Skip to content

Developer Documentation

Build agents, create capability packs, and integrate with the KruxOS platform.

Getting started

Architecture for developers

graph LR
    A[Your Agent] --> B[Python SDK]
    B --> C[Agent Gateway :7700]
    C --> D[Capability Registry]
    D --> E[Built-in Capabilities]
    D --> F[Community Packs]
    C --> G[Policy Engine]
    C --> H[Audit Logger]
    C --> I[Service Proxy]

Key concepts

Concept What it means for you
Capabilities Typed APIs your agent invokes — structured inputs/outputs, no shell parsing
Permission tiers Each capability has a tier: autonomous, notify, approval_required, or blocked
Structured errors Every error includes a type, message, and recovery suggestions
Packs Installable bundles of capability definitions + implementations
Service Proxy Safe external service access (Gmail + Slack adapters in v0.0.1) with read-replica, write buffer, batch protection
MCP Model Context Protocol — the Gateway's native surface; JSON-RPC is the documented fallback

Model connectors

KruxOS works with any LLM. Connector guides:

Model Protocol Guide
Claude (Anthropic) MCP (native) Connect Claude
OpenAI (+ Codex / DeepSeek / Grok / Mistral / Groq / GLM via base_url) Function calling adapter Connect OpenAI
Google Gemini Function declaration adapter Connect Gemini
Local (Ollama / vLLM / LM Studio / llama.cpp) OpenAI-compatible adapter Connect Local

OpenRouter (200+ models) is also supported as a provider — configure it from the dashboard Settings page.

Documentation map

Section Contents
SDK Guide Connection, discovery, invocation, errors, transactions, state
Connectors Model-specific integration guides
Pack Development Create, test, document, and publish packs
API Reference Protocol specs for MCP, JSON-RPC, WebSocket, REST
Capability Reference Auto-generated from YAML definitions
Agent Experience Side-by-side: traditional Linux vs KruxOS
Contributing How to contribute to KruxOS