The first public release. The repository opened today; this is what is in the box since v0.5.0.
The customs declaration
Before any message goes to a remote provider, the web UI shows exactly what would leave your machine — the system prompt, the memories it would inject, the retrieved excerpts, the prior messages, your message — with sizes, and waits. Send, cancel, or stop asking for that provider (revocable in Settings). Local endpoints never ask: nothing leaves. Every answer from a remote provider carries a receipt of how much left and to which host. POST /api/chat/preview returns the same manifest the real request is built from, and a test proves the two are byte-identical. No PII classifier, by design: you read the bytes. (ADR-26)
FreeToken as a recognized local engine
Frontier-class sparse mixture-of-experts models — gpt-oss-20b and 120b, Gemma 4 26B-A4B, Qwen3.6-35B-A3B — on one consumer NVIDIA GPU plus host RAM. sovereign doctor notices a running engine, Settings → Providers enables it, FREETOKEN_BASE_URL wires it in Docker, and the starter shelf sizes the tier against your RAM and your GPU (a zero-dependency probe: nvidia-smi, then Linux sysfs; SOVEREIGN_HARDWARE_PROBE=off to disable). A model's reasoning is shown live in chat and never stored. (ADR-25)
The starter shelf, refreshed and dated
Qwen3.8-27B leads the reasoning group; Nemotron 3.5 Lightning 30B-A3B joins the frontier tier as the first sparse model Ollama pulls directly; LFM2.5-2.6B replaces LFM2-1.2B. Every entry carries its license. Under the shelf, the sizing rule is stated as the rule of thumb it is, with links to independent measured numbers. memory.extractionModel — the cognition role — lets a small local model own memory-writing while chat uses anything. (ADR-23)
Open-weights rails
sovereign byoc gpu serve <runpod|vastai> --model <hf-id> rents a GPU running vLLM's OpenAI-compatible server with the weights you choose and wires it in as a provider (tested against provider APIs, not yet on a live bill). Every answer from a local Ollama model carries a weight-digest receipt. The Hugging Face browser shows each repository's declared license at the point of choice. (ADR-22)
Cognition sovereignty and the public ledger
Machine-written memories name their authoring model. The "cognition stays home" switch restricts memory-writing to local endpoints even when chat uses a remote provider. The Sovereignty Ledger — every layer, its status, its compromise, your exit — is public at https://mysovereign.ai/sovereignty. (ADR-20)
Life Import rail #1: email
sovereign import-email your.mbox scans a mail archive on your machine — no model calls, bodies never stored, no logins — into evidence-backed life records, and the Mind view answers what you pay for monthly and what renews in the next 90 days. (ADR-21)
Security
The August 2026 audit (docs/SECURITY_AUDIT_2026-08.md) with every recommendation resolved: DNS-resolving SSRF guard, ReDoS fixes in the PDF and mbox parsers, forwarding-header refusal in tokenless mode, an explicit extension-origin allowlist, API-key redaction in provider errors, scrypt bounds. SECURITY.md says how to report.
Builds
The container image is now one manifest for linux/amd64 and linux/arm64 (DGX Spark, Apple-silicon Docker hosts). Single binaries for Windows x64, macOS arm64, Linux x64, and — new — Linux arm64. SHA256SUMS.txt covers every asset (and has been attached retroactively to v0.5.0).
Also
The public site at https://mysovereign.ai: the playground, the clickable command center, a blog with a launch series and four hardware guides, and the "Supported today" shelf on the front page. Integration manifests (VS Code, browser, JetBrains, ChatGPT Actions) move to 0.6.0 in lockstep.
Not sovereign yet, stated on purpose: the database is not encrypted at rest (use full-disk encryption); the model layer is borrowed; rented GPUs are tenancy. The first-run wizard does not yet offer FreeToken (Settings → Providers does) — issue #5 is open for it.
Fair source, FSL-1.1-MIT: self-host, fork, and modify freely; this release becomes plain MIT on its second anniversary.