v5.3.3 - Row-Bot In Your Browser, Every Model In One Place

v5.3.3 - Row-Bot In Your Browser, Every Model In One Place

Row-Bot 5.3.3 runs in your browser, in your own Railway account, with nothing to
install. Settings brings providers and models onto one page, adds twelve API
providers and three model servers, and lets images and video come from far more
providers. Apps search puts the app you meant first. From 5.3.2 it arrives
through background updates.

Row-Bot In Your Browser

• Run in your browser - on row-bot.ai, Continue with Railway signs you in to
  Railway and creates your own Row-Bot in your own Railway account in a few
  minutes. It opens already signed in: no address or setup code to find. Railway
  bills you directly, usually $5-10 a month.
• Nothing kept on the way - Railway's sign-in goes straight to Railway. The
  page's calls to Railway pass through a small relay that can only create a
  Row-Bot, never delete or change anything else, and keeps no logs, no tokens
  and no addresses.
• In Railway's words - on Railway, Row-Bot's pages link straight to your
  service's settings, variables and usage. Settings › Updates says your
  automatic updates install new versions, and the Data page says Railway backs
  up volumes only on its Pro plan, so Download backup comes first.
• Fits a small server - below 4 GB, local speech-to-text is marked "Needs a
  4 GB server", and Row-Bot's browser closes ten minutes after its last use. If
  the server runs out of memory, Home and System Monitor say so, with how to
  resize it.
• The deploy template works too, and so does any host that runs the image.

Models

• One Models page - Settings › Providers and Models are now one page: your
  defaults (Main model, Vision, Image, Video), your models, and every connected
  provider, with what each model can do.
• Stars - a star is the one way to put a model in your lists. Every picker
  shows your starred models that can do its job. On first start, your
  favourites and pinned models become stars, and the old files are kept beside
  the new ones (.pre-5.3.3.json).
• Twelve more API providers - DeepSeek, Mistral, Groq, Kimi (Moonshot), Z.AI
  (GLM), Qwen (Alibaba Cloud), Together AI, Fireworks AI, Cerebras, Hugging Face,
  NVIDIA NIM and Vercel AI Gateway.
• Model servers on this computer - Add a provider › On this computer offers
  LM Studio, vLLM and llama.cpp.
• A picker that finds anything - search every connected model as you type.
  Your starred and recent models come first, the rest wait under Browse by
  provider, and a model several providers offer shows once, with a +N routes
  chip to choose another. Connect or Reconnect opens that provider in
  Settings › Models.
• First run lists every provider that takes an API key.
• A backup model - choose one in Settings › Models › Defaults (off until you
  do). When your main model can't answer (its server isn't running, it timed out,
  hit a rate limit, ran out of credit or lost its sign-in), the backup finishes
  the turn, and the chat says which model answered. It never steps in when you
  press Stop or when the request itself was the problem, and an incognito chat
  never goes to a cloud backup. Chats, channels and workflows all use it.
• Thinking remembered - the Thinking level you choose for a model is where
  new chats with it start. Chats you already have keep their own.
• New models know what they can do - where a provider's own list doesn't
  say, Row-Bot fills in context size, tools, vision and output kind from
  models.dev's open catalogue. It's fetched at most hourly when model lists
  refresh, with only the request sent; when that fails, the copy shipped with
  the release is used. Settings › System › Privacy lists it.
• Claude Code stays up to date - while your Claude plan is set up, Row-Bot
  checks Anthropic about once a day for a newer Claude Code and installs it,
  checked against Anthropic's checksum, while Claude isn't answering. Claude's
  model list is then read again, so new models (such as Sonnet 5.5 and Haiku
  5.5) appear without pressing Update Claude Code. A failed update keeps the
  version you have and tries again the next day; Claude Code's own updater
  stays off. Settings › System › Privacy and the privacy page list the check.
• Use another Claude account - when your Claude plan is connected, the
  two-arrows button in its sheet switches Row-Bot to another Claude account. It
  asks first, because Claude Code keeps one sign-in on your computer: this signs
  it out everywhere there, claude in a terminal included. Claude Code signs
  itself out with its own command, Claude's own sign-in opens for the other
  account, and the model picker then shows that account's models. Row-Bot never
  sees either sign-in.
• The right plan name - the Claude sheet names the plan Claude Code's own
  session runs on. Before, it could show an older plan of the same Claude
  account (Claude Max after a switch to Ultra), because claude auth status
  still named it. After a sign-in the plan is right at once.

Images And Video

• From more providers - image and video tools also work with OpenRouter
  (images, edits and video), Vercel AI Gateway (images), Together AI (images and
  video), Requesty (images), Qwen (Alibaba Cloud), Atlas Cloud, MiniMax and Z.AI.
  MiniMax's and Z.AI's model lists leave out their image and video models, so
  Row-Bot adds the ones their docs name. Together AI video asks each model for a
  resolution and aspect ratio, so models that take only their own frame sizes,
  such as Seedance 1.0 Lite, work.
• Models from each provider's own list - a retired model no longer lingers.
  OpenAI retired gpt-image-1 on 23 Oct and retires gpt-image-1-mini and
  gpt-image-1.5 on 1 Dec. A saved model the provider no longer lists says so in
  Settings › Models.
• Settings that fit the model - OpenRouter image and video models get the
  length, size and aspect ratio nearest the one asked for, of those OpenRouter
  lists for each model.
• Outputs - a chat's Outputs are a grid of thumbnails, newest first, each
  labelled with the start of the prompt that made it (the whole prompt on
  hover). Click opens it full size; ⋯ shows it in the chat, saves it to the
  workspace or uses it in code or a design.
• Videos after a reload - a chat shows the videos it made, with the newest
  in full and earlier versions folded under one row, live and after a reload.
• Every file in a chat has its own name - five generated videos are
  generated-video.mp4, generated-video-2.mp4 and so on, so the video editor
  can open each one. In older chats, files that share a name are told apart the
  same way, newest first.

Apps Search

• The app you meant, first - a Best match, then four more good ones, then
  the rest folded under Official, Reviewed by Row-Bot, Built in and Community.
  When nothing official or reviewed matches, it says so.
• Words, not fragments - "reddit" no longer finds "subreddit".
• One card per app - the same server listed in several places shows once.
• Labels and stars - each card says who stands behind it and how
  well-kept it is (★ 1.2k · active), from data shipped with the release.
• Filters - kind, Official & reviewed only, where it runs and No sign-in;
  on a phone, in one Filters sheet.

Fixes

• Security: tool settings that can't be read keep tools off - if
  tools_config.json was cut short or locked, tools you'd switched off came
  back on and the shell's blocked-commands list emptied. Now tools stay as they
  were, or off at start-up, and Settings says so and offers Start from defaults,
  which keeps the unreadable file and your workspace folder. If Row-Bot started
  with the file already unreadable, it can't tell your folder, so it says the
  workspace goes back to Documents › Row-Bot, before and after. Reported
  privately; advisory GHSA-68j6-243q-h2hp (low).
• A long chat that starts with generated media opens again instead of
  showing "Connection interrupted".
• Saving is one way everywhere - every Download and Export asks where to
  save, then offers Show in folder. The notice names the file you chose, and a
  chat exports under its own title.
• Web pages read cleanly - reading a page gives its main text as Markdown,
  with headings, lists, tables and full links, and without menus, sidebars or
  cookie banners. It's done on your computer; nothing goes to a reading service.
• Phones stay signed in - a phone that connected before 5.3.2 kept an old
  cookie its browser left out when Row-Bot was opened from the home screen.
  Each session check now renews the cookie (never the session's length).
• Ollama's models show after Claude Code is signed in on the same computer.
• llama.cpp suggests port 8081 when Row-Bot itself is on 8080.
• Model servers on this computer - LM Studio, vLLM and llama.cpp save on
  the first try and are checked as they're added, so their models are ready to
  use. A server's models count as ready only once its check passed, and when a
  model can't answer, the chat box says why.
• Your Claude plan lasts much longer - every model call in a turn started
  a new Claude Code that sent the whole conversation again, and Anthropic's
  prompt cache stored it again each time instead of reading it back, so a long
  agent turn used up a 5-hour limit in under an hour. Now one Claude Code
  stays with the turn: Row-Bot still runs every tool through its own approval
  check and hands Claude Code the result, and the conversation is read from
  the cache.
• Your Claude plan's full context - Row-Bot sized Claude plan models by
  their names, so with Claude Code 2.1.322 Opus, Sonnet and Haiku 5.5 read as
  200K although Claude Code runs them at 1M, and a long chat was refused before
  it reached Claude. Row-Bot now uses the context Claude Code reports with each
  reply, and Claude's own sizes before the first one. The models are always the
  ones your plan's Claude Code lists, and Row-Bot now keeps Claude Code up to
  date, so newer ones appear.
• No more chats refused on a guessed size - where nothing said how much a
  model reads (Requesty, Atlas Cloud, Ollama Cloud, Grok sign-in, DeepSeek,
  Z.AI and other providers' newer models), Row-Bot assumed 32K to 200K and could
  refuse a long chat with a 1M model itself. It now finds each model in
  models.dev's catalogue (Requesty's and Grok's models included), and when
  nothing knows a model's size it still tidies long chats but lets the provider
  decide instead of refusing.
• Videos in order - while Row-Bot makes videos, each step's words, its
  activity and the video it made show in the order they happened, with the
  newest step at the bottom, live and after a reload.
• A new workflow runs on the first click after it's saved.
• Wikipedia reads again - Row-Bot's page reader names itself
  (Row-Bot/5.3.3) instead of posing as a browser, as Wikipedia asks.
• Settings switches show what's saved when it changes elsewhere, such as
  after Start from defaults.
• Workflow chats move up the sidebar only when something new shows in them;
  a run that didn't finish says why.
• Selecting workflows shows a checkbox in each row, and Select all.
• X replies - X allows a reply or quote from its API only when the post's
  author mentioned or quoted you. Row-Bot now checks before asking you, and
  otherwise gives you the draft and the link to post it on X yourself.
• Security and conduct reports go to syd@row-bot.ai.
• A failed reply says why - instead of only "The response was
  interrupted", the chat shows the reason, such as running out of credits or
  the provider's own words, and keeps it after a reload.
• Gemini answers again on Google's own API and through gateways such as
  Vercel: a few tools offered an empty choice, which Gemini refuses.
• Images and video work with keys saved in the new provider panel (Qwen,
  Together, Vercel and the others).
• Model lists keep themselves current again - as before 5.0, about a minute
  after start-up and then every six hours, a list older than six hours is read
  again from your connected providers (their model lists only: no prompt, no
  tokens). With no cloud provider connected nothing is contacted. Connecting a
  provider or Refresh under Settings › Models › Catalog still updates at once.
• OpenRouter's image and video lists offer only the models its image and
  video services take, so Auto Router and its kin stay in the chat list.
• Thinking controls for more models - a model that reasons gets a thinking
  control on Groq, Cerebras, Fireworks, Together, Hugging Face, Vercel, NVIDIA,
  Z.AI, DeepSeek, Kimi and Qwen, and on newer OpenAI, Anthropic and Google
  models.
• Mistral's thinking goes to the thinking panel instead of showing as raw
  text in the answer. Cerebras answers: it gets Row-Bot's instructions as
  one message, as its models need.
• Cleaner model lists - one model from several providers shows once with
  its other routes (one per provider), Mistral no longer repeats a model per
  alias, Qwen's image makers stay out of the chat list, Fireworks models read by
  name, and a model the catalogue lists without tools says Chat only.
• No more "Part of Row-Bot isn't responding" after tidying up - ten
  minutes after start-up and every six hours, Row-Bot tidies old chat history.
  With a large chat history it also rewrote the whole chats database, which
  stalled every chat for a minute or more. Now it never rewrites a large one
  while Row-Bot runs (the space is reused anyway), tidies in small steps, and
  reading your chats never waits on it. Old chat history is now trimmed every
  hour even while workflows keep Row-Bot busy, so a long workflow chat no
  longer grows the chats database by gigabytes between tidy-ups.
• Smaller fixes - a bold link in an answer stays a link; reading a short
  page no longer sends the model to the browser; Show in folder finds a file
  whose path has a space; exports skip empty steps and say where an image was
  made; a workflow's notice says what went out; outputs are named after their
  subject; the API-key sheet keeps the provider's name and has icons on its
  buttons.

For Hosts

• Railway - Row-Bot knows Railway as a host (ROW_BOT_HOSTED_BY=railway).
  Its pages use Railway's words and, with the IDs Railway gives every service
  (RAILWAY_PROJECT_ID, RAILWAY_SERVICE_ID, RAILWAY_ENVIRONMENT_ID, each
  checked as a UUID), link the service's own Settings and Variables and the
  workspace's Usage. Started as root (RAILWAY_RUN_UID=0), the image hands
  /data to uid 10001 once and drops to it for good. Row-Bot's template and
  its Railway page set 8 GB / 8 vCPU (the Hobby plan's ceiling, billed by use),
  nightly automatic updates (Overnight, 02:00-06:00 UTC), PORT=8080 and no
  health check: Railway's probe gets a 400 and the deploy hangs.
  deploy/railway/README.md has the definition.
• Visitor address on Railway - from a trusted proxy, Row-Bot accepts
  X-Real-IP beside X-Forwarded-For, and on Railway takes the visitor from
  X-Real-IP, since Railway's X-Forwarded-For ends with its edge's own
  address. On Railway set ROW_BOT_TRUSTED_PROXY_CIDRS=100.64.0.0/10. Other
  hosts are unchanged.
• Health checks from inside the container - a check from a loopback
  address with no Forwarded or X-Forwarded-* header now passes whatever
  host names Row-Bot allows. A check through a proxy, or from another
  computer, still needs a host name Row-Bot knows.
• Container logs on stdout - In the container, Row-Bot's console output
  goes to stdout and warnings and errors to stderr; the setup code and its
  sign-in link are on stdout. Log views that colour stderr no longer show
  every line as an error.
• A sign-in link beside the setup code - while nobody owns a server and
  ROW_BOT_PUBLIC_URL is set, the start-up output adds Open this link to sign in: https://<address>/setup#code=<code>. Like the code, it's only in
  the console, never in Row-Bot's log files, and never printed once someone
  owns the server or for a recovery code. The code now appears twice at the
  first start (its line and the link).
• /setup#code=... - opening the setup page with the code in its fragment
  signs in with one click, with the same checks, refusals and pauses; the page
  clears the fragment at once, and the fragment never reaches the server.
• Updates in the host's words - with ROW_BOT_HOSTED_BY set, Settings ›
  Updates says who installs new versions: PikaPods installs each one, and on
  Railway the owner's automatic updates do, with a link to the service's
  settings.
• Memory on a small server - Row-Bot reads the container's limit (cgroup
  memory.max) and keeps, in /data, the highest working set it has seen.
  Nothing is turned off: below 3.5 GB, Settings › Voice marks local
  speech-to-text "Needs a 4 GB server" and asks before it's chosen, and
  Row-Bot's browser closes ten minutes after its last use. If Row-Bot
  restarts because the server ran out of memory, Home and System Monitor say
  so, with the server's size and how to resize it in the host, until the
  owner dismisses it.
• Memory in the host's unit - a limit set in decimal GB (Railway's
  8,000,000,000 bytes) shows as 8 GB, and Docker's --memory 4g as 4 GB.
• Backups on Railway - Railway backs up volumes only on its Pro plan.
  Row-Bot's Data page says so on Railway; on any hosted Row-Bot, Download
  backup is the main action, and Back up now says its copy stays on the server.
• No new required setting. The recommended size for a pod stays 4 GB. The
  hosting smoke's memory high-water mark is about 0.5 GB on amd64 and arm64.

schema: 1
files:
  Row-Bot-5.3.3-Linux-x86_64.tar.gz: sha256=6f4dec9059127f69bbd92938def07b3ebbf541bb00dcea5aa11cf4b8ff4c57fc
  Row-Bot-5.3.3-Windows-x64.exe: sha256=f453b32ebe2bc4af6e5b15734656dc0971ba0db27b2babed6d7c7acf65097993
  Row-Bot-5.3.3-macOS-arm64.dmg: sha256=b204b9d9d84fad660d0367dd365d38a42b6a32ee1e888b5377f92d34147f61b5
  row-bot-assets-5.3.3.tar.xz: sha256=f6a6a5064ed9e95a1f8de2d66bebdef607b2a1630f35c9a371dfd83642ecc8ce
  row-bot-code-5.3.3.tar.xz: sha256=e1546dd906d91bbfec078188af1e06a3c35c7e0203f83681afc850b245ba28ff
  row-bot-container-5.3.3.json: sha256=6fce619b57bf11d06232a5dded7074f382957f7173bf6fcb8e457fbcf7a82dd6
  row-bot-pylock-5.3.3.toml: sha256=8ef67de06906cb1e3822ae44a0e9c9435e9840c1860d6e17716c47e34ca20401
  row-bot-update-5.3.3.json: sha256=4d6898fb90b0a5b5c193c5b3af7595532af62a4e864d368f19182e602fca5ec5