bibo

A coding agent for local models on consumer hardware

Fortified with industry-grade system prompts from Claude & OpenCode

My Server

• GPU 1: NVIDIA V100 SXM2 (w/ PCIe Adaptor) 32GB for attention layers
• GPU 2: NVIDIA RTX 3080 (20GB MOD) for MoE expert layers
• CPU: AMD Ryzen 5 7500X3D
• RAM: 32GB DDR5 (The X3D suffices for RAM shortage and single channel implications D:)

Currently running Qwen-3.6-35B-A3B Q5 with llama.cpp.

Also compile llama.cpp yourself for better custom tailoring!

Quick Start

git clone https://github.com/exoad/bibo.git
cd bibo
npm install

Configure the Model

Bibo uses a provider.json file to configure model providers. The llama-cpp-provider extension reads this file at startup.

First, create your config:

cp provider.json.example provider.json

Then edit provider.json to point at your model endpoints:

Field	Purpose	Default
baseUrl	Model server endpoint	http://127.0.0.1:6969/v1
apiKey	API key (if required)	noop

Local llama.cpp (running on this machine):

{
  "providers": [
    {
      "name": "jackbox",
      "api": "openai-completions",
      "baseUrl": "http://127.0.0.1:6969/v1",
      "apiKey": "noop"
    }
  ]
}

Remote server (your own model host):

{
  "providers": [
    {
      "name": "jackbox",
      "api": "openai-completions",
      "baseUrl": "http://<your-server-ip>:6969/v1",
      "apiKey": "noop"
    }
  ]
}

Tip

provider.json is ignored by VCS — keep it local. Use provider.json.example as a template.

Note

If provider.json doesn't exist, the extension falls back to environment variables: LLAMACPP_BASE_URL, LLAMACPP_API_KEY, OLLAMA_BASE_URL, OLLAMA_API_KEY.

Run

./bin/bibo.sh

Dashboard

Bibo includes a web-based dashboard that shows sessions, brain memory, vault notes, and quests in real-time.

bibo                    # Start with dashboard
bibo "your prompt"      # Start with a specific prompt

The dashboard runs on http://127.0.0.1:3000 and auto-starts when you run bibo.

Global Alias

To use bibo from anywhere, add this to your ~/.zshrc (or equivalent shell config):

alias bibo='~/bibo/bin/bibo.sh'

On Windows, use:

doskey bibo=C:\bibo\bin\bibo.cmd $*

Setup

Model Provider Configuration

The llama-cpp-provider extension (.pi/extensions/llama-cpp-provider/index.ts) reads provider.json at startup and registers providers dynamically.

If provider.json doesn't exist, it falls back to environment variables:

• LLAMACPP_BASE_URL - llama.cpp OpenAI-compatible endpoint
• LLAMACPP_API_KEY - API key (leave as noop if no auth)
• OLLAMA_BASE_URL - Ollama endpoint (default: http://127.0.0.1:11434/v1)
• OLLAMA_API_KEY - Ollama API key (default: noop)

pi Settings

See .pi/settings.json for model profiles, compaction, and retry settings.

Structure

Path	Purpose
bin/	Cross-platform bibo launcher (macOS/Linux/Windows)
dashboard/	Web dashboard for sessions, brain, vault, and quests
.pi/extensions/	20 TypeScript extensions (auto-discovered)
skills/	Markdown cheat sheets injected on demand
AGENTS.md	Project system prompt
SYSTEM.md	Extended system prompt
SELF.md	Internal architecture reference
.pi/settings.json	Per-model profiles, compaction, retry settings

Custom Extensions & Skills

Local extensions (.pi/extensions/):

• benchmark-profiles - profiling benchmarks
• breathing-border - animated chatbox border
• browser - browser extraction
• browser-extract-retention - retention-aware extraction
• checkpoint - session checkpointing
• context-mode - context management (inlined from npm)
• cost-tracker - fake cost counter
• evidence - evidence collection
• evidence-compact - compact evidence bridge
• extra-tools - additional tool integrations
• hello - greeting widget
• knowledge-inject - knowledge injection
• llama-cpp-provider - llama.cpp model provider
• output-parser - structured output parsing
• permission-gate - tool permission gating
• quality-monitor - repeated tool call detection
• shell-session - persistent shell sessions
• skill-inject - skill injection
• thinking-budget - thinking token budget
• timer - session timer
• tool-gating - tool access control
• turn-cap - turn limiting
• write-guard - write operation protection

NPM packages (.pi/settings.json):

• @feniix/pi-code-reasoning - code reasoning
• pi-gsd - GSD workflow system
• pi-web-access - web search/fetch
• pi-markdown-preview - markdown preview
• @rhobot-dev/rho - rho agent framework
• pi-token-burden - token budgeting
• @feniix/pi-statusline - statusline widget
• pi-review-loop - automated code review
• pi-quests - quest/task management
• pi-tool-display - tool display
• pi-mermaid - mermaid diagrams
• pi-rtk-optimizer - RTK optimizer

Customization

• Extensions - each under .pi/extensions/<name>/. Disable by removing the directory or editing .pi/settings.json. Add your own as a TypeScript module.
• Skills - auto-discovered from .pi/skills/ and skills/. Disable with --no-skills.
• Settings - see pi settings docs.

Attribution

Fork of little-coder → built on pi.

• pi - Apache 2.0 / MIT
• little-coder - Apache 2.0