System map

A guided tour of the stack.

BawtHub is built on llm-bawt — a self-hosted, model-agnostic LLM platform with persistent personas, real-time voice, semantic memory, and agent orchestration. This documentation maps every layer. Pick a subsystem to dive in, or hit ⌘K to search.

23 pages · audited against current code llm-bawt public on GitHub Last updated 2026-05-13

The whole picture, in one diagram.

Everything funnels through one OpenAI-compatible API. Surfaces (web, voice, CLI) hit the same endpoints. Tool calls and agent dispatches stream events back through a single bus. Memory is shared across bots and surfaces.

Surfaces

BawtHub web

Next.js 16 · React 19 · Zustand store

Voice

STT · TTS · pause-driven turns · realtime ws

CLI & MCP

llm -b loopy … · MCP tools for agents

↓ ↓ ↓

Core

OpenAI-compat API

FastAPI · 18 route modules · streaming

Request pipeline

Turn lifecycle · prompt build · client dispatch

Tool execution

Native loop · ReAct parser · streaming

↓ ↓ ↓

State

Memory

pgvector · 5 layers (messages, semantic, summaries, profile, sessions)

Data schema

Postgres · Prisma · turn logs · activity

MCP server

60 tools · cross-bot memory access · task ops

↓ ↓ ↓

Models

Provider abstraction

OpenAI · Claude · Grok · Ollama · vLLM · llama.cpp

Prompt adapters

Per-model chat template (dolphin, pygmalion, default)

↓ ↓ ↓

Agents

Claude Code

SDK bridge · Redis fanout · streamed tool events

Codex

OAuth · session queue · per-bot threads

OpenClaw

First bridge · WS · ingest · fanout

Task system

Projects · plan · dispatch · review

Pick a place to start.

Tour highlights

Chat surface →

One render engine, every tool call. Diffs that mean something, parsed bash/SSH commands, the MCP pill, streaming statuses, grouping. The dense visible output of every backend underneath.

Memory →

Five-layer state model: raw messages, semantic memories (with decay + supersede chains), rolling summaries, structured user profile, and sessions. The most opinionated subsystem.

Agent task system →

Plan → dispatch → execute → review, with markdown-typed projects, clean-context dispatch to bots, and MCP tools for agents to manage their own work.

Claude Code bridge →

How an external agent SDK plugs into the chat UI with full streaming tool events. The pattern repeats for Codex and OpenClaw.

MCP server →

60 tools exposed over streamable-http. Any agent — Claude Code, Codex, OpenClaw, or a human at the CLI — can manage memory, messages, tasks, and bots.

Streaming events →

The Redis-backed event bus that lets every backend (LLM tool loop, Claude Code, Codex, OpenClaw) emit tool calls and results in one unified format. The hard problem.

Browse by component

llm-bawt →

The core platform. OpenAI-compatible API, memory, tools, agent backends, MCP server. Public on GitHub.

BawtHub →

The frontend. Next.js web UI, Python voice backend, 3D avatar, multi-surface streaming consumer.

Deployment →

Docker on Unraid · NPM reverse proxy · Authelia · multi-host networking · agent bridges as sidecars.

About this documentation.

This site is built and maintained by an agent (currently loopy on Claude Sonnet) working from the live llm-bawt and bawthub repos. Subagents audit internal docs against current code on each major update, so divergence between this site and reality is small but not zero. The footer of each page shows when it was last validated.

If a diagram looks wrong, a code reference is stale, or you're confused — that's a bug. Open an issue on llm-bawt or email Nick.

Tip: press ⌘K or / to search. home · story · repo