# Ultra AI — Product & Feature Overview **Ultra AI is a self-hosted, agentic AI coding platform.** It lives inside the developer's editor as a single assistant that reads a project, writes and reviews the code, and shows exactly what every request costs — while the whole platform runs on _your own server_, controlled end to end from one admin dashboard: users, models, pricing, limits, languages, releases, and security. You describe what you want — in plain English or 中文 — and Ultra AI builds it. A senior engineer gets a tireless pair-programmer; a non-technical founder gets working software from a sentence. And behind both, the operator gets a complete, self-hosted commercial platform with the margin in their hands. --- ## Why Ultra AI — at a Glance - **Zero code required.** Describe the outcome in plain language; Ultra AI plans, writes, and reviews the code. Built so a product manager or founder can drive it as well as a senior engineer. - **Token & cost optimization, engineered in.** Prompt-cache-preserving routing, self-managing context compaction, cache-hit metering, and per-request cost transparency — cost is treated as an engineering problem, not the user's problem. - **You own the economy.** Set per-token prices (input / output / cache-hit), the points-per-RMB rate, and see your profit live in a margin calculator. Decide exactly how much you earn. - **Self-hosted & secure.** Your code and data stay on your server. Enterprise sign-in, role-based admin, encrypted keys, network rules, and a full audit trail. - **Total live control.** Change limits, models, pricing, languages, and announcements while the platform runs — no redeploys — and ship extension updates on your own schedule. - **Bilingual, region-aware.** The entire experience — editor and dashboard — ships in English and 中文, with native USD/CNY pricing and China/Global routing. --- ## Contents 1. [In Your Editor — the Developer Experience](#in-your-editor--the-developer-experience) 2. [Platform & Control](#platform--control) 3. [Cost & Token Optimization](#cost--token-optimization) 4. [Security & Trust](#security--trust) 5. [What's Coming](#whats-coming) 6. [Changelog](#changelog) --- ## In Your Editor — the Developer Experience Ultra AI lives inside Visual Studio Code as a single, always-available assistant. You describe what you want — in plain English or 中文 — and Ultra AI reads your project, writes the code, shows you every change, and tells you exactly what it cost. Nothing to install beyond the extension, no API keys to manage, no configuration required to get started. > **The short version:** Ultra AI turns "I need this built" into working, reviewed code — while showing you every token it spends and squeezing the most out of each one. --- ### Chat With an Assistant That Knows Your Code Open the Ultra AI sidebar and start typing. This isn't a generic chatbot — it answers with your actual project in view. - **Ask anything about your codebase.** "How does login work here?" "Where is the price calculated?" Ultra AI searches and reads your project to answer with specifics, not guesses. - **One keystroke to get help.** Highlight any code and press a single shortcut to send it straight into the conversation. Another shortcut opens the chat full-screen in its own editor tab for bigger sessions. - **Point it at exactly the right context with `@`.** Type `@` to attach precisely what matters: individual files and folders, the current file, all open files, your file tree, a map of the repository, recent git changes and commit history, terminal output, the errors and warnings in your editor, your clipboard, a URL or web search, indexed documentation sites, GitHub issues, GitLab merge requests, Jira tickets, even a database — plus anything added through connected tool servers. - **Attach real files to the conversation.** Drop in images (logos, mockups, screenshots) and they appear as neat, compact chips — the agent then uses them as assets in your project. PDFs, Word documents, and code files can be attached as context too, with clear size and format guidance so nothing fails silently. - **Your own documentation, indexed.** Add any documentation site by URL and Ultra AI crawls and indexes it, so answers use the current docs for your framework — popular sites are pre-cached for instant setup. - **Sessions that behave like real work.** Automatic session titles, optional session tabs, full searchable history, one-click resume of your last session, and export of any conversation as a Markdown file to share with the team. --- ### Three Ways of Working: Chat, Plan, Agent A single selector switches how much freedom Ultra AI has — from "just talk to me" to "go build it." - **Chat** — answers questions and explains code. It cannot touch your files, so it's completely safe for exploration. - **Plan** — before anything is built, Ultra AI investigates with read-only access and describes the steps it would take. You review the approach first, then let it proceed. - **Agent** — full autonomy: it creates and edits files, runs commands, and carries a task from request to result. - **Effort dial (Low / Normal / Max).** Choose how hard the model thinks per task — fast and economical for simple asks, deep deliberation for hard problems. The UI even warns you that Max uses more tokens, so cost is never a surprise. - **Reasoning on demand.** Toggle model reasoning on or off, and watch its "Thinking" unfold — with the time and tokens it spent clearly labeled. --- ### The Autonomous Agent: Hand Over Whole Tasks This is where the manual work disappears. Give Ultra AI a goal — "add a dark mode," "fix the failing checkout flow," "build a settings page" — and it works the task end to end. - **Multi-file changes, done for you.** The agent reads files, creates new ones, edits existing ones (including large batched edits applied atomically), searches the project, lists directories, inspects diffs, and runs terminal commands — chaining these steps itself until the job is finished. - **It reaches beyond your project when needed.** Built-in web search and URL fetching let it pull in current information — current library versions, error explanations, API references — instead of relying on stale knowledge. - **A status feed anyone can follow.** While it works you see plain-language updates — "Reading file…", "Searching the project…", "Running a command…", "Updating…" — with details one click away. You always know what it's doing and why. - **It asks instead of guessing.** When a request is ambiguous, the agent asks a clarifying question with clickable answer options (or a free-text box), so it builds the right thing the first time. - **It remembers your project.** Important facts it learns — your conventions, decisions, project quirks — are saved as memory, so it stops asking you the same questions over and over. - **Auto-approve, your way.** A three-position control governs autonomy: **Off** (confirm every action), **Edits** (file changes apply automatically, terminal still asks), or **Full** (everything runs, after one explicit, clearly-worded confirmation — and you can switch it off anytime from the toolbar). - **Rules keep it on-brand.** Write project or global rules ("always use our logger," "tests go in /tests") and the agent follows them. Ultra AI can even **generate a rule from your chat history** — one click turns "the way we just did it" into a permanent instruction, with flexible triggering (always on, auto-attached by file pattern, agent-requested, or manual). - **Connect your own tools (MCP).** Plug in external tool servers — databases, internal APIs, ticketing systems — with live connection status, and they become tools the agent can use. - **Per-tool permission policies.** Every tool — built-in or connected — can be set to allowed, ask-first, or excluded, so autonomy always runs inside boundaries you define. - **Custom AI Actions & prompt files.** Save reusable prompts as files with their own syntax highlighting and `@`-completion, and build your own quick actions. --- ### Review Every Change Before It Lands Ultra AI never sneaks code into your project. Everything arrives as a visible, reviewable diff. - **Live diff streaming.** Proposed edits appear right in your editor as color-coded insertions and deletions while they're generated — you watch the change take shape. - **Accept or reject at any granularity.** Approve a single block, a whole file, or everything at once — by click or keyboard shortcut. A familiar "undo" gesture also rejects a pending diff. - **A pending-changes review center.** When the agent touches several files, a dedicated Changes view lists every file with its pending edits — step through them, accept all, or reject all from one place. - **Stop anytime.** A streaming edit can be halted mid-flight, and rejected changes restore your file exactly as it was. > **Coming soon — Checkpoint & Rewind:** every agent step will be snapshotted automatically so you can roll the whole project back to any earlier point with one click. This is the next feature shipping. --- ### Autocomplete & Predictive Editing Ultra AI also speeds up the typing you still do yourself. - **Tab autocomplete everywhere.** Context-aware inline suggestions in every file type — press Tab to accept. Toggle it with a shortcut, force a fresh suggestion on demand, and see its status at a glance in the status bar. - **Tuned to stay out of your way.** Control multiline behavior, response timing, and per-file exclusions (e.g., disable in generated files); it can even pause itself automatically when your laptop is on battery. - **Next Edit (early access).** Beyond completing the current line, Ultra AI predicts your _next_ change — including edits elsewhere in the file — and offers to jump you there. Accept with Tab, dismiss with Escape. --- ### Inline Edit & One-Click Actions For focused changes, you never need to leave the file you're in. - **Edit by describing.** Select code, press the edit shortcut, and say what you want — "make this async," "add input validation" — and the change streams in as a reviewable diff. - **AI chips above every function.** Lightweight ✨ Explain, 🧪 Test, 📝 Doc, 🔧 Refactor, and 🐛 Debug actions sit above each function, method, and class — the Debug action even bundles in the editor's own error diagnostics for that function. One click, no prompt writing. - **Ready-made commands for everyday chores.** Right-click or use the command palette: write comments, generate a docstring, fix this code, optimize this code, generate tests, explain a selection, explain an error — even fix grammar and spelling in text files. - **Debug your terminal in one keystroke.** A failing command? One shortcut sends the terminal output to Ultra AI and asks it to figure out what went wrong. - **Helpful nudges for newcomers.** Select code and a small hint appears showing the chat and edit shortcuts — discoverability built in (and dismissible for veterans). - **Add files as context from the explorer.** Right-click any files or folders and send their contents straight into the conversation. --- ### Slash Commands & Saved Prompts Type `/` in the chat for instant, repeatable workflows: - **/commit** — write a commit message from your staged changes. - **/cmd** — generate a shell command from a plain-language description. - **/test** — generate tests for the highlighted code. - **/doc** and **/explain** — document or explain code on demand. - **/review** — get feedback on code. - **/share** — export the session as Markdown. - **/onboard** — get oriented in an unfamiliar project. - Plus commands for drafting issues and reviewing pull requests, your own **custom prompt files**, and commands contributed by connected tool servers. --- ### Pick Your Power Level: Model Choice & Smart Routing You shouldn't need to know model names to get great results — and with Ultra AI you don't. - **Three simple tiers — Lite, Pro, Max.** Choose by need, not by model trivia: Lite for speed and economy, Max for maximum capability, Pro in between. Behind each tier, Ultra AI automatically routes your request to the best-fitting model. - **Consistent, cache-friendly conversations.** Each conversation sticks with the same underlying model, keeping answers coherent _and_ keeping the prompt cache warm — which directly lowers your cost (cached reads are billed at a fraction of the normal rate). - **Resilient by design.** If a provider has an outage, the platform falls back to healthy models — your work continues instead of stopping with an error. - **Bring your own models when you want to.** A full model-management UI lets you add providers and models by API key, or run entirely local models — with guided setup, model search, and one-click connection tests. - **The right model for each job.** Chat, inline edit, applying changes, and autocomplete can each use the model best suited to that role — configured once, then automatic. --- ### Token & Cost Optimization — Built Into Everything Most AI assistants treat cost as your problem. Ultra AI treats it as an engineering problem — and solves it in layers. - **Prompt caching that actually pays off.** Conversation-sticky routing preserves each conversation's cache, so repeated context is served from cache instead of re-billed. Your **cache-hit rate is shown as a percentage** on every usage row — you can literally watch your savings. - **Context that manages itself.** A live meter shows how full the model's context window is. As long sessions grow, Ultra AI prunes the oldest messages automatically — and one click on **"Compact conversation"** summarizes the older history so the assistant keeps its memory of the task at a fraction of the tokens. It never "loses the thread," and you never pay to resend the whole transcript. - **Oversized content is caught before it costs you.** Attachments and context items that would blow past the model's limit are flagged up front with clear explanations — no silent token bombs. - **Effort control = spend control.** The Low/Normal/Max effort dial and the reasoning toggle let you decide, per task, whether to spend tokens on deep thinking. - **Full cost transparency, down to the request.** An optional Console panel charts cumulative cost over your interactions and breaks every request down: input, output, and thinking tokens, cache reads and writes, cache-hit rate, total cost, time-to-first-token, and tokens per second. Tokens-per-day and per-model charts show where your budget goes. --- ### Your Account, Transparently Everything about your plan and spending is visible inside the editor — no hunting through a web portal. - **Usage panel.** Token totals for the **current session, today, this week, this month, and all time** — each with its cache-hit percentage and an **estimated points cost**, refreshed live. - **Today's budget meter.** A headline bar shows today's usage against your daily allowance, turning amber and then red as you approach the cap — and displaying a clean "Unlimited" badge on uncapped plans. - **Subscription panel.** Your plan (Free / Pro / Enterprise) with its price shown in your own currency, requests made today, per-minute request and token limits, and your **credit balance with a live ≈¥ conversion** so the number always means something. - **"How points are charged."** An expandable breakdown shows the exact per-token rate for input, output, and cache-hit tokens — pricing you can read, not fine print. - **Friendly limits.** Hitting a daily cap produces a clear message with a reset time and an upgrade path — never a cryptic error. Buying more credits is one click away. - **Stay current effortlessly.** A built-in "Update extension" check tells you when a new version is available. --- ### Built for Everyone — Zero Code Required Ultra AI was designed so that a product manager, designer, or founder can drive it as effectively as a senior engineer. - **Describe it, get it.** The welcome screen says it plainly: _"Describe what you want to build or change in plain English — Ultra AI will create the files for you."_ Starter prompts like "Build a to-do list app with a clean UI" or "Make a landing page for my coffee shop" get non-developers producing real results in their first minute. - **A one-minute guided setup.** Onboarding walks you through five simple steps — welcome, region, plan, a live connection test (with measured latency), done. No keys, no config files, no terminal. - **Region-aware from day one.** Choose China or Global routing during setup; the model lineup is identical in both, and deployments in China respect local network compliance out of the box. - **Fully bilingual.** The entire interface ships in **English and 中文**, switchable in settings. - **It speaks human.** Agent activity is narrated in plain sentences, clarifying questions come with clickable answers, and every confirmation dialog explains consequences in normal language. - **Comfort features included.** Optional text-to-speech reads responses aloud; adjustable font size, markdown formatting, code-block wrapping, and theme-aware styling keep long sessions pleasant. Anonymous telemetry is a single, clearly labeled opt-out. --- ### Fast, Focused, Unobtrusive - **Never slows your editor down.** Ultra AI activates _after_ VS Code finishes starting, so your editor opens at full speed. - **Leaner with every release.** The latest version trims resource use and feels snappier, with cleaner and more concise responses. - **Keyboard-first.** Every core action — chat, edit, accept, reject, autocomplete, debug terminal, full-screen — has a shortcut, so power users never touch the mouse. - **Cross-platform.** macOS, Windows, and Linux. --- ### Coming Soon The roadmap is public and ambitious. Highlights already in development (in rough order): - **Checkpoint & Rewind** — roll the project back to any point in an agent run with one click _(shipping first; the underlying snapshot engine is already built)_. - **Background tasks & task board** — hand jobs to agents that keep working on the server while you do something else, and queue several at once. - **Voice control** — push-to-talk your instructions _(the voice input engine is already in the product, gated for rollout)_. - **Screenshot to code** — turn a design image into working code. - **Self-checking work** — the agent runs your build and tests and keeps fixing until everything passes. - **Skill & tools marketplaces, shared team memory, scheduled runs, audit logs**, and more. Everything above the "Coming soon" line is shipping today — this section is where Ultra AI is headed next, not marketing for features that don't exist yet. --- ## Platform & Control Ultra AI is more than an AI coding assistant — it is a **complete, self-hosted AI platform** with a full admin dashboard. Everything your team's assistant does runs through **your own server**, and everything about it is controlled from one place: users, models, pricing, limits, languages, releases, and security. ### Team & User Management **Every user, fully managed.** The admin dashboard gives the operator a complete view and full control over every account on the platform. - **User directory with usage insight** — see every user, their tier, plan, credit balance, and exactly how much they have used, with per-user usage charts and per-model breakdowns. - **Tiers with real limits** — assign users to tiers, each with its own daily token cap, requests-per-minute, and tokens-per-minute limits, plus editable per-tier feature flags. Limits are enforced live at the server. - **Per-user overrides** — grant an individual user a higher (or lower) cap than their tier, applied instantly, no redeploy. - **Plans & purchases** — define subscription plans priced natively per region (USD for global, CNY for China — no confusing currency conversion), link users to plans, and manage purchases and refunds directly from the dashboard. - **Credits ledger** — every credit grant, purchase, and adjustment is recorded in a transaction ledger, so balances are always explainable. ### The Points Economy — You Set the Price **A built-in pricing engine that puts the margin in the operator's hands.** Ultra AI meters every token that flows through the platform, and the operator decides what those tokens are worth. - **Per-token pricing** — set separate prices for input tokens, output tokens, and (cheaper) cache-hit tokens, with optional per-model overrides. - **Points-per-RMB exchange rate** — decide how many points a customer receives per unit of currency, so packages and top-ups map cleanly to real money. - **Live margin calculator** — enter your provider's wholesale cost and the dashboard instantly shows your revenue, cost, and margin percentage per million tokens, so you can tune prices to a target margin before anything goes live. - **Safe by default** — per-request charging (debiting points and blocking at zero balance) ships **switched off**. You can configure and preview the whole economy with no effect on users, then flip enforcement on when you're ready. - **Transparent to users** — each user sees their points balance, its approximate money value, and how points are charged, right inside the editor. ### Usage & Cost Analytics **Know exactly where every token goes.** The dashboard turns raw usage into decisions. - **Platform-wide analytics** — usage over time with selectable ranges (last 24 hours, 7 days, 30 days, all time), request counts, and live server status. - **Full token breakdown** — prompt, completion, and cached tokens are metered separately, platform-wide, per user, and per model — so cache savings are visible, not invisible. - **Per-model analytics** — see which models are used by how many users, with token and request totals, filtered by time range. - **Per-user deep dive** — every user's detail page includes a usage-over-time chart and a per-model table. - **Revenue reporting** — revenue is tracked per currency (USD and CNY separately), matching the region-native plans. ### Model & Provider Control **Your models, your routing, your rules.** The operator decides which AI models the team uses and how requests reach them. - **Provider management** — add and manage AI providers from the dashboard; API keys are **encrypted at rest** and never shown in the clear. - **Live model catalog** — the dashboard pulls each provider's live model list, and keeps serving the last-known-good list even during a provider outage. - **Editable routing control center** — direct which models handle which traffic, with per-model weights and ordering, all editable live. - **Smart blend tiers** — user-facing model tiers (Lite / Pro / Max) are weighted blends of real models, letting you trade cost against quality per tier — and the blend ratios are tunable at runtime. - **Automatic failover** — models carry a fallback order, so if one is unavailable, traffic moves to the next without users noticing. - **The real model stays private** — clients see your product's model names, never the upstream vendor's, keeping your supplier choices your business. ### Live Configuration — No Redeploys **Change the platform while it runs.** Operational settings live in the database and take effect immediately. - **System Settings page** — identity prompt, attachment limits, context compaction behavior, unlimited-access email lists, and more, all editable live with environment values as safe fallbacks. - **Servers & endpoints page** — one consolidated view of regional endpoints (China / global) and the active upstream. - **Live rate limiting** — tier limits and per-user caps apply the moment you save them. - **Announcements & broadcast** — publish messages to users (optionally targeted by tier or region, with start/end times); they appear as a dismissible banner inside the editor. ### Localization Control **Speak your users' language — and change the wording yourself.** - **Bilingual admin dashboard** — the entire dashboard runs in English and Chinese, with a USD/CNY display toggle. - **Push UI text to the extension** — the operator can override any string the extension shows, per language, from the Localization page. Changes merge over the built-in English and Chinese text at runtime. - **Add new languages** — entirely new languages can be added from the server without shipping a new extension build. ### Extension Version Management **Ship updates on your schedule, not the marketplace's.** - **Release management** — upload and manage extension releases from the dashboard, organized into channels. - **Staged rollouts** — release to a percentage of users first, then widen. - **Controlled auto-update** — set the update policy centrally; users also get a one-click "Update extension" button and a manual update check. - **Minimum supported version** — retire old clients cleanly by setting a floor version. - **Adoption visibility** — see which client versions are actually in use across your user base. --- ## Cost & Token Optimization **This is where Ultra AI earns its keep.** Serving AI coding at scale is a cost problem, and Ultra AI attacks it from every angle — then shows you the receipts. - **Prompt caching, engineered in** — conversations are routed _stickily_: the same conversation always lands on the same model, deliberately preserving the provider's prefix cache. Cache-read tokens are dramatically cheaper than fresh ones (on the order of ~6x on supported gateways), and Ultra AI is built to hit that cache as often as possible. - **Cache hits you can see** — cached tokens are metered separately from regular prompt tokens, in platform analytics, per-user views, and per-model tables. Your cache savings are a number on a dashboard, not a guess. - **Cache-aware pricing** — the points economy prices cache-hit tokens at their own (lower) rate, so your customer pricing can mirror your real cost structure and protect margin automatically. - **Context compaction** — long coding sessions are the expensive ones. Ultra AI summarizes older parts of a conversation server-side, so multi-hour sessions keep their memory without re-sending an ever-growing history on every request. - **Cost-aware smart routing** — weighted tier blends send lighter work to fast, inexpensive models and reserve strong models for demanding tasks; ratios are tunable live to match your budget. - **Automatic failover** — provider hiccups don't become outages: requests fall through to the next configured model, keeping the team working. - **Full transparency** — every prompt, completion, and cached token is accounted for, per user and per model, with time-ranged charts. Combined with the margin calculator, the operator can tune the platform to a target margin with real data, not estimates. --- ## Security & Trust **Self-hosted by design: your platform, your data, your rules.** - **Enterprise authentication** — end users sign in through WorkOS AuthKit, an enterprise-grade authentication service, with sessions handled by the extension's secure storage (the OS keychain — tokens are never written to disk in plain text). - **Role-based admin access** — the dashboard has its own admin accounts with role-based permissions, so a billing admin and a superadmin see and control different things. Sensitive operations (like changing pricing) are restricted to superadmins. - **Hardened admin login** — admin sessions use dedicated signed tokens, login is rate-limited against brute force, and every response carries security headers. - **Encrypted secrets** — provider API keys and admin-managed secrets are encrypted at rest with a dedicated key; they never appear in plain text in the dashboard or database. - **Network controls** — a managed CORS origin allowlist plus IP allow/deny rules (with CIDR support), all editable from the Security page. - **Full audit trail** — admin actions are recorded in an audit log — who changed what, and when — covering users, pricing, subscriptions, limits, and extension updates. - **Egress control** — the extension includes an enforceable outbound-traffic allowlist, so client network access can be restricted to approved destinations. A fully locked-down "private mode" packaging of this is on the roadmap. - **Protected prompt IP** — the product's prompt assets ship encrypted (AES-256-GCM) with signed manifests and per-file integrity hashes — three independent tamper checks. - **Honest privacy posture** — source code leaves the developer's machine only as the context they explicitly send; local indexes stay local; telemetry is anonymous, content-free, and can be switched off entirely. --- ## What's Coming Ultra AI's direction is simple: **from an assistant you talk to, to a workforce you manage.** The items below are in active planning or development — Checkpoint & Rewind is being built first, and much of the rest builds on infrastructure that already exists in the platform today. Timing may shift; each feature ships when it's solid. ### Safety Net First - **Checkpoint & Rewind** _(building now)_ — every change the agent makes is snapshotted automatically, so you can roll the project back to any earlier point with one click. Confidence to let the agent work boldly, with an undo for everything. ### Agents That Work While You Don't - **Background tasks** — hand the agent a job and close your editor; it keeps working on your server and notifies you when it's done. - **Task board & scheduled runs** — queue several jobs for the agent to work through, and set recurring work like overnight dependency updates or morning issue triage. - **Team mode** — split a big job across several agents at once: one on the backend, one on the frontend, one on tests. - **Self-checking and self-healing** — the agent runs your build, tests, and even the live app, reads the real errors, and keeps fixing until everything passes. - **Ticket to pull request** — point it at an issue and it plans, writes, tests, and opens a pull request ready for review. - **Plan before building** — on bigger tasks, a clear step-by-step plan you can read, edit, and approve before any code is touched. - **Mid-task steering & resume** — redirect the agent while it works without starting over, and pick up any previous session with full memory intact. ### Skills, Tools & Marketplaces - **Skill marketplace** — browse and install skills with one click, so the agent adopts your test style, conventions, and deployment steps. - **Tools marketplace** — connect databases, design tools, or your ticketing system with a single click. - **Private team registry** — publish your team's own skills and tools internally, so company know-how spreads to everyone automatically. - **Custom specialist agents** — build your own experts (a tester, a security reviewer, a database specialist), each with its own instructions and tools. - **Automatic rules (hooks)** — "auto-format after every edit," "block risky commands" — rules that fire automatically, every time. - **Bring your own models** — plug in the AI models you prefer. ### Built for Teams - **Shared team memory** — what one person teaches the agent becomes available to everyone, including background agents. - **One-step onboarding** — a new teammate enters a single address and instantly inherits the team's models, rules, and shared knowledge. No keys to configure. - **Cost and time guardrails** — spending and time limits per task and fleet-wide, so work stops before it overruns. - **Private mode** — a locked-down configuration where code never leaves your own server, built for regulated industries. - **Headless / CI mode** — run the agent from scripts and pipelines with no chat window, turning it into production infrastructure. ### Deeper Codebase Intelligence - **Project-wide changes** — the agent finds every place a change must land across the whole codebase and verifies the build still passes. - **Built-in code review & security scanning** — catch bugs, vulnerabilities, and exposed secrets before you push. - **Living architecture maps** — auto-generated diagrams that stay current, so new people find their way around fast. - **Test coverage helper** — finds untested code and offers to write the missing tests. - **Ask your codebase** — plain-language answers about how your project works. ### Everyday Polish - **Review all changes on one screen**, **voice control**, **screenshot to code**, **one-click fixes from editor warnings**, **desktop notifications** when long jobs finish, and a **faster, lighter startup**. ### Why This Direction Matters Much of this roadmap is _activation_ rather than invention — background runs, sub-agent dispatch, context compaction, reasoning-effort controls, and workspace isolation already exist inside the platform and are being surfaced as polished features. That means the roadmap is grounded in shipped engineering, not wishful thinking — and it compounds: the same self-hosted server that meters tokens and enforces limits today becomes the control plane for a whole fleet of agents tomorrow. --- ## Changelog A feature-level history of what has shipped in Ultra AI, newest first. ### Latest — The Points Economy - **Admin-configurable pricing engine** — the operator now sets per-token prices for input, output, and cache-hit tokens, defines the points-per-RMB exchange rate, and sees profit margin live in a built-in calculator. Optional per-request charging (with balance enforcement) is included but ships switched off, so the economy can be tuned safely before going live. Users see their points balance and rates inside the editor. ### Recent — Analytics & Insight - **Per-user analytics** — every user's detail page gained a usage-over-time chart and a per-model usage table, filterable by time range. - **Live model catalog + per-model analytics** — the dashboard pulls each provider's live model list (and keeps serving the last-known-good list during outages), with per-model usage: users, prompt/completion/cached tokens, and request counts. - **Full token breakdown** — prompt, completion, and cached tokens are now reported separately across the dashboard, analytics, and per-user views — making cache savings visible. - **Live system status** — a real-time server health card on the dashboard and an online indicator in the header. - **Resilience pass** — the dashboard became blank-page-proof with graceful error recovery, and the test suite was brought fully green. ### Full Admin Control (Part 2) - **System Settings, live** — operational knobs (identity, compaction, attachments, and more) moved into the database, editable from the dashboard with instant effect and no redeploy; plus a consolidated Servers/Endpoints view. - **Localization from the server** — admins can override any extension UI string per language and add entirely new languages, pushed live to users. - **Bilingual dashboard** — the entire admin dashboard now runs in English and Chinese, with a USD/CNY currency toggle. - **Region-native plans** — subscription plans priced natively per region (USD global, CNY China), linked to user accounts, with revenue reported per currency — no cross-currency mixing. - **Announcements** — broadcast messages to users (targetable by tier/region, schedulable), shown as a dismissible banner in the editor. - **Per-user limit overrides & editable tiers** — individual caps that override tier defaults, enforced live, plus editable per-tier feature flags. - **Admin team management** — full admin-account management with role controls, self-service password change, and lockout protections. ### Extension & Account - **Subscription panel in the editor** — users see their plan, usage, limits, and credits without leaving VS Code, plus a one-click extension update button and manual update check. - **Controlled extension updates** — release management with channels, staged percentage rollouts, update policy, and a minimum-supported-version floor. ### The Admin Platform (Foundation) - **Complete admin dashboard** delivered in phases: secure admin login with role-based access; user management with usage visibility; database-backed live settings with a live rate limiter; AI provider/model/routing management with encrypted keys and an editable routing control center; credits and purchase management; network security (CORS allowlist + IP allow/deny rules); analytics charts, audit log, and run viewers; production hardening; and a modern dark UI. ### Smarter Serving & the Editor Experience - **Smart tier models** — Lite/Pro/Max virtual models that blend fast-cheap and strong models by weight, with agentic (tool-using) requests automatically routed to the strong model. - **Server-controlled identity & privacy** — the assistant's identity is injected at the server, and the real upstream model id is hidden from clients. - **Server-controlled chat config & file attachments** — chat behavior and attachment limits managed centrally; image attachments supported. - **Refined auto-approve controls** — an explicit dropdown for agent autonomy levels, with clear confirmation for full autonomy. - **Cost, memory, image & routing improvements** — a coordinated round of serving-cost, session-memory, and routing upgrades. - **Localized product** — professional English and Chinese product documentation and UI. --- ### At a Glance **Security** — the platform has moved steadily toward enterprise posture: role-based admin access, encrypted provider keys, rate-limited hardened login, CORS and IP network rules, a complete admin audit trail, and encrypted, integrity-checked prompt assets. **Optimization** — cache-hit tokens are now metered, priced, and visible end-to-end; conversations are routed to preserve prompt caches; long sessions are compacted; and blended tier routing with automatic failover keeps quality up and cost down. **Potential & scale** — with live configuration, staged extension rollouts, region-native billing, a tunable points economy, and per-user/per-model analytics, the platform is built to grow from a single team to a multi-region commercial service — and the same control plane is ready to manage the fleet of background and team agents on the roadmap.