elffuss-code / README.md
KikoCis's picture
Sync with code.elffuss.com (fc5b0ba): LFM3 on our own engine, installable app, VS Code extensions, new look; README refreshed
4bd8362 verified
|
Raw History Blame Contribute Delete
5.94 kB
metadata
title: Elffuss Code
emoji: 🧑‍💻
colorFrom: purple
colorTo: pink
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
short_description: A VS Code-style IDE with a local AI agent, in-browser
tags:
  - webgpu
  - local-llm
  - agent
  - code
  - lfm
  - transformers.js

🧑‍💻 Elffuss Code: put your computer to code with you

Open a folder from your disk and a full web IDE unfolds around it: Monaco (the VS Code editor), a file tree, tabs, and an AI agent that reads, searches and edits your project for real. The model runs on your GPU through WebGPU. Your code never leaves your machine. Free, no account, no API key.

⚠️ Open it full-screen (direct link). The File System Access API (opening a local folder) is blocked inside the embedded iframe.

demo

In the demo a small model running entirely in the browser reads calc.py, finds a real bug and fixes it with a minimal edit: genuine tool calls, not autocomplete.

What's new

  • Our own engine. The default brain is LFM3 1.7B running on Elffuss Engine, a WebGPU runtime we wrote (source). It is also the default on phones, in a smaller build. Gemma-4 (LiteRT-LM) and Qwen3.5 (ONNX) are still in the model picker.
  • Install it as an app. Code offers to install itself (PWA), so it opens in its own window.
  • VS Code extensions. Search Open VSX from the Extensions panel and install color themes, grammars, snippets and language settings, checked against the SHA-256 Open VSX publishes. Extensions that run code need a sandbox on a separate site, so they work on code.elffuss.com and not in this Space.
  • A new look, a project city view, search fixed in the header, and a console that is never covered.
  • Says what's wrong. If WebGPU exists but the GPU is blocked right now, Settings says so instead of claiming your browser has no WebGPU.

Why it's different

  • Your code stays local. The folder is opened through the File System Access API: no upload, no account.
  • Real tool calling: code.tree, code.read, code.search, code.edit, code.write. You watch every call.
  • Edits, not rewrites: the agent patches the lines that change, even in files of several megabytes.
  • Long tool results don't break the conversation: they are compacted to fit the model's context.
  • Context retrieval that works beyond English, accents and Cyrillic included (measurements below).
  • Real editor: Monaco, file tree, tabs, Ctrl+S, and the git branch and last commit read straight from .git.
  • Zero install, zero build. Vanilla ES modules.
  • Your language: the whole IDE follows your browser's language (English, Spanish and more).
  • External providers (OpenAI, Anthropic, Ollama) are opt-in; keys stay in your browser.
  • Optional local bridge for real node / npm / python runs, downloaded from the main site with its SHA-256.

Chrome or Edge recommended (WebGPU and File System Access).

How it's built

Write-ups on the whole stack (model, runtime, agent harness, editor and file access, all in one tab): https://bitacora.utopiaia.com/#elffuss

Context management: how it's measured

The context packer isn't chosen by intuition. Numbers measured at the same token budget:

Real agent sessions (25 sessions, 174 probes, fact recall, no LLM judge):

recall
tail truncation 7.0 %
previous version 15.1 %
current 65.3 % ± 8.5

Long-term memory (200 questions, F1 from the benchmark's own repo, a real model answering, ~8 % of the context):

F1
the FULL context, uncompressed 22.56
tail truncation 6.62
hand-written heuristics 6.07
BM25 without IDF 20.34
BM25 23.59
embeddings 23.95
rank fusion (BM25 + embeddings) 28.09

Retrieving well beats having everything: 28.09 vs 22.56 with 8 % of the tokens. Irrelevant context isn't neutral ballast; it distracts.

Two more things that were measured, not assumed: turning the heuristics on cost −3.02 F1, and removing IDF costs −3.25. Semantic search doesn't replace lexical search, it covers its blind spot: with no vocabulary overlap between question and answer, embeddings win (24.28 vs 18.60); with overlap, BM25 wins (29.33 vs 23.58), and fusion keeps both.

References: BM25 (Robertson et al., TREC-3 1994) · IDF (Spärck Jones, 1972) · RRF (Cormack et al., SIGIR 2009) · MMR (Carbonell & Goldstein, SIGIR 1998) · attention sinks (Xiao et al., ICLR 2024) · valid time (Snodgrass & Ahn, SIGMOD 1985).

The full story, including what went wrong: https://bitacora.utopiaia.com/posts/16-beaten-by-doing-nothing.html

The Elffuss family