Localforge Prompt Editor

Introduction

Localforge is an agentic, autonomous LLM toolkit that runs on your own machine. If you have used Claude Code or Codex you already know the basic idea, but Localforge adds a real GUI, deeper access to the file-system, and a far wider choice of models. Out of the box it talks to more than a hundred models across fourteen API families, from GPT-4.x and Claude to Vertex.ai and a local Ollama build. The code is MIT-licensed and lives on GitHub.

The Developer's Evening Saver

Power is nice, but the part that actually saves a developer's evening is the prompt editor.

Until now the chat pane was a plain textarea. Anyone writing non-trivial prompts knows the routine: open Sublime or VS Code, shape the message there, copy–paste it back, tweak, repeat. That dance is fine once; it is brutal the tenth time you are hunting a stray instruction that flips the model into the wrong mode.

The Block-Based Approach

The new prompt editor removes the dance. Think of it as a no-nonsense VS Code, but for prompts instead of source files. A prompt becomes a set of stackable blocks—Context, Task, Instructions, Few-shot examples, Primers. Drag blocks to reorder, toggle them on or off, keep experimental variants side by side. Because each block is isolated, you can change a single assumption without rewriting the rest and see immediately what moves the needle.

Conversation Inspector

See what context is being sent to model each time

That granularity pairs well with Localforge's conversation inspector. When a reply comes back sideways, I usually suspect the system prompt. My current manual fix is to paste the whole transcript into GPT-4-o and ask why the answer diverged. The model points out a contradiction or a vague instruction and I patch it. The plan is to turn that workflow into a single right-click: pick any message, hit "Why did it say that," get a diagnosis plus concrete edits.

On the Horizon

Cost and Token Estimator

A cost and token estimator is next. Localforge already tracks usage per provider; wiring that data into the editor will let you see an estimated bill and token count while you type, so the first clue you get about runaway context length is not next month's charge.

Prompt Marketplace

Longer term, the editor will power a prompt marketplace. Different models reward different strategies—what keeps GPT-4.1 laser-focused may confuse Claude 3.7 Sonnet—and Localforge runs three discrete model tiers at once. Shipping and sharing role presets will be simpler when they are just JSON exports from the same editor you already trust.

"Different models reward different strategies—what keeps GPT-4.1 laser-focused may confuse Claude 3.7 Sonnet—and Localforge runs three discrete model tiers at once."

Getting Started

The feature ships today. Install Localforge through npm and get started with a simple command. Signed DMG and Windows bundles will follow as soon as the notarization pipeline is stable. Feedback, bug reports, and pull requests are welcome.

Once installed, just type localforge in any terminal to launch

Published 25 Apr 2025