ENZO is a self-hosted AI workspace and a "bring your own key (BYOK)" tool. Its core feature is that it lets you use your own API keys from various platforms (such as Groq, OpenRouter, NVIDIA, Hugging Face, and Google AI) to run chats, agents, and tools, with the server not taking a cut and not profiting from subscriptions. It officially states that it is MIT-licensed, free forever, with no ENZO account, no usage meter, and no subscription. The request path is: browser → ENZO → the model provider you choose, and you pay the provider's normal price.
In terms of main capabilities, ENZO provides chat with more than 300 models and lets you build agents that can write their own operating manuals; these agents can conduct research and generate and execute code. The agent builder uses a two-stage approach: the first pass analyzes the task domain and derives the tacit knowledge, decision heuristics, and edge cases that the manual needs to cover; when the draft needs a model, it simultaneously launches about 10 free candidates from the providers you have connected and lets the fastest responder win, aborting the rest, and uses a "brain health" scoreboard to adjust the order of subsequent competitions based on how models actually perform; each agent records the model that actually drafted it, and clearly states when no model can be reached. Each agent also has a neural layer that integrates domain-related platform activity on a 90-second cycle, distills experience into memory, and injects a live NEURAL FOCUS block on every run. In addition, it provides six interfaces within the workspace, agents, skills, and tools (including Gmail and Calendar), as well as in-app file conversion, a music player, an equalizer, a model marketplace, and other features.
In terms of usage, ENZO supports multiple deployment paths. One is to run it on Docker Desktop (macOS / Windows WSL 2 / Linux), requiring at least 4 GB RAM to be allocated, and then visit http://localhost:5001，点击 Just log in and paste the key from any provider. The second is running with zero installation via Google Colab; the notebook clones the repository, installs dependencies, builds the UI, and starts the server, finally providing a browser-usable link as well as a Cloudflare tunnel link that can be used on a phone or other device; the notebook also has a built-in anti-disconnect mechanism (resetting the idle timer every 60 seconds, with a watchdog pinging the server every 5 minutes and restarting it if it crashes), and the free Colab session limit is about 12 hours. Keys are stored encrypted in the browser with AES-256-GCM (non-exportable WebCrypto keys), and the optional passphrase mode uses PBKDF2-SHA256 (600,000 iterations) to reseal and delete the device key; on a new self-hosted instance, the first key that passes live verification will "claim" that instance, writing it into the container .env and sealing it into the enzo-memory volume, thereby unlocking all server-side features (agents, skills, memory) and preserving them across restarts. Projects, learned skills, and agent memory are stored in named Docker volumes and remain after upgrades. The update command is docker compose pull && docker compose up -d.
Use cases include: individual developers or teams who want full control over keys and infrastructure; users who need a self-hosted alternative to multi-provider chat, a local RAG workspace, or a visual workflow builder; users who want agents to continuously learn from their own activity and run on a schedule; and users doing research, generating, and running code on Google Colab.
Limitations and security information: official documented threat model (docs/SECURITY.md) explains that self-hosted mode stores the first key you claim in the container .env (sealed in the memory volume) so that scheduled tasks can still run after the browser is closed, and this tradeoff is explicitly documented. Each push runs a 44-assertion black-box penetration test, including authentication bypass, malicious payloads, IDOR, stream integrity, and keyless startup proof. v1.6.0 adds features such as project writes requiring a valid vault session and creator device identity, generated backends running sandboxed under the Node runtime permission model, save endpoints rate-limited by IP (60/minute), a redesigned developer tools preview panel, and in-app release notifications; it also fixes the cd ENZO case-sensitivity issue that could block the 10-second quick start on Linux. Previous versions also added in-chat file conversion, a music player, a 5-band Web Audio equalizer, a redesigned marketplace, an NYC Subway theme, terminal health ECG, onboarding, ambient weather cards, and more. Other features, fees, prices, platforms, or effects not mentioned in the materials are not yet provided.
