Skip to main content
Ultron is built on a five-layer stack. Each layer has a single responsibility. Together they make it possible for a single chat message to trigger parallel web searches, tool executions, memory saves, and real-time streaming — all in one coherent loop.

​
The five-layer stack

Ultron Platform
├── Layer 1: Interaction     — Chat UI, SSE streaming, tool activity feed
├── Layer 2: Orchestration   — Session state, cost tracking, transcript persistence
├── Layer 3: Core Loop       — compress → call Kimi K2 → execute tools → loop
├── Layer 4: Tools           — 50+ built-in tools + dynamic MCP server discovery
└── Layer 5: API             — Kimi K2 via Moonshot AI (Anthropic-compatible)

​
Layer breakdown

Layer 1 — Interaction The chat interface. Captures input, streams tool activity in real-time via SSE, renders results inline. Canvases render as React components inside the conversation thread. Nothing requires a page switch. Layer 2 — Orchestration Manages the session. Tracks token usage across turns, persists transcripts to JSONL, maintains file history snapshots, and holds the permission context for tool calls. This layer keeps multi-turn conversations coherent and resumable. Layer 3 — Core Loop The agent loop. Each iteration: run compression if needed → call Kimi K2 with full context → collect tool calls → execute tools (in parallel where possible) → feed results back → check if another iteration is needed. Repeats until the task is complete. Layer 4 — Tools The execution surface. 50+ built-in tools (web search, browser automation, CRM lookups, email, enrichment, document generation) plus dynamic MCP server discovery. Every session starts with 50+ MCP servers already loaded and available. Layer 5 — API Kimi K2 via Moonshot AI’s Anthropic-compatible endpoint. Model routing selects Kimi K2 for complex reasoning and a fast model for lightweight tasks. All API communication is streamed — responses and tool calls arrive as they’re generated, not after the full response completes.

​
Design principle

The user never leaves the chat. No separate apps for research. No external CRM for outreach. No content tool for publishing. One interface, every capability.

Chat Interface

How SSE streaming, tool activity, and inline rendering work.

Execution Engine

The sandboxed runtime, tool pipeline, and 50-minute window.

Parallel Execution

75 concurrent tasks — how the math works and why it matters.

Model Routing

How Ultron picks between Kimi K2 and fast models per request.