Kopai

Kopai is a serverless cloud platform that lets creators build domain-specialized AI agents and monetize them through a built-in marketplace. It handles the full infrastructure stack — RAG, model orchestration, vector databases, and per-message billing — so builders can focus on the agent's expertise rather than backend code. Agents built on Kopai can be listed on the marketplace, embedded in external products via an upcoming SDK, or kept private for team use.

What It Is

Kopai sits at the intersection of an agent-building platform and a marketplace. Creators write a system prompt, upload documents, connect tools, and run the agent through an evaluation gate before it can be listed publicly. The platform manages model keys, vector indexing, streaming resilience, and per-message billing on behalf of the builder. Buyers spend credits per message; builders keep 70% of their markup on top of runtime cost.

Seven-Step Build Workflow

The platform structures agent creation into eight stages: Prompt & Model, Connect, Preview, Race Models, Evaluate, SEO, Publish, and Analytics. Each step is discrete and interactive:

• Prompt & Model — Write instructions in plain language and pick from models including GPT-4o, Gemini 2.5, Claude Sonnet, Kimi K2, Qwen-3, and DeepSeek.
• Connect — Attach connectors and subagents (up to 6 connectors and 5 subagents on paid plans).
• Race Models — Run the same prompt across multiple models side by side to compare outputs.
• Evaluate — Pass 9 adversarial test suites and 5 evaluation passes before the agent can be listed.
• Publish — List on the marketplace or keep private; set a per-message markup.
• Analytics — Monitor usage and earnings after launch.

Evaluation Gate and Safety Architecture

Kopai's evaluation system has three components that operate outside the creator's control:

• Risk tier assignment — The platform automatically assigns a risk tier (Low, Medium, High) based on the agent's configuration. An agent giving financial advice receives the strict bar regardless of creator preference.
• Groundedness checking — A second model reads each answer back against the sources it cited and fails responses where the two diverge.
• Fingerprinting — A passing evaluation score is tied to the exact prompt version that earned it. Editing the prompt invalidates the pass and requires re-evaluation.

Infrastructure Managed by Kopai

The platform abstracts away the entire runtime layer:

• Model gateway — One integration covers all supported models; API keys are held by Kopai.
• Streaming resilience — Answers resume from where they stopped if a connection drops, with zero tokens lost.
• Document indexing — Uploaded PDFs, DOCX, and XLSX files are indexed automatically and searched at query time.
• Per-message ledger — Every message is priced, charged, and logged in real time; both runtime cost and creator markup are itemized separately.

Marketplace and Monetization Model

Agents that pass evaluation are listed on the Kopai marketplace. Buyers can try each agent for free for its first few messages before spending credits. Creators set a per-message markup and, according to Kopai's published terms, keep 70% of that markup. Credits can be purchased as one-off wallet top-ups or via monthly rolling plans; auto-reload adds a bonus on the standard top-up rate. An Export/SDK feature that allows embedding agents into external products while billing through Kopai is listed as "shipping next" with a waitlist.

Teams and Organizations

Kopai supports multi-user workspaces with seat-based roles, shared knowledge libraries, and private agents visible only to team members. The Free plan supports up to 2 members; paid plans expand seat counts and storage. Enterprise plans offer unlimited members, custom SAML/SSO authentication, priority support, and custom connectors on request.