My Intelligence Stack

A portable, Docker-based AI stack combining LiteLLM (model gateway), OpenWebUI (chat frontend), and Codex (agent CLI), backed by OpenRouter, Fireworks, and OpenAI.

What this packages

Layer	Tool	Purpose
Frontend	OpenWebUI	Web-based chat UI
Gateway	LiteLLM	Routes models across providers, logs spend
Agent CLI	Codex	Terminal-based coding assistant
Agent CLI	Claude Code	Terminal-based coding assistant
Inference	OpenRouter / Fireworks / OpenAI	Model providers
Storage	Postgres	Chat history and LiteLLM database

Architecture

graph LR
    User["🧑 User"] -->|browser| OpenWebUI
    User -->|terminal| Codex
    User -->|terminal| ClaudeCode["Claude Code"]

    OpenWebUI -->|OpenAI-compatible API| LiteLLM
    Codex -->|OpenAI-compatible API| LiteLLM
    ClaudeCode -->|Anthropic-compatible API| LiteLLM

    LiteLLM -->|routes to| OpenAI["OpenAI"]
    LiteLLM -->|routes to| Fireworks["Fireworks"]
    LiteLLM -->|routes to| OpenRouter["OpenRouter"]

    LiteLLM -->|stores chat history| Postgres["Postgres"]

Loading

Quick start

One-liner:

curl -fsSL https://raw.githubusercontent.com/asaini/openmodelstack/main/install.sh | bash

Or clone and run locally:

git clone https://github.com/asaini/openmodelstack.git
cd openmodelstack
./install.sh

The installer will:

1. Check for Docker
2. Create .env from .env.example and prompt for API keys
3. Generate random internal secrets (Postgres, LiteLLM master key, WebUI secret)
4. Copy the Codex and Claude Code configs to ~/.codex/config.toml and ~/.claude/settings.json (each skips if one already exists)
5. Start all services via Docker Compose

After it finishes, open http://localhost:3000 for the chat UI.

Services

• OpenWebUI → http://localhost:3000
• LiteLLM API → http://localhost:4000/v1 (OpenAI-compatible)
• LiteLLM Admin UI → http://localhost:4000/ui

Configuration

• LiteLLM model routing: config/litellm/config.yaml
• Codex model provider: config/codex/config.toml
• Claude Code model provider: config/claude/settings.json
• Secrets and API keys: .env (not committed to git)

For Codex users

The installer copies a minimal config.toml to ~/.codex/config.toml that points Codex at your local LiteLLM proxy. You'll need:

export LITELLM_API_KEY="<your LITELLM_MASTER_KEY from .env>"

Add that to your shell profile. If you already have a Codex config, the installer won't overwrite it — merge the [model_providers.litellm] section manually.

Launch Codex with its configured default model:

codex

Or select any model exposed by LiteLLM without changing your default:

codex -m z-ai/glm-5.3-flash

The value passed to -m must match the model_name in config/litellm/config.yaml. Models are added there; config/codex/config.toml only defines the provider connection and your default model.

For Claude Code users

The installer copies a sample settings.json to ~/.claude/settings.json that points Claude Code at your local LiteLLM gateway. The sample sets:

• ANTHROPIC_BASE_URL → http://localhost:4000 (LiteLLM serves the Anthropic-compatible /v1/messages endpoint)
• ANTHROPIC_AUTH_TOKEN → a placeholder for your LITELLM_MASTER_KEY from .env — replace it after install:

export ANTHROPIC_AUTH_TOKEN="<your LITELLM_MASTER_KEY from .env>"

• ANTHROPIC_MODEL / ANTHROPIC_SMALL_FAST_MODEL → model names matching model_name entries in config/litellm/config.yaml (defaults: gpt-5.6-luna and glm-5p3-flash).

If you already have a Claude Code config, the installer won't overwrite it — merge the env section from config/claude/settings.json manually. To change models, update the names in that file; no gateway change is needed since LiteLLM serves any routed model.

Stopping / restarting

docker compose down      # stop services
docker compose up -d    # restart
docker compose logs -f  # follow logs