Changelog
What we shipped.
Newest first. Beta versions tend to land features fast — expect a few entries a week while we round out the platform.
Your data, your continent
- EU-only AI routing: flip one switch under Settings → Privacy and your agents' text generation runs exclusively through our provider's EU-region endpoint. It fails closed — if no in-region provider can serve a model, the request errors out instead of leaving the EU. Around 70 models are eligible, including the Claude and Gemini flagships.
- EU badges everywhere you pick a model: the model picker marks EU-eligible models, model suggestions for EU-mode users only offer models they can actually run, and the weekly model digest tracks eligibility. Honest scope note: image/video generation stays on standard routing until our provider supports EU media inference.
- ByteDance Seedance videos now have prices: Seedance bills per video token instead of per second, which our estimator couldn't quote — so all five Seedance models failed before generating. The estimator now speaks token pricing (verified against the provider's own published rates), and the model picker shows their real per-second cost.
Claude makes movies now
- Studio over MCP: the connected Claude can now browse the image and video model catalog with live prices, quote a job, generate images and videos into your Studio gallery and list what you've made — "make me a 6-second clip of a foggy forest" in plain conversation.
- Price before pixels: every media tool call surfaces the cost up front, and videos refuse to dispatch without a confirmed quote — a runaway agent can't silently burn your credits. Failed jobs report the reason in plain language and refund the hold, same as the web UI.
Lock the front door
- Two-factor authentication: protect your account (and your credits) with a six-digit code from any authenticator app. Enable it under Settings → Security — QR setup, manual key, and one-time recovery codes for when your phone isn't where you left it.
- Passkeys: sign in with Face ID, Touch ID, Windows Hello or a hardware key — no password typed, nothing to phish. Register a device under Settings → Security; a "Sign in with a passkey" button appears on the login page wherever your browser supports it.
Your factory speaks MCP
- Connect Claude to your account: AgentFactory now runs a remote MCP server. Create a token under Settings → Claude & MCP (shown once, with the ready-made connect command), and Claude Code or Claude Desktop can list your agents, start runs and collect results — conversationally, billed to your credits like any run.
- Failed jobs stop vanishing: the floating jobs indicator now keeps failed image/video jobs visible for 15 minutes with a plain-language reason ("the provider's content filter rejected this — reword it") and the refund confirmation, instead of a toast you could miss.
- Videos are the length you asked for: a silent provider-API mismatch meant every video ran at the model's default duration. Fixed — your 10 seconds are 10 seconds, and billing was always correct either way.
- Security updates across the framework and markdown renderer.
Personas
- Your agents finally know who you are: create personas under Settings — scoped "about me" profiles per field of life. The work persona carries your job context; the hobby persona doesn't need to know your employer. Each agent uses the persona you pick for it, your default, or none at all.
- Scoped on purpose: only the selected persona reaches an agent — no stacking, no leakage between your worlds. Cloned community templates start on your default persona, never the creator's.
Community templates
- Share your agents: a new Share button turns any of your agents into a community template — instructions, model picks, tools, the lot. Submissions are reviewed before going live, and published templates carry your name.
- Clone from the community: the agent wizard now shows what other users are sharing, right next to our hand-tuned templates. One click seeds a fully editable copy — the complete system prompt is on screen before you save. Your clone is independent; nothing links back to the creator's account or data.
Chat that feels alive
- Chat got its heartbeat: your message lands in the thread the instant you hit Send (or Enter — Shift+Enter for a new line), a "Thinking…" indicator holds the floor, and the reply streams in token by token. Live streaming was silently broken before — now you watch the agent type.
- Background jobs show themselves: a floating spinner appears top right while images or videos are rendering. Click it for the list — what's running, on which model, for how long. Follows your light/dark preference.
- Video quotes now match the bill: models that always produce audio (looking at you, Veo) are quoted at the audio rate up front instead of surprising you at settle time, and the audio toggle defaults on where audio is built-in. If a provider charges less than held, you're refunded automatically — the surprise can only be in your favor.
- Prompt polish is fast again: the Studio's meta-agent answers in seconds instead of timing out mid-thought on reasoning models.
- Timezone picker: schedules now offer a searchable timezone dropdown with Europe/Berlin pre-selected — no more guessing identifier formats.
Deutsch, dark mode, smarter wizard
- AgentFactory spricht Deutsch: the whole platform — landing, docs, app, emails — is available in German (informal Du, the way we actually talk). Switch languages in the footer; your choice sticks to your account. Legal pages are German-canonical with English courtesy copies.
- Dark mode everywhere: the public site now follows your system preference, with a manual ◐ toggle in the footer for night owls. The app was always dark; now the front door matches.
- The model digest reads native: German digest entries are written in German by the meta-agent, not translated after the fact.
- Smarter agent wizard: model suggestions now come from a best / mid-tier / cheapest shortlist (like the digest) instead of bargain-bin-only, models are picked from a browsable catalog instead of typing slugs, temperature and max tokens explain themselves, and the browser back button finally does what you expect.
- Studio grew up: paste a whole poem — briefs up to 5,000 characters — and the prompt-polisher turns it into a properly rich image or video prompt instead of two timid sentences. Your generated media now lives in a Gallery right in the sidebar.
Launch hardening
- We're live — AgentFactory now runs in production, with email verification (on-brand mail, finally not default-Laravel-blue) and live job notifications: finished or failed images and videos toast you on whatever page you're on, and failures always refund their hold automatically.
- Playground remembers: reopening an agent resumes your latest conversation, and a "Previous chats" picker switches between recent ones. Chat-only agents no longer show dead Image/Video tabs.
- Honest usage page: free platform-helper calls are labelled
free, failed jobs show refunded instead of a phantom cost, and the spend total matches your balance to the cent.
- Video generation got pickier in your favor: edit-only models no longer masquerade as text-to-video options, models without fixed resolutions quote correctly with a "model default" output, and cost quotes appear instantly.
- Model digest: a curated cheap/balanced/premium shortlist per modality, refreshed weekly (admins can trigger it any time).
Money correctness + legal
- One credit lifecycle for everything: text, image, avatar and video all follow the same estimate → hold → settle → refund path. A background reaper fails and refunds any job whose worker died, so credits can't get stuck.
- Checkout consent: top-ups now ask for your express OK to immediate delivery (§ 356 BGB) before Stripe opens — credits are usable the second you pay.
- Legal pages: legal notice, privacy policy and terms drafted and published at
/legal-notice, /privacy, /terms.
- Security pass across dependencies and friendlier error messages when a provider or model misbehaves.
Top-ups, attachments, scrape
- Stripe top-ups: one-shot credit purchases via Stripe Checkout with idempotent fulfilment — no subscriptions, ever.
- Chat attachments: drop images and PDFs into any chat turn; vision models see them inline.
- Image-to-video start frame on the Video tab, and a
web_scrape tool (Crawlbase) for pages that fight back.
- Wizard template gallery: Monitor URLs, Weekly research, Daily news, Writing coach — one click to a working agent. Plus a duplicate-agent action and a full
/docs rewrite.
Memory, admin top-up, marketing pages
- Agent memory: per-agent summary brief that gets prepended to the system prompt. Edit by hand, regenerate from chat history, or clear. Mirrors how Hermes / OpenWebUI handle memory.
- Admin user management: full CRUD over users — view, create (with optional starter credit), delete (with confirm + self-guard), edit name/email, change password. Per-user transaction ledger and recent runs.
- $5 starter credit on signup, granted automatically by a Registered listener.
- Marketing site rounded out: docs, changelog, status and contact pages. Reusable marketing layout under
components/layouts/marketing.
Brand pass — auth + landing
- AgentFactory landing page: hero with claim triplet, agent-run mock, pay-as-you-go comparison, cost-scale bar, four-cell features grid, three example agents, yellow CTA band.
- Auth pages on the new shell: register, login, forgot-password, reset, verify-email, confirm-password — all wearing the off-white card with hairline-underline inputs and the schematic factory mark.
- Schematic factory logo in the dashboard sidebar + a real SVG favicon.
Run mode + persistence + toasts
- Run vs Chat split in the playground. Run is one-shot autonomous execution; output renders as markdown and survives navigation.
- Top-level Runs section mirrors Conversations: filter, search, view, download
.md, delete.
- Confirmation modals + Flux toasts replace browser
confirm() popups and emerald flash banners across the app.
Tools + memory + media
web_search and web_fetch tools wired up via Tavily. Real Laravel\Ai\Tool implementations.- Conversation memory via the SDK's
RememberConversation middleware. Private toggle skips persistence per-turn.
- Avatar generation on agent create via OpenRouter's image catalog. Letter fallback while the job is queued.
- Video generation with up-front cost estimates + reservation against the credit ledger.
First light
- Build, schedule and run agents end-to-end. Reasoning model picker, system prompt editor, scheduler with plain-English hints.
- Live cost ledger via the
credit_transactions table — every run debits actual usage to the cent.
- Streaming responses via Reverb so the playground feels responsive.