one agent. above every model.

fabius is an autonomous agent layer that sits above the model. It gives Claude, GPT, Gemini, Mistral, DeepSeek, Llama, Qwen, Kimi, Grok, Command, and other compatible models the same operating brain: instructions, routing, fifteen coordinated skills, memory, safety and verification. The selected model remains the execution engine; it follows the fabius contract from the task to the proof.

Open the synapse console, hand fabius a task, and choose the model engine. fabius scopes the job, routes the right skills, gives the operating contract to the selected model, verifies what it returns, and records what mattered. Every run is logged and reviewable.

Yes — and it is the same agent, not a cut-down one: same router, same rules, same contracts, read off disk and handed to the model. A console has no filesystem, so it cannot read the repository you are working in and the task leaves the building; running locally fixes both. It writes nothing and runs nothing until you allow it, the working directory is a boundary it cannot escape, secrets and key files are refused outright, and even in fully autonomous mode it still stops and asks before anything that cannot be undone — a push, a production deploy, a recursive delete. When it hands you code it runs that code first, and a failing exit overrules its own review. It also audits any domain you own without an API key, and can be reached by encrypted message with no server in between.

Yes — in any language. Speak to it in the console instead of typing: it detects the language itself, so you never pick one from a list, and the transcript comes back punctuated and clean. It then reads its answer back out loud — Hebrew in a Hebrew voice. Being straight about the boundary: the mic that covers ~99 languages runs on the console’s own speech service — not on a model key — and costs about $0.0005 per audio-minute, billed to whoever runs the console; outside a console it falls back to the browser’s own speech engine — free, but that is the browser’s language coverage, not ~99. Connect it to Telegram and you can send it a voice note there too.

fabius implements eight runtime paths: Anthropic, OpenAI, Google, Mistral, Groq, Hugging Face, OpenRouter and Ollama. Custom model IDs through the two routers, plus local IDs through Ollama, extend that field to DeepSeek, Llama, Qwen, Kimi, Grok, Cohere Command and many more models exposed by the selected catalog. Availability follows the provider or local runtime you choose; the same fabius brain remains in control.

No. fabius’s brain — the fifteen skills, the routing policy, the memory — is private and cryptographically sealed (SHA-256 + a Merkle root anchored to Bitcoin). The seal is proof of authorship if anyone ever copies it.

It does not retrain or alter the base model. It adds the operating brain around it: instructions, routing, specialist skills, memory, safety and verification. The selected model follows that contract, so the same model can work more consistently and with less waste. In the benchmark — blind, judged against both a bare baseline and a generic “be concise” control — every capable tier (Fable 5, Sonnet 5, Opus 4.8) beats both controls on the four Claude models measured (the roster current at the 2026-07-01 run). Under fabius the smallest model passes 93% of real tests and factual checks vs 75.6% bare, and parameterized SQL goes 67.5% → 100%, on 20–35% less output. The fast tier’s dip on trivial one-liners remains printed as-is.

Against its own written contract. The identity is fixed in the brain: fabius is an intelligence amplification layer, so the only fair question is whether the exact same model achieves better outcomes with less waste. The Fabius Benchmark Suite runs 100 neutral, production-shaped tasks in three modes — bare model, under the shipped fabius files, and with recalled memory added — scored by two blind judges and objective per-task checks. Latest run: the fast tier climbs 25.27 → 26.00 → 26.73 of 28 on 11–14% less output; the mid tier holds 93% quality on 10–12% less output. The misses are printed in the receipt, not hidden.

That is exactly the control it was tested against — and it wins. fabius is structure, not brevity: a scope-control system that knows when to compress and when to expand. Method and caveats are in the whitepaper.

Yes — on content you own. Before anything ships, fabius runs a deterministic hygiene pass: invisible Unicode — zero-width characters, bidirectional marks, tag characters, the same character classes Trojan-Source attacks hide in — is stripped from generated text and code, and file metadata (EXIF, XMP, C2PA, document properties) is cleaned from images and documents with standard tools. Two boundaries, stated straight: statistical text watermarks live in the wording itself, so removing them means rewording — which degrades the copy — and no tool can certify that a vendor’s detector will fail afterwards, so fabius never promises that. The other direction is the point: fabius adds verifiable provenance to your own work — a content-bound seal anchored to Bitcoin — rather than erasing anyone else’s.