Local — PHP / Ollama / LM Studio

About 850 wordsAbout 3 min

Run Spora on your laptop. No Docker, no VPS, no shared host. The skeleton’s composer dev command starts a PHP server + Vite HMR; you can also use the canonical FrankenPHP image via Docker — single container.

Two paths for the LLM:

  1. Hosted LLM (Anthropic Claude, OpenAI GPT) — point an LLM driver config at the hosted endpoint
  2. Local LLM via Ollama or LM Studio — for offline use, privacy, or to avoid per-token costs

This page covers the local LLM path. The PHP server path is the same as the standard operator install.

Prerequisites

  • PHP 8.4+ with pdo_mysql, mbstring, zip, json extensions
  • Composer (latest 2.x)
  • Node.js 18+ and npm (only for the dev server’s Vite HMR; not needed for the built php bin/spora serve)
  • Ollama (ollama.com) OR LM Studio (lmstudio.ai) — only needed for the local LLM path

Path A — Pure PHP server (hosted LLM)

The skeleton’s composer dev starts two processes:

  • PHP server (via php -S localhost:8080 -t public/)
  • Vite dev server (proxies /api/v1 to PHP, hot-reloads the SPA)
composer create-project spora-ai/spora my-spora
cd my-spora
composer install
cp .env.example .env
# Edit .env: set SPORA_SECRET_KEY, leave SPORA_DB_DRIVER=sqlite
php bin/spora spora:install
php bin/spora db:seed   # creates a sample admin user (see output for credentials)
composer dev

The site is at http://localhost:8080. The SPA hot-reloads on any source change.

For a non-dev mode (closer to production), use FrankenPHP:

docker run -d -p 8080:8080 -v $PWD:/app ghcr.io/<org>/<repo>:latest

Or use the Docker — single container path.

Path B — Local LLM with Ollama

Ollama is a single-binary LLM runner. It serves an OpenAI-compatible API at http://localhost:11434/v1, so Spora’s OpenAICompatibleDriver works against it with no code changes.

Install Ollama

# macOS / Linux
curl -fsSL https://ollama.com/install.sh | sh

# Pull a model (a few GB)
ollama pull llama3.2:3b
ollama pull qwen2.5-coder:7b   # good for tool-calling

Configure Spora

In the admin UI:

  1. Go to Settings → LLM drivers
  2. Add a new LLM config:
    • Driver: OpenAI Compatible
    • Name: Ollama — Llama 3.2 3B (or whatever)
    • Base URL: http://localhost:11434/v1
    • API Key: (any value — Ollama ignores it, but the field is required)
    • Model: llama3.2:3b
    • Context window: 8192
    • Max output tokens: 4096
  3. Save

Then attach this driver to an agent in Agents → [agent] → LLM config.

Try it

In the admin UI’s chat view, send a message. The agent will use Ollama locally — no API calls leave your machine, no per-token cost, fully offline.

Models that work well

Ollama supports many open-weights models. The ones that handle tool-calling reliably (essential for Spora agents):

  • llama3.2:3b, llama3.2:8b — small, fast
  • qwen2.5-coder:7b — good at code and tool calls
  • mistral:7b, mistral-nemo:12b — good general-purpose
  • deepseek-coder-v2:16b — best for coding tasks

Larger models (32B+) are smarter but need more RAM. 7B models fit comfortably on 16 GB Macs.

Path C — Local LLM with LM Studio

LM Studio is a GUI app for running LLMs locally. It also serves an OpenAI-compatible API at http://localhost:1234/v1.

Install LM Studio

Download from lmstudio.ai. The app runs on macOS, Windows, and Linux. Once installed:

  1. Search for a model in the LM Studio library (e.g. “Qwen2.5 Coder 7B Q4_K_M”)
  2. Click “Download”
  3. Go to the “Developer” tab
  4. Start the local server (default port 1234, OpenAI-compatible)

Configure Spora

Same as Ollama:

  • Base URL: http://localhost:1234/v1
  • API Key: (any value)
  • Model: the model name shown in LM Studio (e.g. qwen2.5-coder-7b-instruct)

Path D — Both: PHP server + local LLM

The default composer dev is the dev loop. For a more production-like local server, use the Docker single-container and set the LLM to point at Ollama / LM Studio:

# docker-compose.yml
services:
  spora:
    image: ghcr.io/<org>/<repo>:latest
    ports:
      - '8080:80'
    environment:
      SPORA_APP_ENV: development
      SPORA_DB_DRIVER: sqlite
      SPORA_SECRET_KEY: ${SPORA_SECRET_KEY:?required}
    volumes:
      - spora_storage:/app/storage

  # No mariadb / phpmyadmin — SQLite + local LLM
volumes:
  spora_storage:

Start Ollama separately, then in the admin UI add a new LLM config pointing at http://host.docker.internal:11434/v1 (the special Docker hostname for the host machine).

Performance expectations

Local LLMs are 5-20× slower than hosted Claude/GPT-4 for tool-calling tasks. Expectations:

  • A 3B model: 5-15 seconds per tool call
  • A 7B model: 10-30 seconds per tool call
  • A 12B+ model: 30-60 seconds per tool call

For exploratory dev and demos this is fine. For production agent workloads, hosted Claude or GPT-4 is dramatically faster.

What’s next