Six months, told properly — the releases we shipped and never announced
A newsroom record of what 🤫 shipped between February and July 2026, drawn from the commit history rather than memory, with the people who did it named.

Between February and July 2026 we shipped roughly 6,600 changes across four repositories, from 65 people. We announced almost none of it.
This is the correction. Every item below was derived from the commit history, not from recollection — a script clusters the record into bounded milestones so nothing quietly disappears because no one remembered it in a meeting. Where something is a design rather than working code, it says so.
July 2026 — One Voice becomes a router, not a demo
52 new capabilities. Seven people.
The largest single month of the half-year, and the most consequential. 🤫 Agent One stopped being a voice surface with a fixed script and became something that routes: One now dispatches to any specialist by intent, rather than by a hardcoded path. Remote MCP and the developer API went live in production, matching UAT — which means an outside developer can reach the agent rails on the same contract we run ourselves. Hushh One MCP 0.1.3 shipped as a released package.
Also this month: active-screen awareness, so the agent knows what the person is looking at when they speak, and a setup watchdog that catches an onboarding flow stalling instead of leaving someone stuck.
The reason this matters more than the count suggests: an agent that routes by intent can gain a new specialist without a release. That is the difference between a product and a platform.
June 2026 — the site learns to be found, by people and by agents
26 new capabilities in discoverability. Eight people.
Two audiences, one month. For search engines: robots, sitemap, and a public-route allow-list, so what we publish is deliberate rather than incidental. For AI agents: JSON-LD @graph, FAQ schema, and llms.txt — a machine-readable index of what we do and how to reach it.
Publishing llms.txt is a small file with a real position behind it. If agents are going to act for people, they need to discover services the way people do. We would rather be legible to them on purpose than scraped by accident.
Alongside it, One Voice moved to Gemini Live on a Vertex relay, added a realtime-voice benchmark harness, and put ADK delegation behind a flag with persona-aware framing. A benchmark harness before a launch is the unglamorous ordering, and the right one.
June 2026 — 🤫 Puppy One becomes a product line
15 new capabilities. Four people.
The compute ladder took shape: Puppy One as a supercomputer a person owns, in a home or a garage, with the economics of buying versus renting written down rather than asserted. This is the north star made concrete — a distributed edge grid owned by the people hosting it, not by us.
February 2026 — Kai goes from prototype to regulated-ready
33 new capabilities. Two people.
Two engineers, one month, and the finance specialist grew up. A resumable debate task centre and run manager, so a long analysis survives a closed laptop. Portfolio import with schema preloading. Streaming, caching, and context gating. iOS flow hardening. Regulatory hardening and a regulated cutover, with consent-session ownership fixed as part of it.
We are naming the month rather than the individual commits because that is the honest unit: this was one arc of work with a beginning and an end, and it ended with Kai able to operate under the constraints a regulated context imposes.
February and July 2026 — native parity, twice
17 new capabilities in February, 14 in July. Three people, then eight.
iOS and Android are not ports here; the same contracts have to hold on every surface. February brought safe-area handling, native polish, and Apple iOS colour tokens applied without disturbing the underlying design structure. July brought a second wave, with parity reports generated rather than claimed.
July 2026 — the guards that exist because something already failed
19 new capabilities in reliability. Eight people.
The least announceable and most valuable category. A declarative agent.yaml manifest with a sync guard. Location sharing restricted to connections only. An LLM proposal confirm gate in KYC, so a model's suggestion needs a human yes. Features gated as "coming soon" rather than shipped half-working.
Every guard in this codebase exists because the thing it prevents happened once. We publish them because a company's reliability posture is visible in what it refuses to ship, not only in what it launches.
What we are not claiming
PCHP — the Personal Consent Handshake Protocol — runs in production, with a receipt behind data access.
hu_ssh is a published design, not shipped code. An earlier post of ours implied otherwise and we corrected it.
FedRAMP High and DoD Impact Level are in pursuit, not held. No 3PAO assessment, no agency ATO.
🤫 Fund A is pre-launch. No PPM, LPA, or subscription agreement; no Form D filed; the investment manager is not SEC-registered.
How this record is made
A script reads the commit history across all four repositories, scores each change on what it created rather than how many lines it touched, and clusters the result into month-bounded milestones. It also folds duplicate git identities, because the same person appears under several and crediting someone twice under two names is worse than not crediting them.
It deliberately does not write these paragraphs. Generated announcements read as generated, and the work was done by people.
The next edition covers August. It will be shorter, because it will be on time.
The 🤫 hussh magazine
Written by 🤫 hussh, and built to read beautifully here — and to travel to 🤫 One on your phone, your glasses, and visionOS, as one immersive magazine you own.