GC
Overview / OverviewMenu
3 min read·Updated 2026-04-10

Overview

Self-hosted agent runtime organized around a small set of primitives. This page explains the shape; the rest of the docs explains each piece.

What it is

GolemCore Bot is a self-hosted agent runtime. You deploy one container, open the dashboard, and talk to an agent that has access to tools, memory, and external services. Everything the agent does — model calls, tool execution, state persistence — happens inside your infrastructure.

This page describes how the runtime is shaped. It does not try to sell it. If you want to install it, jump to Quickstart. If you want to understand the internals, start with Architecture.

How the runtime is shaped

The runtime is organized around a small set of primitives. Each primitive has a clear role. You do not need to use all of them, but understanding how they relate is the fastest way to read the rest of the docs without getting lost.

Runtime (GolemCore Bot)

The process that runs agent sessions, executes tools, and persists state. Everything else plugs into it.

Core

Skill

A sticky overlay for a session: instructions plus optional MCP tool server. Shapes what the agent does in a given context.

Behavior

Plugin

A capability pack contributed to the runtime: tools, channels, voice, RAG. Adds things the agent could not do before.

Capability

MCP server

An external tool server launched from a skill. Its tools are exposed to the agent alongside native tools.

External tools

Memory

Structured recall across sessions with progressive disclosure. Keeps context small while preserving depth.

State

Model router

Abstract tier names (balanced, smart, coding, deep) mapped to concrete models. Tier choice is a separate decision from model choice.

Routing

Channel

Dashboard, Telegram, webhooks, Hive command flows. Any entry point that starts an agent turn.

Input

Hive

Optional fleet orchestrator for running multiple bots with approvals and shared inspection.

Optional

What happens in a turn

An agent turn is a straight line through the primitives above. A channel delivers input. The runtime selects a tier, resolves the concrete model through the router, and runs a tool loop. During the loop, the model can call built-in tools, plugin tools, MCP tools, and memory operations. When the loop finishes, the response goes back through the channel and state is persisted.

Every step is observable through Sessions, Logs, and Diagnostics in the dashboard. Nothing in the runtime is hidden from the operator.

Where to go next