Architecture

A single picture of how the platform fits together and how products plug in — and what makes it production-ready. The guiding principle: apps talk to Kiwi over its API; everything else is the shared core.

flowchart TB
    subgraph Consumers["Consumers"]
        direction LR
        KiwiUI["Kiwi UI"]
        subgraph Modules["MRCL Make modules"]
            direction LR
            BrandCheck["BrandCheck"]
            GenCanvas["GenCanvas"]
            GenScale["GenScale"]
            GenFlow["GenFlow"]
        end
        MRCLIdeate["MRCL Ideate"]
    end

    Kiwi["Kiwi Agent"]
    SP["System prompt"]
    Skills["Skills (coming)"]
    MS["Media Studio"]
    Kiwi --- SP
    Kiwi --- Skills

    subgraph Models["Models"]
        direction LR
        M1["OpenAI"]
        M2["Google Gemini"]
        M3["Anthropic<br/>(coming)"]
    end

    subgraph Tools["Built-in tools"]
        direction LR
        T1["Media"]
        T2["Files"]
        T3["Spreadsheets"]
        T4["Web"]
        T5["Code"]
    end

    Conn["Connectors MCP (coming)"]
    Sandbox["Sandbox (coming)"]

    Consumers <==>|API| Kiwi
    Kiwi <==>|API| Models
    Kiwi ==> Tools
    Kiwi ==> Conn
    Kiwi ==> Sandbox
    T1 <-. API .-> MS

    classDef kcore fill:#E8F5E9,stroke:#2E7D32,stroke-width:1.5px,color:#1B5E20;
    classDef kext fill:#E3F2FD,stroke:#1565C0,color:#0D47A1;
    classDef ktool fill:#F4F6F4,stroke:#6B7C6E,color:#1B2A1E;
    class Kiwi,MS kcore;
    class KiwiUI,BrandCheck,GenCanvas,GenScale,GenFlow,MRCLIdeate,M1,M2,M3 kext;
    class T1,T2,T3,T4,T5,Conn,Sandbox,SP,Skills ktool;

How to read it

• Consumers — the Kiwi UI (people), the MRCL Make modules, and the separate MRCL Ideate reach Kiwi over an API.
• Kiwi Agent — the core. It reasons over a system prompt, a model, its tools, and conversation memory.
• Models — the reasoning engines, called over an API: OpenAI and Google Gemini today.
• Built-in tools — twelve capabilities across five areas: media, files (video, audio, image, documents), spreadsheets, web, and code.
• Media Studio — generates image, video, and audio; the media tools call it over its API.
• Connectors · MCP (coming) — add external tools, and expose each module's own tools back to agents.
• Sandbox (coming) — a workspace the agent controls to run code, analytics, and dashboards.
• Skills (coming) — packaged, reusable instructions the agent loads for a task.
• Anthropic (coming) — a third model provider, alongside OpenAI and Gemini.

For the value behind this shape, see A shared core for faster innovation.

Request lifecycle

A single request follows the same path whether it comes from the web interface or a product calling the API:

1. In — a message arrives for a chosen agent, authenticated through SSO and authorized against what that agent and user may do.
2. Reason and act — the agent reasons about the request and calls the tools it has been granted, looping reason → act → observe (the ReAct pattern) until it can answer.
3. Tools reach out — tool calls fan out to the connected systems: documents and retrieval, spreadsheets, the web, code execution, Media Studio, and any MCP connectors.
4. Stream out — the response is streamed back as it is produced — reasoning, tool calls, text, and any media — rather than delivered all at once.
5. Persist — the session's history and any generated media are stored, so the conversation can be resumed and outputs reused.

Integrate over the API

Apps talk to Kiwi over its streaming API. A product sends a message and receives the agent's response as a live stream of events — reasoning, tool calls, text, and media — so it can render the agent at work in real time.

The platform provides the essentials so a product builds on top rather than rebuilding them:

• Authentication and access control — SSO and role-based permissions.
• Persistent sessions — conversation history kept and resumable.
• Shareable media outputs — generated media stored and returned as links.
• Inherited improvements — building on the shared core, a product gains new capabilities and fixes as they land.

This is how a product gains its agentic dimension — its features and actions driven by Kiwi agents. Explore it in the web interface first.

Built for production

These capabilities are provided by the platform, so teams building on the core inherit them rather than implementing them separately:

• Security — single sign-on; role-based access for agents and collections (owner and member roles, with some actions reserved for administrators); and encrypted credentials for connected tools.
• Reliable storage — your files, generated media, and conversation history are stored and remain available across sessions.
• Scale — Kiwi runs on managed enterprise cloud infrastructure (Azure), so teams using it don't operate the underlying services.
• Observability — activity is monitored and traced, so the operating team can diagnose issues and keep the platform healthy.

---

Related: Shared core · Key concepts · Getting started