gemma-server

A zero-config local Gemma server for the DoppelCheck browser extension. One command from clone to a working OpenAI-compatible / Ollama endpoint at http://localhost:11434.

Two install paths. Pick the one that fits your machine.

Option A: native (Linux / macOS) - smallest footprint

Requirements: bash and curl. Nothing else.

git clone https://github.com/doppelcheck/gemma-server.git
cd gemma-server
./start.sh

If Ollama is not already installed, the script asks for confirmation and runs the official installer (Linux: curl -fsSL https://ollama.com/install.sh | sh; macOS: brew install ollama if Homebrew is present, otherwise it points you at https://ollama.com/download). Then it pulls the model and starts the server.

Press Ctrl+C to stop.

Option B: Docker (any OS, fully isolated)

Requirements: Docker Desktop or docker + docker compose.

git clone https://github.com/doppelcheck/gemma-server.git
cd gemma-server
docker compose up

Heavier (Docker itself is ~1 GB installed) but self-contained: nothing is installed on your host, the model cache lives in ./ollama_data/, and docker compose down cleans up.

Windows

Option B (Docker Desktop) is the cleanest path on Windows.

If you would rather install Ollama natively: the OllamaSetup.exe from https://ollama.com/download installs Ollama as a background service that auto-starts. To make it accept requests from the DoppelCheck extension you have to set OLLAMA_ORIGINS for that service (a plain $env:OLLAMA_ORIGINS = ... in PowerShell only affects the current window, not the service). The simplest way:

[Environment]::SetEnvironmentVariable("OLLAMA_ORIGINS",
  "chrome-extension://*,moz-extension://*", "User")
# then quit Ollama from the system tray and start it again,
# or reboot, so the service picks up the new variable.
ollama pull gemma4:e2b-it-q4_K_M

After that the model is reachable at http://localhost:11434 like on the other platforms.

What you need

• Disk: about 10 GB free (the default model is 7.2 GB on disk).
• RAM: 8 GB minimum, 16 GB comfortable.
• Network: the first run downloads the model from the Ollama registry.

The first start takes 5 to 20 minutes depending on your connection. Subsequent starts reuse the cached model and come up in seconds.

Use it from DoppelCheck

In the extension settings:

• Tier: Network
• Provider: Ollama
• Base URL: http://localhost:11434
• Model: gemma4:e2b-it-q4_K_M (or whatever you set OLLAMA_MODEL to)

Click Test connection. You should see a green check.

Use it from anything else

The OpenAI-compatible endpoint works with any client that lets you set a base URL:

curl http://localhost:11434/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{
    "model": "gemma4:e2b-it-q4_K_M",
    "messages": [{"role": "user", "content": "Say hi"}]
  }'

The native Ollama API also works:

curl http://localhost:11434/api/generate \
  -d '{"model": "gemma4:e2b-it-q4_K_M", "prompt": "Say hi", "stream": false}'

Switching models

cp .env.example .env
# edit .env and change OLLAMA_MODEL to any tag from
# https://ollama.com/library, e.g.:
#   gemma3:1b-it-q4_K_M   (~815 MB, much smaller)
#   gemma4:e4b-it-q4_K_M  (~9.6 GB, larger)
#   llama3:8b
#   qwen2.5:7b

Then restart:

• Native: stop the script (Ctrl+C), re-run ./start.sh.
• Docker: docker compose down && docker compose up -d.

The new model is pulled on the next start. Old models stay on disk until you remove them (ollama rm <tag> for the native install, or rm -rf ollama_data/ for the Docker install).

Stopping

• Native: Ctrl+C in the foreground terminal. The script handles SIGTERM / SIGINT cleanly. If you ran with nohup or similar, find the pid with pgrep -f 'ollama serve' and kill it.
• Docker: docker compose down. Add -v to remove the named volumes (does not touch the bind-mounted ./ollama_data/). To free disk: rm -rf ollama_data/.

Troubleshooting

Extension says CORS error (HTTP 403 from the daemon). Ollama's built-in default origins are localhost-only - they do not include chrome-extension://* or moz-extension://*. The daemon needs to be told about them via OLLAMA_ORIGINS. Both start.sh and the compose file set this; it only fails when something else is serving 11434.

A common cause: the official Ollama installer on Linux sets up a systemd service that runs without OLLAMA_ORIGINS, and that service auto-starts before you ever run start.sh. start.sh then detects the existing daemon and reuses it - inheriting its (insufficient) CORS settings. Two fixes:

1. Take over the port: stop the systemd-managed daemon and let start.sh own it.

   sudo systemctl stop ollama
   sudo systemctl disable ollama   # optional, prevents auto-start
   ./start.sh

2. Patch the system service: keep the systemd-managed daemon and give it the right origins.

   sudo systemctl edit ollama
   # add:
   #   [Service]
   #   Environment="OLLAMA_ORIGINS=chrome-extension://*,moz-extension://*"
   sudo systemctl restart ollama

Verify either way with a preflight from a fake extension origin:

curl -sS -o /dev/null -D - -X OPTIONS http://127.0.0.1:11434/api/chat \
  -H 'Origin: moz-extension://abc' -H 'Access-Control-Request-Method: POST' \
  -w 'STATUS=%{http_code}\n' | grep -iE '^(STATUS|access-control-allow-origin)'

Expect STATUS=204 and Access-Control-Allow-Origin: moz-extension://abc.

Pull is stuck or slow. Check disk space (df -h) and that you can reach https://registry.ollama.ai. The full model is 7.2 GB; on a 10 Mbit connection that is over an hour.

Port 11434 is already in use. Most often this is because you already have an Ollama daemon running (e.g. installed via the official installer, which sets up a systemd service). start.sh detects this and falls back to just pulling the model through the existing daemon - no second ollama serve is started. The model is then served by your existing daemon at the same http://localhost:11434. If that daemon's CORS configuration rejects extension origins (rare on Ollama 0.20+), stop it (sudo systemctl stop ollama on Linux) and re-run start.sh so this script can own the daemon and apply its own OLLAMA_ORIGINS.

If port 11434 is held by a non-ollama service, set OLLAMA_HOST_PORT=11500 (or any free port) in .env and update the extension's Base URL to match.

"failed to pull model". The tag in OLLAMA_MODEL does not exist. Check https://ollama.com/library and update .env.

Native: "ollama: command not found" after install. Open a new shell (the installer added it to PATH), or run hash -r in the current shell.

Why a separate repo?

This is a companion to the DoppelCheck extension, not part of it. The extension already supports any Ollama or OpenAI-compatible endpoint, so hosting the model elsewhere (a remote box, a different runtime, an existing Ollama install) is fully supported - you would just point the extension at that endpoint instead. This repo is for the common case where you want a local server and do not want to think about it.

License

MIT. See LICENSE.