Help

The parts of Ocadox worth knowing about that you might not find on your own.

Chat tabs

Open several chats at once: press + above the chat, or Ctrl+T (Cmd+T on a Mac). Each tab has its own folder and runs on its own, so one can work on a task while you chat in another. A dot on a tab means it's working, has finished, or (amber) is waiting for your OK. Opening a chat from the list shows its tab if it's open, or opens it in place of an idle tab. Closing a tab stops anything it's running; the chat stays in Chats.

Local models take turns on the graphics card, one reply at a time, which keeps the fan guard working; cloud models (Claude Code, Codex, OpenRouter) can answer side by side. Long jobs like a video edit or a song put their result in the tab they started in.

Terminal, editor and moving computers

All in Settings > Terminal, editor and moving computers.

Keeping it quiet

Fans follow temperature. Ocadox keeps the temperature down in four ways, all in Settings:

No graphics card: CPU mode

On an office desktop or a laptop with built-in graphics, models run on the processor. CPU mode turns on by itself (Settings > CPU mode shows why) and:

Use a small model: qwen3:1.7b on 8 GB of memory, qwen3:4b on 16 GB. The model library labels what fits.

Project rules (AGENTS.md)

Put an AGENTS.md, OCADOX.md or CLAUDE.md in the root of your project folder. The agent reads it as standing instructions: how to build, test, what to leave alone. Two lines do more than instruct:

# AGENTS.md
Use pnpm, not npm. Keep components small.

Check: pnpm lint && pnpm test
After edit: npx prettier --write {file}

Both go through the usual "run this command?" approval, unless you've set an auto-approve rule.

Helpers and self-checks

Helpers

For a big task, the agent can hand a question ("where is the login handled?", "why does this error happen?") to a helper that starts with a fresh, empty context and can only look, not change anything. It answers in a few lines, which keeps the agent's own context small: a big help with local models. Helpers run one at a time on the same model, under the same fan guard, so the fans stay calm.

Checking its own work

When the agent has changed files and thinks it's finished, Ocadox runs the project's check and hands any errors back to it to fix, up to twice. Without a Check: line it picks one: cargo check for Rust, tsc --noEmit for TypeScript, go build for Go, a syntax check for Python.

The agent's browser

To test a web app it's building, the agent can open it in a hidden copy of Chrome or Edge (whichever you have), click, type, and read the page and its console errors. Start your dev server, then ask, for example: "open http://localhost:5173, log in as test@example.com and check the dashboard loads without errors". Pages on your own PC always work; other websites only while web access is on (Settings).

From the terminal

ocadox run "fix the failing tests" --folder ~/code/app --model qwen3:8b

On Windows, the terminal doesn't wait for a windowed app to finish. Use start /wait ocadox run "…" in Command Prompt.

Bots, routines and triggers

A bot has its own name, chat, folder and model. Give it a job and close the window; the result is waiting when you come back. With Ocadox Pro, a bot can also work on its own with routines:

A bot's own edits never set it off again. Turn on Save each result to get a file per run in its folder's Ocadox results.

Sharing bots

In a bot's editor, Share as a file saves a small .ocadox-bot.json: its name, look, instructions, starter prompts and routines. Never its folder or chat. To use one, choose New bot > From a file…. Routines from someone else arrive switched off; turn them on when you're happy with them.

The phone app

  1. Install the Android app from the releases page (the _android.apk).
  2. On the PC: Settings > Other computers > turn on "Let my other computers use this PC's bots" > Pair a computer or phone, and scan the code with the phone.
  3. Away from home: install Tailscale (free) on both and it keeps working, no router changes.

Full control: next to a paired phone, tick Full control to use the whole desktop app from it, running on the PC. Off, the phone has bots and chats only. Notifications work with the app closed (a quiet "Watching your bots" notification keeps it connected; switch it off on the list of computers).

Making images and videos

Settings > Image creation > Set up image creation downloads an image engine and the Z-Image Turbo model once (6.7 GB, every file checked). Then type /image_create and what you want, e.g. /image_create a lighthouse at dusk, oil painting. About a minute per image on a good graphics card. It needs 16 GB of memory on a PC without a graphics card. Already run Forge? Put its address in the same settings and Ocadox uses it instead.

Videos

Settings > Image creation > Video creation downloads the Wan 2.2 video model once (9.4 GB). Then /video_create a cat walking through autumn leaves makes a clip, or press Animate under any generated image to bring it to life. Choose the length in Make a video… (command palette), 2 seconds to 2 minutes, or type it: /video_create 30s a cat…. Over 5 seconds, the video is made as 5-second parts, each starting where the last ended, and saved as an MP4; write one line per scene for a storyboard. Each 5 seconds takes a few minutes and works the graphics card hard, so the fan guard and power cap apply. It needs a graphics card (12 GB or more is comfortable).

Speaking in your voice

Settings > Image creation > Speech, in your voice downloads Qwen3-TTS once (2.3 GB). Under Voices, add a 5–30 second WAV or MP3 of clear speech, give it a name and confirm it's your voice or you have the person's permission. Then:

/speak voice=Me Your tests all pass, and the fans stayed quiet.

Without voice= it uses your latest voice. Voices stay on your PC. Never use someone's voice without asking them.

Voice-overs

Press Voice over under a video (or Voice over a video… for a file). Type the script, or describe the video and press Draft with AI; pick a saved voice and whether to add captions. A short clip loops under a longer voice; a longer video holds its last frame until the voice ends.

Editing pictures and videos

Pictures: set up Picture editing once in Settings > Image creation (7 GB). Then press Edit under any picture and say what to change, or Upscale to make it 4 times bigger. Edit a picture… in the command palette works on your own files.

Remove the background: in Edit this picture, press Remove the background for a PNG of the subject with a see-through background. To erase something, say what to remove ("remove the person on the left").

Videos: press Edit under a video, or Edit a video…. Describe the edit and press Fill in, or set it by hand: trim, shape (9:16 for TikTok, Reels and Shorts; 1:1; 16:9) by fitting over a blurred background or cropping, clips to join after it, music and its volume, and text on screen. Ocadox downloads FFmpeg once on Windows (115 MB); on Linux install it with sudo apt install ffmpeg.

Edit by transcript and overdub

Tools > Edit by transcript (or Transcript under an MP4 in the chat). Whisper writes down every word with its time, on this PC; it needs a speech model, set up once in Settings > Voice.

In Edit video you can also cut out parts from the middle by time, and add timed text that shows between chosen seconds, at the top, middle or bottom.

Long video to Shorts

Tools > Long video to Shorts (or Shorts under an MP4 in the chat). Choose the video, how many Shorts (up to 10) and the longest each can be, then press Find the best moments. Whisper writes down what's said and your chosen model picks the moments that work on their own: a hook, one complete idea, a clear ending. It needs a speech model, set up once in Settings > Voice.

Free and Pro

Every creative tool works in the free version, up to a size: talking avatars up to 15 seconds, songs up to 30 seconds in fast quality, videos up to 10 seconds, dubbing videos up to 1 minute long, 2 Shorts at a time, and videos edited by transcript up to 2 minutes. Ocadox Pro (or the 7-day trial) has no limits. Each window says what the free version allows. Everything else in the free version, from the coding agent to images, speech in your voice and dictation, has no limit.

Talking avatar

Set it up once in Settings > Image creation > Talking avatar (about 15 GB: Wan 2.2 S2V 14B, its audio encoder and VAE, all Apache 2.0; it shares video's text encoder). It needs a graphics card with 16 GB or more; 12 GB works slowly. Then open Tools > Talking avatar: choose a photo of you (or one of your characters), write what to say in your saved voice or choose a recording, pick vertical or wide, and confirm the face is yours or used with permission. The speech is made in 5-second parts from the same photo that dissolve into each other over the same words, with captions if you like, and saved in the project's Avatars folder. Each 5 seconds takes several minutes on a 24 GB card with nothing else using it, and much longer if another program holds the card's memory. For speed, add Talking avatar: fast mode in the same place (1.2 GB, lightx2v's 4-step Lightning LoRA, Apache 2.0) and choose Fast: 4 steps instead of 20, about 5 times quicker, a little less lively. Label it as AI where you post it.

Design

Open Tools > Design to make posts, YouTube thumbnails, stories, posters and banners. The panels on the left hold Templates (66, in categories from YouTube thumbnails to flyers and menus, with search), Elements (shapes, photo frames and backgrounds), Text (styles and font pairs), Uploads (or drag pictures onto the page), AI, your Brand kit and your saved Projects. Click anything on the page and the bar above it changes to its font, size, colour, effects, transparency and position; Delete removes it, Ctrl+D copies it, Ctrl+Z undoes, and the pink lines show when it's centred. Resize changes the page size and keeps your design. Share downloads a PNG or JPG, saves the design in the project's Designs folder to edit later (Ctrl+S), or puts it in the chat.

Design it for me, at the top of Templates: describe what you want (“a flyer for my bakery's weekend sale”) and a model writes the words and picks a mood, then designer layouts (split, centred, framed, photo band, colour block, badge, editorial, card) arrange them in matching colours, sized to fit. You get six; Shuffle shows more with the same words, and changing the size redraws them. Pick the model under Writes with: one on your PC or a cloud model. Click one to open it and change anything. Tick Also paint a background picture and Image creation paints a background to match. Even small models do well, since they only write the words. Tick Paint the photos too and each photo frame gets the photo the model described, made on your PC. With Pro, Use my brand kit makes them in your brand's colours and font, with your logo.

Pages: at the bottom, + Page, Duplicate and Delete make carousels, menus and multi-page flyers. Share downloads a PNG or JPG of the page you're on, a PDF of every page, every page as PNGs, or an animated MP4: pick Rise, Fade, Pop or Slide and the seconds per page, and things come in one after another, every page in turn (made on your PC with FFmpeg).

The AI panel works on your PC: Make a background or Make a picture from a description (Image creation), Remove background from a selected photo, Change the picture by asking, and Upscale (Picture editing), and Write the text with the chat's model. Free has 14 of the 66 templates and puts a small “Made with Ocadox” on downloads; Pro has every template, the brand kit (your colours, font and logo) and clean downloads.

Notes

Open Tools > Notes (or Ctrl+K, “Notes”). Pages sit in a tree on the left, pages inside pages, with favourites, search and a Trash you can put pages back from. Every page is a plain Markdown file in the project's Notes folder, so your notes are yours: searchable, easy to back up and fine to commit. Pictures go in Notes/_files.

In a page, type / for blocks: headings, to-do lists, bulleted and numbered lists, toggles, callouts, quotes, code, tables, pictures, dividers and today's date. Type [[ to link another page (or make a new one); the linked page shows Linked from at the bottom, and links follow a rename. The handle at the left of a block moves it up or down, duplicates it, turns it into another kind of block or deletes it.

Writing help uses the model picked at the top (one on your PC or a cloud model): select text and press AI to improve it, fix it, shorten or lengthen it, simplify it, make it professional or friendly, translate it, summarize it, find the action items, explain it or continue it. / Write with AI drafts a page from a description.

Databases (Pro) are pages whose pages are rows, with properties (select, tags, date, checkbox, number, person and link) at the top of each row. See them as a table (sort, filter, add properties) or a board grouped by a select property, and drag cards between columns. Ask your notes (Pro) answers a question from everything you've written and lists the pages it used. Templates: Meeting notes, To-do list and Daily journal are free; Project plan, Task board, Content calendar, Reading list, Weekly review and Brainstorm come with Pro.

The agent can read and search your notes; with Pro it and your bots can write them too (asking first when Ask before each edit is on). Meeting notes are also saved as a page under Meetings, and the ⋯ menu of a page puts it in the chat or sends it to Poster.

Presenter videos

To be on screen while you show something, like a Loom video: make the recording first (Tools > Record Ocadox, Record a website or Record a demo), then open Tools > Presenter video (or tick Put me on a screen recording in Talking avatar). Choose your photo, write what you'd say over the recording, and pick the recording (a file, or the video in the chat), the corner, the size, and a circle or square. Ocadox makes your talking avatar and sets it in that corner, with your voice under the whole video and captions across the bottom. The video keeps the recording's shape; if you talk longer than the recording, it holds the last frame until you finish. Saved in the project's Avatars folder. Free: up to 15 seconds of you talking.

Making music

Set it up once in Settings > Image creation > Music (about 5.7 GB: the ACE-Step 1.5 models and the acestep.cpp engine, both MIT licensed). Then open Tools > Make music: describe the style, mood and instruments, and choose Instrumental, Write lyrics for me or My lyrics (mark parts with [Verse] and [Chorus]), the length (30 seconds to 4 minutes) and the language it's sung in. Choose Fast or Best quality: best uses ACE-Step's XL model and its 4B planner (one more 9.8 GB download, under Music: best quality in Settings) for fuller, more detailed songs. Your chat model first turns a short idea into a full description with a tempo that suits the style. A song takes under a minute on a good graphics card. Songs play in the chat, and is saved as an MP3 in the project's Music folder. Use it as the music in Edit a video. The agent can make music too, asking first.

Dubbing

Tools > Dub a video (or Dub under an MP4 in the chat). Pick the language (English, Spanish, French, German, Italian, Portuguese, Russian, Japanese, Korean or Chinese) and a saved voice. Whisper writes down what's said, your chosen model translates it, and it's spoken in that voice right where the original was, sped up a little if a part runs long. Keep the original sound quietly underneath, and add captions in the new language, if you like. A picked video's dub is saved beside it as "<name> (Spanish).mp4". It needs voice input (Settings > Voice) and speech in your voice (Settings > Image creation > Speech). Only use your own voice, or one you have permission for.

Meeting notes

Tools > Meeting notes. Record from the microphone (up to 4 hours, saved in the project's Meetings folder as it goes), or Choose a recording. Press Write the notes: Whisper writes down what's said and your model writes a summary, key points with their times, decisions and action items. They're saved as "<name> notes.md" beside the recording, with the transcript under them, and shown in the chat. Tick Put the action items on this project's board to add each one as a card. Ask everyone's permission before recording them.

Dictate anywhere

Turn it on in Settings > Voice. In any app, press Ctrl+Alt+Space (or your own shortcut), talk, and press it again: what you said is typed where your cursor is. It's pasted using the clipboard, and whatever you had copied is put back straight after. It works while Ocadox is open or in the tray. On a Mac, allow Ocadox under Privacy & Security > Accessibility; on Linux install xdotool (X11) or wtype (Wayland), or the text is left on the clipboard to paste.

Recording a website

Record a website… in the command palette: type the address, pick the length, desktop or phone size, and whether to scroll to the bottom. It opens in a private browser of its own (no sign-ins, nothing shared with your browser) and saves an MP4 you can edit or voice over.

Books

Tools > Make a book, for every kind of book: novels, short stories, poetry, screenplays and stage plays; memoir, biography and history; nonfiction guides, cookbooks, textbooks and workbooks, and essays; and picture books for children. Say what it's about, pick the kind, the number of parts and the page size, and your chat's model plans a title and an outline. Each kind is written and laid out its own way: poems keep their lines, scripts are in script format, recipes have ingredient lists and numbered steps, lessons end with exercises and answers. Then, for each part, or all at once:

Export PDF lays the book out at its page size (8.5 × 8.5, 8 × 10, 8.5 × 11, 6 × 9, 5.5 × 8.5 or 5 × 8 inches, all Amazon KDP sizes) with a cover, title and copyright pages, contents and page numbers, ready to print or upload as a paperback interior; it needs Chrome or Edge. Export EPUB makes an e-book for Kindle, Apple Books and Google Play. Post about it drafts a post with the cover in the Poster. Everything is in the project's Books folder. The agent and bots can add chapters and export too (book_add_part, book_export): "every morning, write the next chapter of my book".

History, biography and other true stories

History, biography, nonfiction, textbooks and essays get Sources and accuracy. Research searches the web for each chapter and keeps the best pages as its sources (only with Settings > Web access on); Add a file (PDF, Word or text) and Add notes add your own, which stay on this PC. Chapters are then written from the sources and cite them by number, like [2], and the PDF and EPUB end with the numbered list. Fact-check reads each chapter against its sources and flags dates, names, numbers and quotes they don't back up. History and biography books also get a Timeline page. The fact-check helps, but it isn't a guarantee: check flagged claims, and every quote, yourself.

Selling a book made with AI? Amazon KDP asks you to say when text or pictures are AI-generated.

Characters (AI influencers)

The Characters button in the top bar. Describe what your character looks like and press Make a portrait (or use a picture you have the rights to); add a personality, a saved voice and the Poster brand to post as. You'll be asked to confirm the face is made up, yours, or used with the person's permission.

Free has 1 character, Pro 5. Videos post to Mastodon, Discord (up to 10 MB) and Bluesky.

The Poster

The Poster button in the top bar writes, schedules and publishes posts.

  1. Accounts: connect each place, following the short instructions there (Mastodon: type your server and press Sign in with Mastodon; the others have a button that opens the right page: Bluesky an app password, Discord and Slack a webhook link, X your developer app's keys, Reddit a "script" app). Press Test to check it.
  2. Write a post: type it, or describe it and press Draft with AI (your local model). The counter shows the shortest limit among the places you picked (X 280, Bluesky 300). Attach an image, pick the accounts, then Post now or set a time and Schedule.
  3. Queue: posts waiting for your OK (from bots), scheduled ones, and what went out, with links. Failures can be retried.

Bots: ask a bot or the agent to "write a post about this release" and it drafts one into the queue. With Pro, tick Bots may post on an account to let them post there without asking (never Reddit, which bans automatic posting). A routine such as "every Monday, post what we shipped" makes it a weekly habit.

Limits: Free has 1 brand and 20 posts a month to Bluesky, Mastodon, Discord and Slack. Pro has 5 brands and 100 posts a month, plus X and Reddit. A post to three accounts counts three; the count resets on the 1st.

The board

Tools > Board opens a Trello-style board for the chat's folder, kept on this PC (not in the project, so it never ends up in git). Add cards with + Add a card, drag them between columns, rename or add columns, and click a card for notes, labels, a due date and a checklist.

LM Studio in the model library

With LM Studio's server added under Settings > Local servers, the model library's LM Studio tab lists its models with size, context and whether they fit your graphics card, with Load, Unload and Download (a catalog name or a Hugging Face link). Its loaded models show in the In VRAM panel, and ones Ocadox used are unloaded when idle, like Ollama's. Needs LM Studio 0.4 or later. The Cloud tab lists OpenRouter's models with prices, and your Codex.

Ocadox in Claude Code, Cursor and Codex

Other agents can use Ocadox's local tools as an MCP server: pictures, picture edits and background removal, upscaling, video, voice-overs in your saved voice, long video to Shorts, music, website recordings and (with Pro) transcription. They run on this PC with the models you set up in Ocadox. Open Settings > Use Ocadox from Claude Code, Cursor or Codex and copy the line for yours, for example:

claude mcp add ocadox -- "C:\Program Files\Ocadox\Ocadox.exe" mcp

The other agent asks you before each tool call, as it does for any MCP tool, and files are read and saved in the project it's working in. Ocadox's fan guard still applies to the heavy tools. From a terminal, ocadox mcp --help shows the options.

Bots for creators

New bot templates set up common jobs in one click (their weekly routines come with Pro): Shorts channel turns new long videos into Shorts and drafts the posts; Podcast publisher writes show notes and chapters for each episode and makes clips; Music for my videos makes a background track to fit each new video; Website demo maker records your site, voices it over and drafts a post. Posts always wait for your OK unless you let bots post.

Git and GitHub

Tools > Git works on the chat's folder:

Signing in to GitHub uses what Git already has: on Windows, Git for Windows' own sign-in opens in your browser the first time you push or press Sign in with GitHub; the GitHub CLI's login (gh auth login) works too. Ocadox keeps no GitHub password or token of its own.

Cloud models (OpenRouter or your own key)

For the odd task a local model can't handle. Settings > Cloud models: turn them on, keep OpenRouter as the provider and press Sign in with OpenRouter; approve in your browser, then pick models from the list (search, prices per million tokens, and Agent-ready for models that can use the agent's tools). Or choose OpenAI, Anthropic or another OpenAI-compatible provider, paste your API key and the model names you want. They appear in the model list marked CLOUD. When you use one, your messages and the files the agent reads go to that provider, and it bills your account. The key is kept only on your PC. Off until you add a key.

Context: a cloud model gets its real context window (as OpenRouter reports it, or 128K tokens if the provider doesn't say; Claude Code and Codex 200K). The whole conversation goes with each message, and older turns are compacted into a summary only at about half the window. Local models keep a small budget (compacting past about 16,000 characters) so the graphics card stays quiet. Type /compact to summarise a chat now; on a paid key it makes each following message cheaper, since the whole conversation is sent every time.

Anthropic key: with Anthropic as the provider, Ocadox uses Anthropic's own API. The conversation is cached, so the repeated part of a long chat is billed at a fraction of the price; agent runs keep Claude's reasoning between steps; and current Claude models get their full 1M-token window. Spending estimates each reply's cost from Anthropic's list prices, including cache reads and writes.

Spending: under each cloud reply Ocadox shows the tokens it used, how many came from the provider's cache, and what it cost (OpenRouter reports the exact cost), with the chat's total. Settings > Cloud models > Spending shows this month by model and the months before, and lets you set a monthly limit: once it's reached, cloud requests stop until next month or until you raise it, and local models keep working. Claude models on OpenRouter reuse the conversation from the cache, which makes long chats much cheaper. Claude Code with your Anthropic Console key is counted too, from the cost it reports for each run, and stops at the same limit; a Claude plan and Codex are subscriptions and aren't counted.

Where everything is: the Tools menu at the top has video, picture editing, voice-overs, recording, Characters, the Poster and Git; Ctrl+K finds anything by name.

Local servers (LM Studio, llama.cpp and more)

Settings > Local servers > Find servers checks the usual ports on this PC for LM Studio, llama.cpp's server, vLLM, Jan, KoboldCpp and text-generation-webui (start the server first; in LM Studio: Developer > Start server). Or add one by address, including another computer on your network (like http://192.168.1.20:1234/v1). Their models appear in the list as local:server/model, without the CLOUD tag; the fan guard and power cap apply as with Ollama.

Use your Codex (ChatGPT plan)

With a ChatGPT Plus or Pro plan, you can use OpenAI's Codex as a model on your plan's limits. Install Codex once (npm install -g @openai/codex, needs Node.js), then Settings > Cloud models > Sign in with ChatGPT; Codex's own sign-in opens in your browser. Its models appear in the list marked CLOUD. In agent mode Codex works in the chat's folder and asks through Ocadox before anything risky; in plain chat it only reads.

Use Claude (your Claude Code, with your Anthropic key)

Install Claude Code from Anthropic (or npm install -g @anthropic-ai/claude-code), then open Settings > Cloud models > Use Claude. The recommended way in is an Anthropic Console API key: make one at console.anthropic.com, paste it, and Claude Code uses it, billed per use to your Console account. The key stays on your PC. Signing in with a Claude plan is offered as an alternative, in Claude Code's own window (Ocadox never sees it); Anthropic's rules for using a plan through other apps have changed several times, so it may be limited or billed as extra usage. Then claude-code:opus, claude-code:sonnet and claude-code:haiku appear in the model list, marked CLOUD. In agent mode Claude Code edits files in the chat's folder, and asks through Ocadox before commands, the web or anything outside the folder; in plain chat it has no tools. You can also use Anthropic's models directly with your key under Cloud models, without Claude Code.

Recording Ocadox itself

Tools > Record Ocadox records the app's window (the red timer at the top stops it) and puts the video in the chat. Tools > Record a demo asks for a task, records while the agent does it, and stops when it's done. On a Mac, allow Ocadox under System Settings > Privacy & Security > Screen Recording the first time. Some Linux desktops (Wayland) don't let apps record the screen; use an X11 session or the desktop's own recorder.

Something's wrong

Settings > Help > Send feedback or report a bug copies a short report of your setup (system, graphics, memory, CPU mode, Ollama and the loaded models, the last errors) and opens a GitHub form; paste it in and say what happened. It has no names, folders or chats, and you see all of it before sending.