Chat with 300+ models and build self-drafting agents
Self-hosted AI workspace that unifies 300+ models across 9 providers, builds agents that draft their own operating manuals, and runs research
1.4.0Add to Favorites
Why it matters
ENZO lets you chat with 300+ AI models across 9 providers using your own keys, then build agents that write their own operating manuals from plain-English task descriptions. Everything runs on your infrastructure with no middleman, no subscription, and no usage meter-just direct provider billing.
Outcomes
What it gets done
Stream chat responses from 300+ models across OpenRouter, Google, NVIDIA, Groq, HuggingFace, Cloudflare and Gemini in one unified catalog
Draft agent operating manuals automatically by describing a task in plain English, with 74 bundled domain skills the agent pulls in per run
Run deep-research loops that write their own queries, read results, and decide when they're done under hard token budgets
Generate, boot, and preview coding projects with built-in code-gen mode and live health monitoring
Source
Get it from source
Spark does not host a copy of it.
Open sourceReports
Agent outcome reports
No reports yet
Overview
ENZO
ENZO is a self-hosted AI workspace that unifies 300+ models from 9 providers into a single catalog with live health-checking. It builds agents that draft their own operating manuals from plain-English descriptions, executes autonomous research loops that write queries and decide when they're done, and generates code projects with boot and preview capabilities. Keys are encrypted in-browser with passphrase-protected vaults, and the first validated key claims a fresh instance to unlock server-side features. Use ENZO when you need to compare models across providers without switching interfaces, when you want full control over API keys and infrastructure, or when you require agents that self-document their operational logic. It fits teams running deep-research workflows under hard budgets or organizations that need BYOK (bring your own key) deployments with no intermediary service taking a cut.
What it does
ENZO is a self-hosted AI workspace that consolidates 300+ models from 9 providers into a single catalog, enables agents to write their own operating manuals from plain-English task descriptions, and executes research and code-generation workflows. When you send a message, the request goes from your browser through ENZO to the provider you selected, and you pay that provider their normal price - nothing sits in between taking a cut. There is no ENZO account, no usage meter, no subscription.
When to use - and when NOT to
Use ENZO when you need to compare models across multiple providers without switching interfaces, when you want agents that self-draft their operational logic from natural language, or when you require full control over API keys and infrastructure. Use it for deep-research loops that write their own queries under hard budgets, or for code-generation projects that boot and preview in the same workspace.
Do NOT use ENZO if you need a managed service with built-in rate limits and usage tracking, or if you cannot self-host Docker containers on your own infrastructure.
Inputs and outputs
You provide API keys from supported providers (OpenRouter, Google AI Studio, NVIDIA NIM, Groq, HuggingFace, Cloudflare, Gemini), which are encrypted in your browser using a passphrase-protected vault. On a fresh self-hosted instance, the first live-validated key you paste claims the instance and unlocks server-side features (agents, skills, memory).
You receive streaming chat responses across 300+ models, agent operating manuals generated from task descriptions, research reports from autonomous query loops, and executable code projects with live previews. The workspace includes 74 bundled injectable agent skills (domain playbooks the agent loop pulls in per run).
Integrations
ENZO integrates with 9 AI providers through their native APIs:
- OpenRouter: API key authentication, many free models available
- Google AI Studio: API key authentication, generous free tier
- NVIDIA NIM: API key authentication, free credits
- Groq: API key authentication
- HuggingFace: API key authentication
- Cloudflare: API key authentication
- Gemini: API key or OAuth authentication
The unified catalog health-checks models live and displays platform cover art, download counts, licenses, and benchmarks in the marketplace interface.
Installation
Self-hosted deployment via Docker Compose:
git clone https://github.com/theguysudo/ENZO.git
cd enzo
docker compose up -d
# → http://localhost:5001
Alternatively, run on Google Colab with zero local install - the notebook clones the repo, installs dependencies, builds the UI, boots the server, and provides both a Colab link and a Cloudflare tunnel URL for access from any device. Free Colab sessions cap at approximately 12 hours; the notebook includes keep-alive and watchdog mechanisms to maintain uptime.
Who it's for
ENZO serves developers and researchers who need to evaluate models across providers without vendor lock-in, teams building autonomous agents with self-documenting logic, and organizations requiring full key custody and infrastructure control. The codebase is approximately 44,000 lines of strict TypeScript.
The workspace includes Terminal (streaming chat with live health traces), Model marketplace (unified catalog with search and platform cover art), Music player (song search with 5-band equalizer), Agent builder (task descriptions to operating manuals), Research mode (autonomous query loops under budget), and Code-gen (project writing with boot and preview).
Source README
ENZO
Quickstart · Live demo · What's inside · Security · What's new · Usage guide · Changelog
Chat with 300+ models. Build agents that write their own operating manuals. Research, generate code, run it all - on your keys, on your infrastructure. When you send a message, the request goes from your browser through ENZO to the provider you picked, and you pay that provider their normal price. Nothing sits in between taking a cut. There is no ENZO account, no usage meter, no subscription.
What's inside
| Inside ENZO | Count | What it gives you |
|---|---|---|
| Models in one catalog | 300+ | across 9 providers, health-checked live |
| Injectable agent skills | 74 bundled | domain playbooks the agent loop pulls in per run |
| Self-drafting agents | 2-pass builder | plain-English task in, operating manual out |
| CI pipeline stages | 7 | including a black-box security pentest on every push |
| Pentest assertions | 44 | auth bypass, IDOR, hostile payloads, stream integrity |
| Unit + security tests | 298 | agent, vault, crypto and model suites |
| TypeScript (strict) | ~44,000 lines | one language, strict mode throughout |
| Releases | 5 | v1.0.0 → v1.4.0, everything in the changelog |
Quickstart
git clone https://github.com/theguysudo/ENZO.git
cd enzo
docker compose up -d
# → http://localhost:5001
That's the whole install. No accounts, no mandatory env, no database server. Open the app, press Login, and pick any provider:
| Provider | What you need | Free tier |
|---|---|---|
| OpenRouter | API key | many free models |
| Google AI Studio | API key | generous free tier |
| NVIDIA NIM | API key | free credits |
| Groq, HuggingFace, Cloudflare, Gemini | API key / OAuth | varies |
Keys are saved encrypted in your browser (passphrase-protected vault, with a recovery file you can download). You can wipe them anytime from the Vault.
On a fresh self-hosted instance the first live-validated key you paste claims the instance - it's written to the container .env and sealed into the enzo-memory volume, so every server-side feature (agents, skills, memory) unlocks immediately and survives restarts. No master key to configure, no setup wizard - paste a working key and go. (Pre-seed a provider key in compose env instead if you'd rather not have the claim window at all; the threat model states this trade plainly.)
Run on Google Colab - zero install
Don't want the local hassle - or want ENZO reachable from any device? The Colab notebook does the whole setup for you: it clones this repo, installs the dependencies, builds the UI and boots the server, then hands you a URL.
- Click the button - the notebook opens in Colab (a free Google account is enough).
- Run all cells (
Runtime → Run all, orCtrl/⌘ + F9) - install + build takes ~4-5 min the first time; every cell is idempotent, so a re-run reuses what's already there instead of starting over. - Take your URL - the last cell prints two links: a Colab link that works in the browser you're already in, and a Cloudflare tunnel link that works from your phone or any other device (free, no account). The tunnel URL changes each session; the Colab link is bound to your session.
Stays up for the whole session. The notebook arms two disconnect-prevention mechanisms before handing you the URL: a keep-alive that resets Colab's ~90-minute idle timer every 60 seconds while the tab is open, and a watchdog thread that pings the server every 5 minutes and restarts it automatically if it ever dies. Free Colab caps a session at ~12 hours, so a 6-hour run fits comfortably; when a session does end, one click on Run all brings everything back.
Six surfaces, one workspace
| Surface | What it does |
|---|---|
| Terminal | Streaming chat with 300+ models - normal, thinking, research and coding modes - with a live ECG-style health trace in the toolbar that flatlines red the moment the catalog is unreachable |
| Model marketplace | One unified catalog across all 9 providers - cards now carry the platform's own cover art and a research panel with real download counts, licences and benchmarks |
| Music player | Search any song and play it in the marketplace - keyless, no YouTube API key - with a 5-band equalizer that actually re-shapes the audio |
| Agent builder | Describe a task in plain English; ENZO drafts the agent's full operating manual, and the agent keeps training itself on your activity from then on |
| Research mode | A deep-research loop that writes its own queries, reads what it finds, and decides when it's done - under hard budgets so it can't burn your key |
| Code-gen | Writes a coding project, boots it, previews it live, and tells you when it's broken |
| Vault | Every key sealed in the browser, attached per-request, wipeable in one click |
How the agent builder works
- Pass 1 - analysis. A two-pass drafter reads the domain of your task ("an agent that researches MUN country positions") and derives what the manual needs to cover: tacit knowledge, decision heuristics, edge cases.
- The race. When a draft needs a model, ~10 free candidates from your own providers fire simultaneously - the first to answer wins, stragglers are aborted, and a brain-health scoreboard reorders future races by which models actually deliver. Dead or rate-limited free-tier models can no longer collapse a draft.
- Honest provenance. Every agent records which model actually drafted it - and says so plainly when nothing was reachable.
- It doesn't stop. A per-agent neural layer folds in domain-matched platform activity on a 90-second cadence, distills lessons into memory, and injects a live NEURAL FOCUS block into every run. Watch it in the agent's Neural tab.
Why ENZO stands out
Most "AI workspaces" hold your keys, meter your usage, or need a subscription to exist. ENZO is built the other way around:
| ENZO | Hosted AI apps (ChatGPT, Poe, …) | Typical self-hosted AI tools | |
|---|---|---|---|
| Who pays the model | you, directly to the provider | the vendor (plus markup) | you |
| Where API keys live | your browser, AES-256-GCM sealed under a non-extractable key | vendor's servers | server-side env/config |
| Server reads your keys | hosted mode: never - relay-only, CI-enforced. Self-hosted: only the key you explicitly claim, for scheduled agents - stated in the threat model | yes | usually yes |
| Middleman fee | none | subscription / per-seat | none |
| Install | docker compose up -d, zero config |
none (it's hosted) | often multi-service setup |
| Agents improve themselves | neural layer learns from your activity | no | no |
| Security testing in CI | black-box pentest, 44 asserts, every push | opaque | rarely |
(Competitor column is about the category, not specific products - details vary.)
Security
The full threat model is written down - checkable, with the code that makes each claim true - in docs/SECURITY.md. The short version:
- Keys are sealed in your browser with AES-256-GCM under a non-extractable WebCrypto key. It can be used to decrypt your keys while never being copied - no JavaScript can export its bytes, ours or an attacker's. Optional passphrase mode re-seals everything under PBKDF2-SHA256 (600,000 iterations) and deletes the device key entirely.
- One module touches key storage, and CI enforces it. A pipeline stage greps the frontend for any raw
localStoragekey read and fails the build on a hit - a missed key-access site is a red build, never a production bug. - Every push runs a 44-assertion black-box pentest against a booted server - auth bypass, hostile payloads, IDOR, stream integrity - plus a keyless-boot proof: the server must start with zero provider keys. That's the BYOK guarantee, tested, not promised.
- The limits are stated up front. Self-hosted mode stores the first key you claim in the container
.env(sealed in the memory volume) so scheduled agents can run while your browser is closed - that trade is documented, not hidden. docs/SECURITY.md covers what's protected, what isn't, and why.
What's new in v1.4.0
- In-chat file converter - attach a PDF, spreadsheet, CSV, JSON, TXT or Markdown file in the terminal chat and it's parsed to real text/rows in your browser (files never leave the device; only the reasoning step uses your own key, like normal chat). The agent extracts, merges, or cross-converts - and CSV / Excel download buttons appear right on the reply. Built for research papers: "extract every table and merge into one CSV" now works end to end, and scanned PDFs report their missing text layer instead of failing.
- Run it on Google Colab - one click on the Open In Colab button (see Run on Google Colab): the notebook clones, installs, builds and boots ENZO on Google's hardware, hands you a URL for your browser plus a tunnel URL for your phone, and arms a keep-alive + watchdog so it stays up for 6+ hours.
- Self-healing Docker pulls - the container now verifies its dependencies on every start and installs anything missing before the server boots (
ENZO_AUTO_INSTALL=0to skip). A pulled image can't boot broken.
What's new in v1.3.0 (previous)
- Music player - search any song and play it straight from the marketplace, keyless (no YouTube API key, no quota): a collapsed corner pill expands into a full player card - vinyl disc hero, queue walking, shuffle/loop/like, keyboard controls. For You turns your own listening history (kept device-local) into song seeds through your own provider key.
- Real equalizer - a 5-band Web Audio EQ (bass / low-mid / mid / presence / air, ±12 dB, preamp, five presets) that genuinely re-shapes the frequency response when you opt in. Enhance starts off - normal playback is untouched. Tracks the enhancer can't stream fall back to the YouTube engine automatically; playback never breaks.
- Marketplace, redesigned - every model card now carries the platform's own cover art, brand colour on hover, live health dot + latency, and a research panel with real facts: HuggingFace download counts, licences, knowledge cutoffs, Artificial Analysis scores and a Wikipedia-backed family summary - pulled keyless from public endpoints, never guessed.
- NYC Subway theme - the workspace's new flagship backdrop: an AI-animated subway ride through a tunnel, with a handheld-camera tremble, monochrome film grade and animated recording grain added in code (the video ships clean).
- A quieter interface - the whole workspace went monochrome + a single coral accent: the terminal toggle switch rebuilt (was a 385-line component with dead animations), the weather chip moved up beside the catalog header, the nav collapses on scroll and springs back, and the top bar got a cursor-reactive dot grid.
- Previously in v1.2.0: the terminal health ECG, the onboarding stepper, ambient weather, the smoke top bar.
- Full history:
docs/CHANGELOG.md.
Usage guide
macOS
Requirements: Docker Desktop for Mac (Apple Silicon or Intel). Allocate at least 4 GB RAM in Docker Desktop → Settings → Resources (the model catalog + agents like headroom).
git clone https://github.com/theguysudo/ENZO.git
cd ENZO
docker compose up -d
Open http://localhost:5001, press Login, and paste a key from any provider (OpenRouter, Google AI Studio, NVIDIA NIM - all have free tiers; links are in the app). You're in.
Everyday commands
docker compose logs -f # follow what the server is doing
docker compose restart # bounce the app, data survives
docker compose pull && docker compose up -d # upgrade to a new release
docker compose down # stop (add -v ONLY to wipe all data)
Where your stuff lives: projects, learned skills and agent memory are in named Docker volumes (docker volume ls | grep enzo) - they survive upgrades and down. Your provider keys never touch the server: they're sealed in your browser's vault, and on a fresh install the first key you paste claims the instance for server-side features (scheduled agents).
If the app feels slow on a MacBook: the video themes are GPU-composited; on battery or an older machine, flip the Lite/Full chip (bottom-right) - it swaps video backgrounds for pure shader ones with one click.
Updating: docker compose pull && docker compose up -d. Releases are tagged at github.com/theguysudo/ENZO/releases.
Windows
Requirements: Docker Desktop for Windows with WSL 2 (Docker Desktop's installer sets this up; reboot when it asks). Give it ≥ 4 GB RAM in Settings → Resources.
In PowerShell (no clone folder needed - git comes with Docker Desktop's WSL distro, or use Git for Windows):
git clone https://github.com/theguysudo/ENZO.git
cd ENZO
docker compose up -d
Open http://localhost:5001 in your browser, press Login, paste a provider key - done.
Everyday commands
docker compose logs -f # follow the server log
docker compose restart # bounce the app
docker compose pull; docker compose up -d # upgrade to a new release
docker compose down # stop (add -v ONLY to wipe all data)
Windows notes
- If
http://localhost:5001doesn't load, check Docker Desktop is running (whale icon in the system tray), thendocker compose ps- the port is listed there. - Anti-virus software occasionally slows the first boot (image extraction). The second start is fast.
- Everything else - volumes, keys, the Lite/Full chip - works exactly as on macOS.
Linux (same everywhere)
git clone https://github.com/theguysudo/ENZO.git
cd ENZO && docker compose up -d # → http://localhost:5001
What's new in v1.2.0 (previous)
- Live terminal health ECG - the static ONLINE label is now a heart-monitor trace sweeping the terminal toolbar; reachable catalog keeps it beating, anything else freezes a red flatline.
- Onboarding, rebuilt - an animated stepper walks the three connect-provider steps (numbers morph into checkmarks, completed steps are click-back-navigable), with a liquid save switch that ticks when your key lands.
- Ambient weather card in the marketplace sidebar - one keyless IP geolocation + Open-Meteo, cached 30 minutes, degrading quietly.
- Smoke behind the glass - the top bar carries a slow violet/cyan drift (WebGL fbm, low-power context) and a ~20% slimmer silhouette.
- Previously in v1.1.0: the custom agent builder, race drafting, the neural layer, and ~20 hardening fixes.
- Full history:
docs/CHANGELOG.md.
CI
Every push to main runs the full pipeline in .github/workflows/ci.yml - 7 stages:
- Security checks - no key literals in tracked files,
.envnever committed, keys never read from rawlocalStorage, no onboarding bypasses - Backend - strict TypeScript, every imported file tracked, unit tests (298 assertions across agent, vault, crypto and model suites)
- Dependency audits - backend + frontend, fail on any high/critical vulnerability
- Black-box security pentest - 44 live assertions against a booted server: auth bypass, hostile payloads, IDOR, stream integrity
- Keyless boot proof - the server must boot with zero provider keys
- Frontend - strict TS + production build with an enforced gzipped bundle budget
- Repo hygiene - no large binaries, changelog and agent docs present
Two editions
latest / lite |
full |
|
|---|---|---|
| Homepage background | Nebula drift - animated WebGL | + 8 anime video themes |
| Workspace/terminal background | Default Particles (three.js) + the NYC Subway ride | + 9 cinematic video themes |
| Image download | ~150 MB | ~470 MB |
Both animate by default - the lite themes are GPU shaders, not static images. To get every theme:
ENZO_IMAGE=ghcr.io/theguysudo/enzo:full docker compose up -d
What's stored where
Everything you make lives in Docker named volumes, safe across upgrades:
enzo-projects- generated coding projectsenzo-skills- skills the agent learned from GitHub reposenzo-memory- the agent's durable notes about your work
Your provider keys are not in the volumes - they're browser-side (encrypted at rest with your passphrase).
Optional configuration
Everything works with zero environment variables. A few features want server-side values - put them in a .env next to docker-compose.yml:
# Extra origins allowed to call the API (comma-separated)
ENZO_CORS_ORIGINS=https://enzo.example.com
# "Connect with Cloudflare" OAuth button (optional — pasting a token works too)
CLOUDFLARE_OAUTH_CLIENT_ID=...
CLOUDFLARE_OAUTH_CLIENT_SECRET=...
# HuggingFace OAuth app for the HF onboarding step (optional — token paste works)
VITE_HF_CLIENT_ID=...
HF_CLIENT_SECRET=...
VITE_HF_CLIENT_IDonly takes effect when building the image from source (it's baked into the frontend at build time).
Building from source
docker build -t enzo:mine --build-arg THEME_VARIANT=full .
What's different from the hosted deployment
This image is generated from the same codebase that runs https://enzo-hub.duckdns.org, with exactly two feature differences:
- No Google sign-in. The hosted site offers Google OAuth as a convenience; here, login is setting your provider keys. Everything else - providers, research, coding agent, vault, memory, skills - is identical.
- Default themes (lite image). The first homepage and workspace themes run as pure WebGL/three.js so the image stays small. The
fullimage has the complete set.
Discussion
Questions & comments · 0
Sign In Sign in to leave a comment.