What does “bring your own model” mean?

You use the API key of whichever LLM provider you choose (OpenAI-compatible), or a self-hosted model reachable over the internet. Harnix does not sell models.

Where is my data stored, and who can see it?

Each business gets its own data space; end users only ever see their own conversations; personal data can be deleted on request.

Is it tied to a particular industry?

No. An agent's expertise comes from your documents and your configuration.

Not announced yet. Design partners use it free for the length of the pilot.

AI AGENT OPERATIONS PLATFORM

In development · M2

Harness every run.

Harnix lets businesses put AI agents into the apps they already have, and see every answer end to end: what the agent read, which tools it called, and what it cost.

Join the waitlist Watch the demo

console.harnix.vn / runs / run_8c41f2

Run trace

completed

00.00suser_message

00.14stool_call knowledge_search

00.61stool_result · 3 chunks

00.72sllm_call · gpt-4o-mini

02.38stoken_usage · 1,240 tokens

Step detail

tool_result · knowledge_search

Chính sách bảo hành.pdf · page 4

“Customers may return an item within 30 days of delivery…”

Tokens in

1,024

Tokens out

216

Sample content, for illustration.

Building an AI assistant is easy. Operating one is not.

The AI answers wrong — who finds out?

An off-the-shelf chatbot is a black box. You only learn there is a problem when a customer complains.

Cost that moves with every question

If you do not know what each answer costs, you cannot keep the budget under control.

Where does customer data go?

Internal documents and customer conversations have to stay separated, and be deletable when a customer asks.

How it works

From nothing to an assistant inside your app

01Available

Load the knowledge

Upload documents (PDF, DOCX, TXT, MD) into a knowledge base.

02Available

Create an agent

Pick your model, write the system prompt, attach knowledge bases and tools.

03M2

Embed it in your app

Your backend exchanges the API key for a short-lived token; your frontend mounts the chat widget.

See the whole flow in 3 minutes

From an empty organisation to an AI assistant answering customers inside a sample app.

Demo video coming soon

Video chapters

01Sign in, empty organisationM0
02Load documentsM1
03Create an agentM0
04Try it and read the traceM1+M3
05Create an API keyM0
06Integrate it into an appM2
07An end user asks a questionM2
08Back to the consoleM3
09Wrap-up—

Notify me when the video is up

Anatomy of a Run

Every answer is a Run, recorded step by step.

Sample content, for illustration.

01user_messagecompleted

A customer asks: “How long is the return window?”

02tool_call knowledge_searchrunning

The agent searches the “Chính sách bảo hành.pdf” document.

03tool_result · 3 chunkscompleted

Three relevant passages found, with citations.

04llm_call · model gpt-4o-minicompleted

Drafts the answer from those passages.

05token_usage · 1,240 tokensmeasured

What this one answer cost, counted as tokens in and out.

06state_change · completedcompleted

Finished. It can be stopped partway if you need to.

Manage AI the way you manage a colleague

See every step

A full trace for every Run.

Stoppable

Cancel a Run while it is running.

Know the cost

Tokens and cost, answer by answer.

Bring your own model

An OpenAI-compatible endpoint, or self-hosted.

Data kept apart

One data space per business; personal data deleted on request (Decree 13/2023).

Durable

If a worker crashes, the Run carries on and finishes.

For developers

Preview · M2

// 1. Your backend — never expose the API key to the browser
const res = await fetch("https://api.harnix.vn/v1/end-user-tokens", {
  method: "POST",
  headers: { Authorization: `Bearer ${process.env.HARNIX_API_KEY}` },
  body: JSON.stringify({ end_user_id: user.id }),
});
const { token } = await res.json();

Illustrative code. The API is not final.

The API key never leaves your server

Streaming + citations

CSS-isolated widget (shadow DOM)

Read the integration docs →