few lines.
Your application.

In your own project, install the downloadable package directly. This alpha is distributed as a tarball; it is not an npm-registry release.

npm install https://gh0st.fyi/Agentkit/downloads/gh0st-agent-kit-0.3.0-alpha.1.tgz

Create example.mjs, then run it with node example.mjs while the local API is running:

import { createQwenClient } from '@gh0st/agent-kit/client';

const qwen = createQwenClient({
  baseURL: 'http://127.0.0.1:8790/v1'
});

const response = await qwen.generate(
  'Give this idea a voice.'
);

console.log(response.content);

The client returns content, finishReason, token usage and gh0st metadata. TypeScript declarations are included. Supply the full conversation through qwen.chat({ messages, maxTokens, temperature }) when you need context.

03 / PYTHON

The same model.
Your language.

From the extracted kit directory, install the Python client into your project's Python environment:

python3 -m pip install ./python

from gh0st_qwen import GH0STQwen

qwen = GH0STQwen(
    base_url="http://127.0.0.1:8790/v1"
)

print(qwen.generate("Give this idea a voice."))

generate() returns text. chat(messages, max_tokens=256, temperature=0.3) returns the complete JSON response. The Python client has no runtime dependencies beyond the standard library.

04 / HTTP

Your client.
A small interface.

Use the local API from any backend that can send JSON over HTTP:

curl --fail-with-body --silent --show-error \
  http://127.0.0.1:8790/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "gh0st-qwen3-1.7b",
    "messages": [{"role": "user", "content": "Welcome a new developer in one sentence."}],
    "max_tokens": 128,
    "temperature": 0.3,
    "stream": false
  }'

Read answer text at choices[0].message.content. The response includes actual token usage and model-artifact metadata. GET /v1/models lists the supported model ID.

05 / BOUNDED TOOLS

Let the model ask.
Your code decides.

The kit's application-side tool loop accepts registered functions with argument validation, step limits and approval rules. The model proposes bounded JSON actions; generated text is never evaluated as code.

npm run example:agent

The included project-catalog example runs against your local API. Native tool-call fields and tool-role messages are not supported by the HTTP endpoint; tool orchestration happens in your application through @gh0st/agent-kit/agent.

06 / WHAT THIS ALPHA SUPPORTS

Know the
boundaries.

Model

Qwen3-1.7B Q4_K_M · gh0st-qwen3-1.7b

Context / output

4,096 context tokens · up to 512 output tokens

Messages

Text-only system, user and assistant messages

Delivery

Non-streaming. The API does not retain conversations.

Execution / payment

Local inference. No production GH0ST-token billing.

Model output is untrusted text. The model hash identifies an artifact; it is not an ICP execution proof or proof that an answer is correct. Hosted inference, public capacity guarantees and production billing are separate work.

The server binds to loopback. An optional GH0ST_QWEN_API_KEY requires a matching bearer key in your client. Remote hosting, TLS and access control are separate deployment work.

GH0ST supplies this integration layer. Qwen is an upstream model; review the upstream model card and quantized model source for their licensing and attribution.

Full integration notes ↗Pinned model details ↗Python client guide ↗