What a self-serve AI platform is, and what serverless really means for live agents

Direct answer: A self-serve AI platform lets a developer sign up, get an API key and put a live Human AI Agent on their own site in minutes, with no sales call. The word to watch is serverless. It means provisioning is invisible to you, not that GPU cold starts have gone away.

Sign up, drop in a snippet, and talk to a live agent. A self-serve AI platform promises exactly that.

No sales call, no onboarding deck. You hit the docs, get a key, and within minutes a real-time Human AI Agent is running on your own site. Serverless is the word to watch, and the rest of this page explains why.

What a self-serve AI platform is

A self-serve AI platform lets a developer go from curiosity to working integration without talking to a human on the vendor side. You read the docs, grab an API key, run a quickstart, and ship. The product teaches you instead of a sales engineer doing it for you. Some integrations only need a plain AI API to call from a backend script, others want the full AI agent SDK, with embed components and event hooks wired into a front end. Either way the interface has to be legible without a human explaining it, which is the job the OpenAPI Specification, maintained by the OpenAPI Initiative under the Linux Foundation, was written to do for HTTP APIs.

In most software categories that is now normal. For live agents it is rare, because what you spin up is a running conversation: a face rendered in real time, a voice generated as the person speaks, and presence that holds attention across turns. A Human AI Agent is the bundle that covers those layers: speech recognition, reasoning, a face model and speech synthesis. The face model is Oris Portrait or Oris Presence, and the sub-200ms budget belongs to that model alone. Two modes cover how you plug it in. Ojin Agent runs the whole bundle, or Third-Party Agent lets you bring your own speech provider, with several speech providers supported out of the box and listed in the docs, and the face model added on top. The layer-by-layer definition sits in what a Human AI Agent actually is.

The pull toward self-serve is a broader buyer shift rather than an Ojin idiosyncrasy. Gartner predicted in September 2020 that 80 percent of B2B sales interactions between suppliers and buyers would occur in digital channels by 2025. Developer tooling has led that shift for years: reading the docs and shipping a quickstart is simply what "evaluating a vendor" now means for the people building the integration.

From sign-up to a live agent in minutes

The flow is short on purpose. Create an account at ojin.ai/signin and get a key, pick a base agent or upload a portrait and voice reference, then drop the <ojin-agent agent-id="your-agent-id"> embed snippet into your page or call the developer API from your backend. Open the page and the agent loads and starts talking, live. There is no queue, no export step and no wait for a video to render, because the agent is a session rather than an artifact. The docs at docs.ojin.ai also ship an MCP server, built to the 2025-06-18 revision of the Model Context Protocol specification, so an AI coding assistant can read the API surface directly instead of you pasting reference pages into a prompt.

The integration detail sits in the real-time agent API guide, and the rest of the library is indexed at Ojin Insights. A docs-first experience is how a developer platform earns trust in the first place, through documentation and technical depth rather than a glossy landing page.

Self-serve real-time embed versus render-queue tools

Pre-rendered avatar tools do one job well: type a script, pick a presenter, a video renders. The output is a finished file. It is ideal for training clips and cannot hold a conversation. Template-based avatar tools push closer to interactivity but lean on preset flows. Voice-only platforms are superb at speech and stop there. They give you no face and no presence.

A real-time embeddable agent is a different shape of product. The face, voice and presence are generated live, in response to whatever the person says, and there is no render step because there is no final file. The bar for that kind of exchange is old and well documented: ITU-T Recommendation G.114 (International Telecommunication Union, May 2003) notes that if one-way delay is kept below 150 ms, "most applications would not be significantly affected". If your use case is shipping a polished video, a render-queue tool wins. If it is letting a visitor talk to someone and see a face do it, the live embed is the right category. Developers weighing voice-only self-serve APIs deserve a straight answer here too: those platforms are strong, well documented and quick to integrate for audio-only agents, and worth using if a face adds nothing to your product. The shape of a live exchange is unpacked further in the conversational AI agent guide.

Why a live agent cannot be fully scale-to-zero

Most self-serve platforms market themselves as serverless: never think about machines, just call an endpoint. That holds up for stateless inference, where a request arrives, runs, and returns, and a cloud AI agent can scale those calls to zero between requests with nobody noticing.

Why cold starts break a live conversation

Real-time presence does not fit that model. A live agent is a held-open, low-latency session rather than a single request, with a face and voice generated continuously. To hold the face model inside its sub-200ms budget, the capacity behind it has to be warm. A truly cold, scale-to-zero box cannot spin up a GPU, load the model, and start rendering a face inside the delay budget the ITU sets for interactive speech. Cold starts are fine for a webhook. They are a problem the moment someone has just said hello.

The production traffic that scale-to-zero was designed around looks nothing like a conversation. In their study of the Azure Functions workload, presented at USENIX ATC in July 2020, Shahrad and colleagues report that "most functions are invoked very infrequently, but there is an 8-order-of-magnitude range of invocation frequencies", which is exactly the sparse, bursty shape a cold pool suits and exactly what a held-open session is not. Modal, which sells serverless GPU capacity, reports in its own engineering write-up on GPU memory snapshots that cold starts fall from 45 seconds to 5 seconds for a small language model served with vLLM, and from 20 seconds to 2 seconds for a speech transcription model. Those are real engineering gains, and, as the ServerlessLLM work on checkpoint loading presented at USENIX OSDI in 2024 also shows, they still land in whole seconds, nowhere near what a live conversation can tolerate on a fully cold box.

Serverless infrastructure works beautifully for the stateless parts, and Ojin uses it there, while real-time presence needs warm capacity standing by. You get the serverless developer experience with a warm pool underneath, the arrangement AWS documents for Lambda as pre-initialised execution environments. Any platform claiming fully scale-to-zero live presence is describing a cold start the caller will hear.

Pricing, the free tier, and building on it

Self-serve only works if you can try before you pay. Ojin gives new accounts $10 in free sign-up credit, enough to stand up an agent and watch it hold a real conversation before deciding anything. Pricing for real-time presence is usage-based, because the cost driver is conversation time and concurrency rather than a flat per-video fee, which makes it a different model from how render-queue tools charge. Current rates are published at ojin.ai/pricing rather than repeated here, since usage-based figures change faster than a guide like this gets updated. On the build side, the agent wires into your own systems through webhooks and event hooks. The first time your agent answers a question you did not script, the effect is disconcerting. That reaction is the point.

When you outgrow self-serve

Self-serve is the front door. The ceiling is somewhere else, and teams usually find it in procurement. They start self-serve, ship something, and only talk to a human when they need volume commitments, custom models, data residency guarantees, or compliance paperwork. Being EU-based in Berlin, where Ojin works out of AixHaus, matters here, because EU data residency is a hard requirement for many teams and the EU AI Act entered into force on 1 August 2024.

The infrastructure underneath

The same infrastructure sits under every Ojin entry point, and self-serve is simply the door you can open on your own: a real-time inference platform built for streaming responses across face, voice and reasoning rather than a text-only pipeline. The same platform seen from a commercial angle is covered in the AI sales agent guide.

Common questions about self-serve live agents

What is a self-serve AI platform?

A platform you adopt without a sales call. You read the docs, get an API key, run a quickstart, and ship. Ojin applies it to live agents, so you go from sign-up to a running real-time Human AI Agent in minutes.

Can I embed a live AI agent on my own site?

Yes. Ojin gives you an embed snippet and an API, and the agent runs on your page as a live conversation rather than a pre-rendered video.

Is a live AI agent really serverless?

The stateless parts run on serverless infrastructure. Real-time presence needs warm capacity to hold the face model's sub-200ms budget, so it stops short of fully scale-to-zero. You get the serverless developer experience with a warm pool underneath.

How is Ojin different from recorded and template avatar tools?

Those render video from a script or run template flows. Ojin generates face, voice and presence live, in response to unscripted input. The face model is Oris Portrait or Oris Presence, working inside its sub-200ms budget, and it arrives bundled with speech recognition, reasoning and speech synthesis as a Human AI Agent.

Is Ojin's self-serve platform better than voice-only agent platforms for developers?

That depends on what you are building. Voice-only agent platforms are strong, well-documented self-serve platforms for audio-only agents. Ojin is built specifically for when the agent needs a face, using the same real-time latency discipline those voice-only platforms compete on.

Can I bring my own LLM?

Yes. The reasoning layer stays yours; Ojin provides the real-time face, voice, and presence around it.

How do I start?

Start free at ojin.ai/signin, using docs.ojin.ai to guide you. New accounts get $10 in free sign-up credit, and you can have a live agent running in minutes.

Start building

Start free at ojin.ai/signin and use docs.ojin.ai to guide you. New accounts get $10 in free sign-up credit, and your live agent runs in minutes. Sign up, drop in the snippet, and talk to it.

Read next