DSH Vibeify

Turn your AI work into a living magazine.

DSH Vibeify is an open-source, local magazine for DeepSeek Harness. It gives AI work a visual, content-first home: use DeepSeek, ChatGPT, or both; keep Chat for control; and share only the finished article you choose.

Download Vibeify 0.15.2 · Watch the 72-second walkthrough · See how it works · Get installation help

DSH Vibeify's magazine is not a public website. It turns DeepSeek Harness—a local runtime for agents, models, tools, permissions and sessions—into a creator-first streaming experience that behaves like a website. Vibe is the visual, content-first home; Chat remains underneath for detailed requests, progress, approvals, Queue, Steer and technical evidence. An optional, deliberately separate share service can publish one reviewed article at a time without exposing the local magazine or requiring the recipient to have DSH, ChatGPT, or DeepSeek.

The idea: generative AI beyond turn-by-turn chat

Most AI products begin with an empty box. The person writes a prompt, waits for one answer, and repeats. DSH Vibeify begins with something useful to explore and turns Chat work from any thread into one continuous edition. On every launch, an evergreen welcome issue explains—in light, visual articles—what the reader chose, how Vibe and Chat fit together, how to update, personalise, save and share, why DSH and open source matter, and how to make the first personal Vibe. It is magazine content, not a setup wizard, and it makes no model call. For an explicit public-content request, each complete verified card can arrive while the rest of the answer is still being made; ordinary final answers join after completion. It generates extra magazine material only when the reader asks for an update.

The one sentence to remember: open Chat and say “Make a Vibe about…” when you want a deliberately edited Vibe article or series. That explicit wording tells the active DSH lead to use Vibeify's visual, linked, complete-chunk publishing contract. An ordinary Chat answer may still join Vibe after it finishes, but it is not guaranteed to receive the richer article structure or to stream card by card while the turn is running.

flowchart LR
    OPEN["Open local DSH"] --> WELCOME["Evergreen welcome issue\nwhy · how · try it in Chat"]
    WELCOME --> VIBE["Vibe already contains a visual edition"]
    VIBE --> EXPLORE["Read · watch · save · respond"]
    RADAR["Shared public radar\ntrends · culture · technology"] --> CACHE["Local hidden reserve\nsignals · candidates · ready pages"]
    CACHE --> VIBE
    EXPLORE --> SIGNAL["A Chat request, pull-to-update, or Update click"]
    SIGNAL --> HARNESS["DSH harness\nsessions · tools · models · permissions"]
    HARNESS --> LEAD["Active lead agent\nplans and defines acceptance"]
    LEAD --> INSTANT["Local reserve\nready pages + visual short + questionnaire"]
    LEAD --> WORK["Independent specialist lanes\nquick · culture · media · deep"]
    WORK --> CHECK["Lead verifies each finished lane independently"]
    CHECK --> STREAM["Each closed editorial chunk streams immediately"]
    INSTANT --> VIBE
    STREAM --> VIBE
    STREAM --> STOP["One bounded update completes, then stops"]
    VIBE --> PREVIEW["Share one article\nprivate preview first"]
    PREVIEW -->|"reader presses Publish"| PUBLIC["Public article link\nno DSH account required"]
    EXPLORE -. "open when you want control or detail" .-> CHAT["Chat\nask · queue · steer · approve · inspect"]
    CHAT --> HARNESS

The visible surface is website-like, but the machinery underneath is an agent harness:

Layer	What it does
DSH	Runs sessions, agents, model routes, tools, permissions and approvals on the user's machine.
Vibeify	Adds the continuous editorial stream, visual presentation, local content buffer and safe route between generated work and the stream.
Vibe	Presents every thread's completed answers—and complete verified cards from an active public-content request—as one newest-first edition. A hidden local reserve may be prepared while Vibe is in recent use; only Pull to update, Update, a closed public-content card, or a completed Chat answer changes the visible magazine. It is not a separate public site.
Chat	Lets the user ask directly, queue or steer work, approve protected actions and inspect progress or technical evidence.
Lead agent	Plans, sets the acceptance bar, validates work and owns the final result. This is Codex in governed ChatGPT/combined mode, or the native DSH agent in DeepSeek-only mode.
Workers	Optionally perform bounded research, coding, analysis or media work; their output is not published merely because they say it is finished.
Optional share service	Receives one allow-listed, reader-reviewed article only after Share is clicked, shows a private preview, and performs a public write only after a second Publish public link click. It never receives the Chat prompt, reasoning, session identity, settings, local history, or credentials.

“Streaming” here means a continuous stream of complete, formatted content items, not merely text appearing token by token. The edition is available immediately from local bundled, saved, and ready-reserve material. A deliberate pull or Update click releases ready pages immediately, including visual and questionnaire formats, then uses one foreground batch only when the reserve is short. During an explicit public-content Chat request, the first short verified card arrives before the whole series is researched and each later closed card follows independently; ordinary un-enveloped final answers join after durable completion. Neither route makes an extra AI call. Opening, scrolling, changing settings, or finishing a batch never changes the visible edition by itself.

This is the different paradigm: Chat is an available control surface, not the whole product. Each Chat thread remains request-driven and becomes idle after it answers. Vibe is the shared presentation layer across those threads, with an explicit update gesture when the reader wants more.

For a plain-English walkthrough of the two surfaces, instant local reserve, parallel agent lanes, complete-page streaming, stop conditions, provider modes, and privacy boundary, read How DSH Vibeify works.

Choose the provider setup that suits you:

• DeepSeek only: native DSH/DeepSeek runs the agent and the provider-neutral Vibe package supplies the experience.
• ChatGPT only: a ChatGPT-authenticated Codex agent leads and works without an OpenAI API key.
• Both (recommended governance mode): Codex plans, manages, verifies, integrates, and answers while DeepSeek performs eligible bounded execution.

Neither account is individually compulsory, and installation can finish before either is connected. At least one working provider is required before asking the agent to perform AI work. DeepSeek billing and ChatGPT plan limits remain separate.

The first screen is Vibe itself: a full-screen editorial feed opens with the current welcome issue, followed by a bundled content well and the reader's older saved magazine. No guide, tile, recipe, or prompt must be selected. The welcome issue repeatedly points to one useful next move: open Chat, say “Make a Vibe about…”, watch the completed visual pages enter Vibe, then preview and share a good one. A new answer made during the current visit arrives above the welcome issue. Older saved pages remain further down rather than being deleted or mistaken for today's front page. Chat remains the detailed source view with the conventional DSH conversation and technical controls.

DSH is a fast-moving developer preview. Vibeify pins the versions it has tested and includes a health check, but you should still expect upstream changes.

Start here

I am curious, but not a developer

Think of DSH as a control room for AI agents. DSH Vibeify changes that control room in three important ways:

1. The product starts with content, not a task form. Read, scroll, save, watch, listen, or answer an optional question; open Chat only when you want to make or solve something.
2. Every tile is visual. A stable locally bundled photograph is available without a model or network wait. Generated magazine pages require subject-relevant public photography, long features can carry several photographic beats, and occasional AI-assisted graphics remain explicitly labelled.
3. Provider choice is real. DeepSeek, ChatGPT, and combined Codex-led governance are separate install modes rather than one account being hidden as a requirement.

It does not switch provider mode without telling you, send your images elsewhere, approve emails or other external actions for you, or promise equivalent quality where no Codex verification step is present.

I want to try it

For the verified Apple-silicon macOS route, download the installer, unzip it, and open Install DSH Vibeify.command. It installs or updates DSH and Vibeify, lets you choose DeepSeek, ChatGPT, both, or later, runs checks, and opens DSH. It never asks for an API key.

The repository also contains friendly Windows and Linux preview installers. They use the same public-source, immutable-package, provider-choice, non-billing-check, and do-not-interrupt-active-work contract. They will remain visibly labelled preview until the complete path has passed on real Windows and representative Linux machines; a Mac-only test is not treated as cross-platform evidence. See Installation for the current downloads and Installation and help FAQ for check-only mode and private, free-chat troubleshooting.

For the developer route you need Node.js 22.19+ in the 22.x line, or Node.js 24+, and Git:

git clone https://github.com/N9-Developer-Empowerment/DSH-Vibeify.git
cd DSH-Vibeify
./scripts/install-dsh.sh --latest
./scripts/install-vibeify.sh --provider deepseek   # or chatgpt / auto
./scripts/doctor.sh
dsh web

The installer does not restart a DSH process that is already running. Finish active work first. If DSH is already open on port 3080, reuse that page instead of starting a second copy.

For screenshots, alternative profiles, DeepSeek credentials, updating, migration, and removal, read the full installation guide. For the product lifecycle before or after installation, see How it works.

What you get

• A continuous editorial website. Vibe covers the DSH work surface completely and opens directly into one long, responsive edition. There is no catalogue-to-player-to-result transition.
• A welcome issue that sells the choice again. Twenty-one visual, reader-facing panels explain the benefits and everyday use of Vibe, Chat, deliberate updates, editorial direction, local learning, imagery, saving, sharing, Queue/Steer, Think/Trajectory, connected apps, DSH, plugins, open source, provider and capability choices, permissions and safe updates. Each panel offers a concrete prompt and a direct route into Chat. The issue is rebuilt from the active plugin rather than cached, so an upgrade cannot reopen with last version's onboarding.
• A deep local well. Twenty-four deterministic editorial chunks—text, credited photography, AI-assisted graphics, and questionnaire cards—are available synchronously on a cold visit. Every tile receives one of twelve bundled visuals with varied crop and treatment, later images decode lazily, and up to 160 completed magazine items persist for return visits.
• Instant, bounded updates. Pull down from the top of Vibe—including a two-finger pull on a Mac trackpad—or choose Update. Ready pages are released from the hidden reserve synchronously; a local visual short and questionnaire ensure an immediate response even when the reserve is empty. If fewer than four ready pages exist, one bounded foreground batch fills the gap and streams complete pages independently. Stop update cancels only that dedicated foreground turn, and a 20-minute ceiling stops a stuck update.
• A shared radar, private editing. A transparent GitHub Action publishes a public, content-only trend catalogue every 30 minutes. It contains headlines, links, regions and broad tribe hints—never reader history. Each installation keeps its selected tribes, free-text editor note, interactions, questionnaire answers, candidates, ready pages and spending ledger in that browser.
• A deliberately overstocked reserve. While the browser is open, Vibe was used within 24 hours, background editing is enabled, and the local ready reserve is below target, a separate hidden editor may prepare 6–8 pages. Combined mode asks bounded DeepSeek Flash workers to do most discovery and drafting while Codex verifies; native mode keeps its distinct no-Codex-verification boundary. The conservative reservation ledger hard-stops at the chosen daily maximum, never above US $2.
• Newest-first, append-only depth. Fresh chunks join the top of the same edition while storage remains append-only. They never replace earlier material; deeper research becomes a visible editorial follow-up above the earlier page.
• Questionnaires are content. One-tap cards create useful engagement time and softly shape later selection. Saves, opens, plays and deliberate skips also teach the local editor; the learning record stays on-device and has a one-click reset.
• One magazine across every real Chat thread. Vibe passively follows ordinary reader sessions for complete verified chat- cards, then reads durable DSH history as the fallback and for un-enveloped final answers. It never reopens or resumes a thread. Internal subagents, the dedicated foreground update and the hidden background editor are never mistaken for reader conversations. Transport envelopes are removed during cache migration; card titles are not repeated in their bodies, and Markdown tables remain structured. Raw prompts, half-written envelopes, worker reports, candidate lists, attachments, reasoning, tool activity, approvals, session identifiers, aborted work and incomplete progress are never copied.
• No empty wait or hidden work. Bundled and saved content render in the first local frame. A content-free ledger measures first frame, restore, explicit updates, chunk arrival, and engagement. Simply reading the magazine consumes no model quota.
• Real photography, labelled AI graphics. Six locally bundled photographs and six labelled local AI-assisted SVG compositions provide a zero-wait offline fallback; they are not the live editorial catalogue. Every generated non-questionnaire page begins with verified subject-relevant public photography, while features above 500 words may show two or three separately credited photographs. Each batch considers at least 18 potential images from three or more credible source families, ranks exact subject/entity relevance, informative value, credit clarity, composition, freshness and recent-use diversity, and publishes only the best choices. AI graphics are occasional, story-specific and labelled—never generic documentary substitutes. The browser accepts reviewed catalogue hosts or a direct first-party image paired with a separate official page on the exact same HTTPS host; arbitrary third-party images stay blocked. It remembers 80 recently used remote image URLs, suppresses refe