REAL-TIME WEB SEARCH · 65 MODELS · ONE THREAD

Ask the web. Keep the receipts.

Live web search with numbered citations, built into a workspace that moves between Claude, ChatGPT, Gemini, DeepSeek and 60+ more models without losing the thread.

Open the workspace See how search works

65models

14providers

4search backends

0demo replies

Live sources, not guesses Bring your own key Local history by default New: effects drawn in code

Eltes/Model workspace

Local-first · BYOK · Web search

CHOOSE A MODEL

66 models · 14 providersBrowse all

ANTHROPICClaude Sonnet 5.5

BALANCED

Anthropic Messages API with real token streaming. Requires an Anthropic API key.

APIMessages

Context200K

ToolsSupported

01Your keyStored in this browser

02Provider APIDirect or trusted proxy

03Your historySaved on this device

Selection preview · no request is sentOpen in workspace

ONE CONVERSATIONYOUR MODEL CHOICEYOUR DATA, YOUR TERMS

Real APIs, real keys, no demo mode

Live web search

Answers with
receipts.

Turn on the Web chip and Eltes searches the live web through your own search key — Brave, Tavily, Serper or Exa. You get a written answer, numbered inline citations and every source link on the page.

What is WebGPU and which browsers support it? Web on

Searching the web Brave · past week 1.4 s

Understanding the question18 ms
Searching the web · “WebGPU browser support”720 ms
Read 4 sourcesmdn · chrome · w3.org
Writing an answer with citations6 ms

Eltes cited answer

WebGPU is a browser API that exposes modern GPU features — compute shaders and rendering — to JavaScript 1 2. It ships in current Chrome, Edge and Safari releases, with Firefox support rolling out behind flags 3. Check the browser support table before relying on it in production 4.

1 WebGPU API — MDN Web Docsdeveloper.mozilla.orgInterface for accessing GPU hardware from web pages, with usage examples. 2 WebGPU — W3C Working Draftw3.orgThe specification defining the JavaScript API, WGSL and validation rules. 3 WebGPU arrives in Chromedeveloper.chrome.comPlatform notes on availability, migration and feature status. 4 WebGPU browser supportcaniuse.comVersion-by-version support table across desktop and mobile browsers.

Interface illustration · real searches use your own key and return live links Try a live search

Numbered, clickable citations

Every [1] in an answer links to the page it refers to. A numbered source list sits underneath with the provider's own snippet.

Live thinking, not theatre

The steps above the answer are recorded as they happen — the query, the provider, how many results returned and the time each step took.

Your index, your key

Pick Brave, Tavily, Serper or Exa in Settings. The search key stays in this browser and goes only through the backend you configure.

No invented sources

No key or no results means an explicit message, not a guess. Eltes never fabricates a URL, a snippet or a citation.

The model wall

One workspace.
Many ways to think.

Real provider logos, real endpoints, real streaming. Pick a model, paste a key, and the same conversation carries across labs.

One key per provider Switch mid-conversation See all 66 models

Eltes Lab

A model you train
in your own tab.

Not a demo script. A real character-level network — embeddings, a hidden layer, softmax — trained by backpropagation and Adam in plain JavaScript, then saved to this browser.

Real gradients

51,245 weights, hand-written backprop, dropout, weight decay, gradient clipping and Adam with bias correction. The loss falls because the maths ran.

Curves, not claims

Training and held-out validation loss are drawn live, with the best validation point marked. On a corpus this small the dashed line turns up — overfitting, shown rather than hidden.

Yours after one save

The checkpoint lives in localStorage. Eltes Lab-1 then appears in the workspace picker and continues your text — honestly labelled, no key, no upload.

~9 s for 1,500 steps zero network requests Open Lab Read the guide

The workspace

Built for people who
switch models.

No account walls, no model lock-in, no fake answers. A sharp interface that gets out of the way of the work.

Compare labs on the same prompt

Send a prompt, switch models, and keep the same context without losing track of which provider answered. Replies stay labeled by model.

Claude Sonnet 5.5 GPT‑6 Sol Gemini 2.5 GLM-4.6 Grok 4 DeepSeek V4.1 Qwen3 Llama 3.3 Mistral NVIDIA NIM OpenRouter

Agent tools that actually run

Native tool calling with a real loop: the model asks, Eltes executes locally and feeds the result back — with timing for every step.

available · runs in your browser

calculate(expression)
web_search(query) → sources
convert_units(value, from, to)

Your keys, your browser

Each provider key is stored locally and attached at request time. Requests go directly to the provider by default; you may opt into your own Vercel/local proxy. Provider billing remains with the provider.

this browser only

ELTES_KEYS.openai = "sk-••••9f2c"
ELTES_KEYS.zai = "••••7a1e"
// sent direct or via your trusted proxy

History that stays local

Threads, titles and pins live in local storage. Search them from the palette, export them as Markdown, delete them in one click.

Realtime sync designnow

Refactor auth middleware2h

Postgres index strategy1d

Real numbers, not vibes

Every reply records time to first token, total latency, input and output tokens, and tokens per second straight from the provider's usage payload.

Livefirst token

Measuredtotal time

Providerusage data

Failure states with answers

401, 429, CORS and offline each get a specific explanation and a retry that keeps your prompt. No silent fallbacks, no invented replies.

error contract

401 key rejected · check Settings
429 rate limited · retry in 4s
CORS blocked · use a proxy

Studio: code, run and preview in one window

A real workspace with a real shell. Open any HTML file and the preview follows it, edit and the page reloads, run npm run build and watch it land. The terminal runs locally with npm start — never on the deployed site.

Open Studio 5-step workflow

~/.studio · bash

$ npm run build
vite v6 building for production…
✓ 12 modules transformed
dist/index.html · 1.42 kB
preview reloaded

Agent mode

It calls tools.
It doesn't pretend to.

Eltes runs a real tool loop through your provider: the model requests a tool, the browser executes it, the result goes back into the conversation. Every call is timed and visible.

Local tools — calculate maths, convert units, read the clock, search your own chat history.
Live web search — numbered sources and inline [1] citations, using your own search key.
Provider-native — OpenAI tool calls and Anthropic tool_use, streamed.
Interruptible — stop mid-loop and keep every token already received.

How the tool loop works

Visible by designHow a tool call moves through EltesLOCAL EXECUTION

The model requests a tool

For example: calculate, read the clock, or search chat history.

MODEL

structured tool call

Eltes runs it in your browser

No shell access. No arbitrary terminal commands. The result is shown to you.

LOCAL

result returns to the conversation

The model continues with that result

When enabled, the provider receives the tool result to finish its answer.

TRANSPARENT

Every step is inspectableRead about the tool loop

Why Eltes

One workspace.
Your rules.

Pick the model, choose how the request is sent, and keep control of your local workspace.

Capability	Eltes	What that means
Model choice	65 catalogue entries	Hosted providers plus your own OpenAI-compatible endpoint.
Provider access	Bring your own key	Direct from the browser by default; optional trusted proxy.
Conversation continuity	Switch mid-thread	Choose a different model without discarding the chat.
History	Local by default	Export or clear it from this browser at any time.
No-key option	Eltes CAI	A small, bounded on-device helper—not a frontier LLM.
Agent tools	Visible and local	Calculator, units, clock, chat search, and cited web search; no shell access.
Usage visibility	Latency and tokens	Metrics depend on what each provider reports.
Hosted billing	Provider-controlled	Provider pricing and account limits still apply.

Pricing

BYOK is free.
Hosted plans are next.

The workspace is free. Hosted model replies require your provider API key and may be billed by that provider. CAI runs locally at no per-token cost. No Eltes hosted plan or free hosted credits are available yet.

Available now

BYOK

$0forever

Connect your provider key for real streamed replies. Eltes checks for a saved key before sending; without one, use CAI locally for free.

All 66 models, 14 providers
Agent tools and streaming
Local history and exports
No Eltes account required

Open the workspace

In design

Plans & credits

TBDpricing not announced

Hosted keys, one bill and no setup. The same workspace, without pasting anything.

Hosted model access
Usage dashboard and caps
Shared threads and teams
Waiting on billing infrastructure

In design

Team

—talk to us

For teams that want centralised keys, audit trails and SSO.

SSO and role controls
Shared prompt library
Self-hosted deployment
Design partner programme

Create an account

Pricing is a placeholder while the beta runs. Nothing is charged, and no card is ever requested. Read the pricing notes.

Accounts

Sign in with
the account you use.

Eltes connects OpenAI, Anthropic and DeepSeek directly. Sign up with the provider that already pays for your models, or keep a local profile with no account at all. The profile lives in this browser. Hosted model requests may optionally use a Vercel/local proxy that you configure; the profile itself remains in this browser.

OpenAI Anthropic DeepSeek

Connect an account

local

Create a local profile

No password. The welcome email is sent by your own Eltes server, if you run one. Provider keys stay in this browser. How accounts work

Downloads

Eltes as a real
desktop app.

A native window with the workspace, Studio, the terminal and the local API inside it. No server to install, no Node runtime to find, no browser tabs. Windows, macOS and Linux builds come from the same harness and the same site.

Eltes Desktop v3.18.0 · signed builds optional

One app, three platforms

The shell starts the bundled local server on a free loopback port, then opens the workspace in a hardened window. Chats, keys and the Studio folder stay on your machine.

Windows 10/11Eltes-Setup-3.18.0-x64.exe · also arm64; or the portable .exe with no install
macOS 12+Eltes-3.18.0-arm64.dmg for Apple silicon · x64 .dmg for Intel
LinuxEltes-3.18.0-linux-x86_64.AppImage · eltes_3.18.0_amd64.deb

Get the installers

Built by GitHub Actions on Windows, macOS and Ubuntu · how the harness works

Windows · portable zip

No installer, just files

Unzip and double-click eltes.bat. It finds Python or Node, serves the folder on 127.0.0.1:4173 and opens your browser.

Download eltes-windows.zip

Build it yourself

Two commands to an installer

Clone the repository, install once, then build the installers for the platform you are on.

npm install && npm run desktop:build

Build and signing steps

Checksums and build logs live on the release page for each tag. Unsigned Windows builds show SmartScreen’s “More info → Run anyway”; unsigned macOS builds need a right-click → Open the first time. The desktop app adds no telemetry and no update pings.

No demo mode

Connect an account.
Get a real answer.

Every hosted answer comes from a live provider you connected. Or start with Eltes CAI and get a real on-device answer before you add a key — no simulated replies, no fake latency.

Open the workspace How keys work

Runs in your browser Setup takes 30 seconds