Gemma Forge: Local AI Without the Setup Wall

title: Gemma Forge: Local AI Without the Setup Wall

published: false

description: A local Gemma 4 work harness that turns local AI into a guided, inspectable project workspace.

tags: devchallenge, gemmachallenge, gemma, localai

cover_image:

New work starts with a plain-language project seed, then moves through protocol cards instead of disappearing into a loose chat thread.

Gemma Forge also leaves receipts on disk. This workspace screenshot shows the output of a real session: project context, GSD planning, research, execution notes, extra reviews, verification, handoff files, screenshots from browsed sources, and the generated index.html artifact.

Code

Repository:

https://github.com/TheRefreshCNFT/gemma-forge

Quick start on macOS:

git clone https://github.com/TheRefreshCNFT/gemma-forge.git

cd gemma-forge

./launch_forge.command

Then open:

http://127.0.0.1:5005/

The macOS launcher is the recommended first-run path. It installs or verifies the local toolchain, starts Ollama, pulls the default source model gemma4:e4b, creates the local Forge Brain alias gemma-4-e4b-it, stages bundled skills, prepares support tools, and starts the harness.

The first-run install default is fixed so setup is predictable. It does not lock users in. After setup, users can import installed Ollama models, search Hugging Face, provision other compatible repos, and choose a different active local model from Settings.

Manual/development launch:

python3 -m venv .venv

source .venv/bin/activate

pip install -e .

gemma-forge

How I Used Gemma 4

Gemma 4 is the planning and orchestration brain inside Gemma Forge.

I chose the E4B / 4B-class lane as the default because Gemma Forge needs more reasoning headroom than a tiny model for project planning, skill routing, repair loops, and verification, while still staying realistic for local hardware. The one-command installer pulls gemma4:e4b and aliases it locally as gemma-4-e4b-it.

The harness presents the expected local footprint before provisioning: about 10 GB of disk budget and 8 GB RAM for readiness, with the current quantized Ollama artifact around 5 GB on disk.

Gemma Forge uses Gemma 4 to:

• Convert a raw user request into a structured project contract.


• Decide which protocol cards and bundled skills are relevant.


• Write project-scoped context and GSD-style phase plans.


• Generate or repair deliverables through the Project Execution card.


• Evaluate outputs against acceptance criteria and deterministic validation.


• Explain next steps when the user does not know which control to use.


• Produce handoff notes so work can resume cleanly.

Gemma Forge also records model-route proof: Forge Brain selection to Flask harness to Ollama /api/chat. That proof matters because the authenticity rule is strict. A valid result means the selected local Gemma model actually did the work through the harness workflow. Deterministic scripts, screenshots, code intelligence, and validators can verify or package the result, but they do not replace Gemma 4 doing the task.

Gemma Forge…Where local agents want to work