Notice
What Onyx is, what it is not, and the terms you accept by downloading or using it.
Onyx is independent software built by TWAAI. It is not affiliated with, endorsed by, sponsored by, or connected to any AI provider — including but not limited to OpenAI, Anthropic, Google, xAI, Mistral, or OpenRouter. Provider and product names are used only to describe compatibility.
Onyx ships with no API keys and makes no provider calls on your behalf. You supply your own key for the provider you choose. You are the customer of record for that provider and are responsible for your account, your usage, and any charges it incurs.
Onyx routes the requests you write to the provider you select. How you use a model — and whether that use complies with your provider’s terms of service and your local law — is your responsibility. Onyx is a client; it does not change your obligations to any provider.
Onyx is provided “as is”, without warranty of any kind. There is no guarantee of fitness for a particular purpose, of uninterrupted availability, or of any particular result. You run it at your own discretion.
Builds are unsigned. Use SHA256SUMS-macOS.txt for macOS and SHA256SUMS.txt for Windows and Linux; compare the matching list before running anything. Verify with certutil -hashfile <file> SHA256 on Windows, or shasum -a 256 <file> on macOS and Linux.
Onyx keeps keys, settings, vaults, and encrypted chat history on your machine, under ~/.onyx/. There is no telemetry and nothing is transmitted to TWAAI. Outbound connections go only to the provider endpoint you configure.
TWAAI Patreon memberships are voluntary recurring support. They are separate from the one-time price of any Onyx build. Supporting on Patreon does not change your rights to the software.
What Onyx is
In plain English: Onyx sits between you and the AI model. You choose the provider and model; Onyx supplies the persona, remembers the conversation, detects common refusal patterns, and can retry with a cleaner route when a model shuts down too early.
Onyx Lite 1.2 is the current streamlined chat app over the Obsidian engine. Onyx V4 Obsidian is the complete desktop suite: Onyx chat and Forge prompt drafting together on one encrypted core. Original Onyx Lite 1 is also listed; its release artifact identity is not verified.
The focused chat app for the Obsidian engine. Version 1.2 adds bounded file attachments, clearer request routing, stronger completion checks, safer log sharing, and better recovery progress while keeping its own window, chats, model pin, and settings.
The complete product. Onyx handles conversation while Forge drafts targeted system prompts; both use one provider layer, encrypted prompt store, and per-target memory.
Current releases
A focused chat app or the complete Onyx + Forge suite.
Choose your product and operating system. Onyx Lite 1.2 is £20 one-time; other product pricing is to be announced.
Checkout coming soon — checking service configuration.
£20.00
One-time purchase
Unsigned, self-contained release. Checkout is coming soon. Bring your own API key; provider usage is billed separately.
Buy once, keep using your version. Same-major updates are included; a future major upgrade may be an optional paid purchase. Patreon support and AI provider charges are separate. Checkout availability is shown above; no live sales before service setup. Full policy in Notice FAQ.
The original Lite established the separate lightweight chat window. Lite 1.2 keeps that simple shape and improves the work around it. Obsidian remains the larger suite for people who also need Forge and its full workspace.
| Compare products | Onyx Lite 1.2 | Onyx V4 Obsidian |
|---|---|---|
| What it is | Focused chat app over the Obsidian engine | Complete Onyx + Forge desktop suite |
| Onyx chat | Dedicated chat window | Dedicated Onyx room |
| Engine relationship | Uses the same Obsidian persona, providers, HOLD logic, and vault | Owns the shared core used by both products |
| Chats and settings | Separate Lite chats, model pin, and configuration | Separate Obsidian chats, model pin, and configuration |
| Vault and API keys | Shared with Obsidian | Shared with Lite |
| Forge drafting | Not in the Lite interface | Targeted system-prompt drafting |
| Encrypted history | Local Lite conversation store | Vault-passphrase encrypted conversations |
| Pricing | £20 one-time | Pricing to be announced |
Onyx Lab
Onyx Lab documents the research behind the refusal-dampening layer. Its first paper, Bandwork, studies a pipeline where planning, isolated generation, validation, and deterministic assembly are measured as one system rather than treating one prompt as the whole capability.
The paper separates simple prompt scanners from models that judge a request by meaning, records model-specific behavior, and sets a stronger standard for cross-model proof: frozen versions, blinded tasks, repeated trials, complete failure logs, costs, confidence intervals, and independent replication.
A model may reject a complete request while still producing individual components. Evaluation has to examine the assembled result.
Planning, isolated model calls, validation gates, memory, and assembly can change what the overall system delivers.
Single-prompt testing misses behavior that emerges across a sequence of requests and local assembly.
Vocabulary filters and semantic refusals behave differently, so one universal retry strategy is a weak fit.
Onyx Lab treats repeatability, failure logs, cost, timing, and independent replication as part of the result.
A workflow-level study of planning, isolated generation, validation, deterministic assembly, and the evidence required to measure the resulting system.
Frequently asked
Clear answers about the products, refusal dampening, pricing, and TWAAI videos.
No. Abliteration changes a model internally by suppressing refusal-related activation directions or modifying weights. Onyx is application-layer refusal-dampening middleware: it manages personas, prompting, provider routing, refusal detection, HOLD retries, and conversation state without modifying the underlying model.
Lite 1.2 keeps the focused chat design and adds bounded document attachments, clearer chat/question/build routing, explicit completion checks for requested artifacts, warnings before shared logs expose common secret shapes, and more visible recovery progress. It still has no Forge room, Anvil interface, or terminal chrome.
Onyx V4 Obsidian is the complete desktop product. Onyx handles multi-provider conversation while Forge handles targeted system-prompt drafting, both on one provider layer, encrypted prompt store, and per-target memory.
Pay once for the latest release and receive all updates within the same major version. Lite 1.2 includes future Lite 1.x releases at no extra charge. A new major version may be an optional paid upgrade, usually discounted for existing customers. Your purchase does not expire, and you do not have to upgrade to keep using the version you bought. Onyx Lite 1.2 is £20 one-time; other product pricing is to be announced. Purchases remain coming soon until the service enables checkout. Test mode is labelled and does not deliver a live purchase.
No. Patreon is separate, optional recurring support—not a requirement to use the software. AI provider usage is also separate from the software purchase: you supply your own API key and pay any applicable provider charges directly.
Use the Rumble card in Links to open TWAAI’s channel on Rumble.
Follow TWAAI