Chat-only Mode (Mobile & Web) 📱💬

Up to now, the AI Agent was "desktop only," right? Now you can chat with AI on iPhone, iPad, Android, the browser, and the Mac App Store version too! Just add an API key and you're off! 🎉

← Home: ObsiKO 101 · Parent: 16 AI Agent

---

What is chat-only mode? 🤔

The desktop's full agent is a "worker" that installs and logs into the CLI (command-line tool) for Claude, Gemini, or Codex, and runs tools right on your computer. But you can't install a CLI like that on an iPhone or in a browser.

So chat-only mode skips the CLI and connects directly to each provider's (Anthropic, Google, OpenAI) API for conversation instead (this is called BYOK — Bring Your Own Key).

💡 The desktop app's engine list shows entries marked 'Chat', like Claude · Chat, too. If installing a CLI feels like a hassle, you can use chat mode on desktop as well by just entering an API key.

---

1. Enter your API key 🔑 (just once!)

1. Click the ≡ (Agent Settings) button at the top of the right-side AI panel.
2. Expand the 'API Keys' section (it's collapsed by default) and paste in a key for whichever engine you'll use — Gemini / Claude / GPT (your OpenAI key goes in the GPT field).
3. That's it! The key is stored safely — Mac/Windows use the OS keychain (safeStorage), iPhone/iPad encrypt it in the iOS keychain, and Android/web keep it in the app settings. ObsiKO never peeks at your key. 🛡️ In the browser version, turn on 'Do not remember in this browser' under the key box and the keys are cleared when you close the tab (for shared computers).
4. Bonus: the moment you save a key, that provider's model list refreshes automatically — ObsiKO pulls the latest list from the provider's API and neatly shows only the newest model of each family (no more errors from picking an outdated model! The list also auto-refreshes once a day).

Where do I get a key? (All free to sign up for — you're only billed for what you use)

• Claude: console.anthropic.com → API Keys
• Gemini: aistudio.google.com/apikey
• OpenAI: platform.openai.com/api-keys

Click the '?' help button to get the links above as buttons (on iPhone, they appear as a nice popup in the center of the screen).

🍎 Simple chats work even without a key: an iPhone or iPad on iOS 26 or later with Apple Intelligence turned on (and a Mac on macOS 26) can chat for free, with no key, using the 'On-device AI · simple chat' engine in the engine list. Your text never leaves the device. Being a small model, it has no note tools, web search or memory recall. Android doesn't support on-device AI. → On-device AI

⚠️ OpenAI note: your ChatGPT Plus subscription and API usage fees are separate. The API only works once you've topped up prepaid credit in Billing. If you see a "quota" error, check your credit balance.

Token usage shown right away 📊

Right below the key input field, ObsiKO shows the total input/output tokens this app has used calling each provider. Counts are kept separately per provider, and you can reset them with the [Reset] button. (This is separate from the numbers on the provider's own billing dashboard — it's "how much I've used within ObsiKO.")

💸 The longer the same chat goes, the less Claude costs: a chat with note tools on resends the tool descriptions, instructions and conversation each time, but Claude remembers that front part and re-reads it at a discount (usually a tenth of the normal price). The re-read part is included in the token totals. If you pause for more than a few minutes (about 5) before continuing, the memory is cleared and it's read at the full price again, and the first time it's remembered, that part costs 1.25×.

Claude   [········key········]
         Input 12,345 · Output 6,789 tokens   [Reset]

---

2. Just ask 💬

Once you've entered a key, you can chat right away like an ordinary AI chat. Replies appear slowly, like a typewriter, which makes them easy to read.

Example prompts

• "Explain quantum computing in one paragraph a middle schooler could understand."
• "Make this sentence sound more natural: …"
• "Summarize the meeting into 5 key points: …"

Under each reply bubble there's a Copy button and a Save as Note button, so you can turn any reply you like straight into a note.

🎙 You can ask out loud too: press the microphone button in the input box and talk — your words land in the input box without being sent, so fix them first if needed. Phones transcribe on the device, no key needed. On the web and computers, the microphone button only appears with an OpenAI or Gemini key, and it listens for up to 5 minutes at a time.

💡 It knows today's date: chat mode always knows the current date, so it answers accurately for questions like "this week," "recently," or "as of today."

⚡ Quick suggestion chips: above the input, chips suggest actions for the note you have open ("Summarize this note," "Find related notes," "Vault topics at a glance"). One tap when you're not sure what to ask!

---

3. Attach material for it to read 📎

Attach a vault file in the chat (via the reference button or drag-and-drop) or drop a file from outside, and it'll answer based on that.

Readable formats:

• Text files: .md · .txt · .csv · .json, etc.
• Documents: PDF, Word (.docx), Excel (.xlsx), Hangul (.hwp/.hwpx) — the body text is extracted and read (Hangul files come through as plain text, without formatting or table structure).
• Images: see 'Viewing images' below 👀

Examples

• Attach a report PDF → "Summarize this PDF in 10 lines or fewer."
• Attach a spreadsheet → "Pull out the top 3 months by sales from this table."

⚠️ Note on scanned PDFs: a scanned PDF whose text exists only as an image has no extractable text. In that case ObsiKO will honestly tell you "couldn't extract text (this may be a scan)." When that happens, turn it into text first with the app's text recognition (OCR) — right-click that PDF in the file list ▸ 'Convert to Markdown (.md)', and for a scan it offers to recognize the text (up to the first 30 pages). For photos and screenshots, right-click ▸ 'Extract text from image (OCR)'. iPhone, iPad and Android get free on-device 'Fast' recognition; elsewhere it uses 'Smart (AI vision)' recognition with an API key. → 07 Search & Navigation (Attaching the page as an image also lets the model view it via vision.)

🖼 Viewing images (vision)

Attach an image (png, jpg, gif, webp, etc.) and the model looks at it directly. It can read and explain text in photos, handwriting, tables, screenshots, and charts. (This is on by default, no setup needed.)

Examples

• A photo of a whiteboard → "Transcribe what's written on this board."
• A photo of a receipt → "Turn the items and amounts into a table."
• A chart screenshot → "Explain what trend this graph shows."

---

4. 🌐 Web search for the latest info and news

Turn on ≡ Settings → 'Allow web search' and the AI will search the internet when it needs to, then answer with source links attached. (This uses each provider's built-in search — supported by Claude, Gemini, and OpenAI.)

💡 Can I use search and note work together? — Only Claude can. With Gemini and OpenAI, turning on 'Allow note editing' drops web search from that reply (it focuses on the note tools instead). Each provider's API blocks using your own tools and their built-in web search at the same time, so there's no way around it. If you want to search for up-to-date info with Gemini or OpenAI, switch 'Allow note editing' off for a moment.

Examples

• "Give me the top 5 AI industry news stories from today, with sources."
• "The latest specs and price for product ○○, with source links."
• "Summarize recent research trends on this topic."

⚠️ Web search is billed to your own API key, and on top of token costs there can be a per-search fee — Claude charges $10 per 1,000 searches (pricing ↗); Gemini gives a monthly allowance free on the paid tier and bills after that (free-tier keys can't use search with the latest Gemini 3 models — pricing ↗); OpenAI charges at its search model's rates (pricing ↗). Prices change often, so check the official pricing pages. That's why it's off by default.

---

5. ✏️ Read, create, and organize notes directly

Turn on ≡ Settings → 'Allow note editing' and the AI will directly use tools to work with your vault notes. (When off, it's pure chat plus attached material only.)

💡 If you ask for something like "make me a mindmap" while this is off, the chat shows a single line — 'Allow note editing' needs to be on — plus an [Open agent settings] button. ObsiKO won't flip it for you: it's permission to read and write your vault, so turning it on is always your call.

Things you can do once it's on:

📖 Read & find

• "Find every note that mentions to-dos" (note search)
• "What's in the Materials folder?" (list contents)
• "Summarize ○○.md" (read it directly — including PDF, Word, Excel, Hangul)

🆕 Create & fix

• "Make me a new meeting note for today" → creates a new .md file
• "Add a line to the end of my diary" with this content → appends only to the end rather than rewriting the whole thing
• "Change 'tomorrow's meeting' to 'Friday's meeting' in this note" → replaces** just that part (safe and token-efficient)

📖 Korean dictionary & spelling

• "Check what '도안' means in the Korean dictionary and use it" → looks it up in the ObsiKO Korean Dictionary (Standard Korean Language Dictionary · Urimalsam) and uses the meaning and examples as grounds
• "Replace '기뻤다' with 즐겁다" → fits the form (즐거웠다) and particles
• "Check the spelling of this paragraph" → runs the ObsiKO spell checker (the text goes to the ObsiKO language server — only after you've confirmed the Korean (ObsiKO) check once in the spell-check panel)

🗂 Organize & view structure

• "Show me this note's table of contents" (heading outline)
• "Find the notes that point to this one" (backlinks)
• "Show my vault's most-used tags / gather notes tagged #project"
• "What topics is my vault made of?" → answers with topic clusters computed from your link-and-tag network
• "Expand the notes connected to this one" → explores neighbors along links and relations
• "Find notes with content similar to 'flow'" → semantic search finds them even when the words differ (index required → Search & Navigation)
• "Rename this memo and move it to the Archive folder" → rename/move (missing folders are created automatically)
• "Make a 2024 Materials folder"

🔗 Links, photos & recordings

• "Summarize this link https://…" → reads only links you've pasted into the chat (the AI doesn't wander off to other addresses). On the web version (in a browser), it may not be able to read a site that blocks it — the computer and phone apps can.
• "Copy out the text in this photo" → reads text from a photo in your vault (free and on-device on phones; elsewhere the AI looks at it)
• "Transcribe this recording" → on-device on iPhone and iPad; elsewhere with an OpenAI or Gemini key

🧰 A few more things it can do

• "Make me a morning routine" → it only saves the routine; you run it yourself → 15-2 Routines
• "Roll a main character for me" → it draws blanks like name, birthday, MBTI and verbal tics from the Random Box character pool, keeping any values you already set → 18-4 Random Box
• "Set up just the skeleton for episode 1 of my web novel" → it picks from the Writing Helper's structures (skeletons) and creates a skeleton note with every step intact → 16-12 Writing Helper
• "Find the conversation where we talked about X last week" → it searches your saved past conversations → 16-11 Personal Memory
• "What time is it? What's the date this Friday?" → it answers from the device's current date, time and weekday

🛡️ When the AI changes notes: the Answer only · Ask for big changes · Change right away setting at the top of the AI panel's options (≡) applies to chat as well. With the default 'Ask for big changes', it shows a plan card first before changing two or more existing notes, rewriting a whole note or moving notes, and in a chat where you've allowed web search or it read a pasted link, it asks first even when changing a single note. → 16 AI Agent

🛡️ Safety: there's deliberately no delete tool (too hard to undo). And the AI only ever operates within the current vault. It won't overwrite a file you're currently editing (unsaved) without care.

↩️ Undo (AI Change History): every file the chat AI creates or edits is snapshotted per run. Open the 'AI Change History' panel in the right sidebar to see line-by-line diffs and revert per file or per run — now on iPhone, Android, and web too.

✅ Auto-check after saving: whenever the AI saves a note, ObsiKO instantly lints it (unclosed code fences, broken tables…) and the AI fixes problems on the spot. A safety net with no extra cost.

🗺️ Mind maps, data tables, and story scripts too: ask "make a mind map of the World Cup," "turn my reading log into a kanban board," or "convert this story into a messenger script" — ObsiKO teaches the AI the exact file formats, so it produces proper .canvas mind maps, .base tables, and story-view documents. A lightweight LLM wiki works too — say "build me an LLM wiki" and it sets up the rulebook, index, and folder skeleton; in a vault that already is a wiki it follows the rules on its own. → 16-4 Building an LLM Wiki Ontology organizing works too — say "organize this folder as an ontology" and it follows the kind & relation rules (dictionary first, frontmatter only, plan approval before bulk edits). → 16-5 Zettelkasten & Ontology

🔎 Sources shown: when it answers from your vault, sentences carry source numbers like [S1], and a collapsible 'Sources for this answer' under the reply shows exactly which notes were read. Numbers the AI makes up are marked “unverified”. → AI Sources & Privacy

💡 Order tip: for something like "go through all the PDFs in the Materials folder and summarize them onto one page," where you need it to read multiple files directly, it's easier to turn on 'Allow note editing' rather than attaching every file one by one (the AI can see the list and open them itself).

---

6. What works / doesn't work in chat mode ✅🚫

✅ Works

• Conversation, reading attached/vault documents (PDF, Word, Excel, Hangul), viewing images (vision)
• Saving replies as notes
• Creating mind maps (.canvas), data tables (.base), story scripts, and a lightweight LLM wiki (note-editing toggle)
• Making charts — give it numbers and ask for a chart, and it writes a ```chart block into a note (a real chart, not a picture). Flowcharts and gantts go to Mermaid.
• AI Change History with revert, source numbers ([S1]) with "sources of this answer," auto lint check after saving
• Memory (memory proposals, auto-capture, conversation logs) at parity with desktop, plus finding past conversations and continuing saved ones (phone and web match words, not meaning; the 'On-device AI' engine doesn't recall memories) → Personal Memory
• Web search (toggle)
• Reading, creating, editing, and moving notes, creating folders (toggle)
• Semantic search · vault topic clusters · connection-graph exploration (toggle, index required)
• ObsiKO Korean Dictionary lookups · fitting word forms · spell check (toggle)
• Slash commands: type / in the input to see the built-in helper commands. You can also make your own — create _ObsiKO/commands/name.md in your vault with your prompt as the body ($ARGUMENTS is replaced with what you type after the command; a description: between --- lines shows in the list) and /name works on phone and web too. When you use a chat engine in the desktop app downloaded from the website, .claude/commands/ is read as well (phone, web and the Mac App Store version read only _ObsiKO/commands/).
• AI image generation, read-aloud (TTS), and audio/video transcription — these aren't in chat but are used directly from right-click in the editor or the command palette (uses your OpenAI/Gemini key, available on every platform including iPhone, Android and web; audio transcription on iPhone/iPad runs on the device, free, with no key). → Details
• Awareness of the current date

🚫 Not available in chat mode (desktop full agent only)

• Large-scale material collection (roaming the web and automatically piling up many notes), fetching YouTube captions
• Automatic Git integration
• Running OCR on a scanned PDF directly in the middle of a chat — instead, turn it into text first with the app's text recognition and attach that (see section 3 above)

The desktop app is a full agent that uses file tools via a CLI/SDK. If you need that, go to 16 AI Agent, 16-1 Basic Usage

---

7. When you're stuck ❓

• The '?' button at the top of the panel: sign-up links, feature limits, and differences from desktop, all at a glance. (Pops up centered on screen on iPhone.)
• Entered a key but it's not working: check that the selected engine matches the key (e.g. a Gemini key needs the Gemini engine), and for OpenAI, check whether you've topped up credit.
• The reply says "the vault is empty": pure chat can't see the vault → attach material, or turn on 'Allow note editing'.
• "The reply hit the length limit and was cut off": the answer grew too long. What arrived stays, and asking again continues from there. A safety block is reported the same way — either way, a cut-off reply is never presented as a finished one.
• The completion card says "No files changed": the AI looked around your vault but did not actually create or edit anything — e.g. the write failed, or the new content was identical to what was already there. We say so instead of hiding it.
• You saw "No response yet — trying again in n s": this appears when the connection drops for a moment or the server is busy. Leave it and it sends again by itself, up to three times. While the app is in the background (phone and web), it waits until you come back to send. If that still fails, a [Continue] button (when work was under way) or a [Try again] button appears. A block from running out of credits isn't sent again; it tells you right away.
• You saw "This run stopped because it cost roughly $20 or more": a run costing that much may have gone wrong, so it stops. This is an estimate, not the exact bill. What's saved stays there, and [Continue] carries on, but costs more.
• You saw "It kept repeating the same step, so it stopped here": the AI stops when it does the same thing over and over. What's saved stays there. Telling it a bit more about what you want may change things.
• Does [Continue] start over from the beginning?: No. Within 10 minutes of the break and with the same model, it picks up right where it stopped, without re-reading notes it already read. If the app closed or 10 minutes have passed, it tells the AI which notes it read and saved, and has it do only what's left.
• You pressed [Stop]: it stops right away, even on a phone, without waiting for a reply. If a note was being saved, only that save finishes. Requests already sent may still be charged.
• Can I give it a job on my phone and go to another app?: For a little while, yes. The screen doesn't turn itself off during a job, and it holds on a bit longer after you leave the app (on Android a notification shows meanwhile). Stay away for long and it stops, but when you come back it sends again or you can pick up with [Continue]. It's no guarantee of reaching the end, so keep the app open for long jobs.
• You saw "Stopped here": there's a limit on how many tool steps one run can take (24 if it's only looking things up, up to 48 once it actually starts writing). If the job is big, the AI stops there and then sums up what it found and what to do next. Press [Continue] to carry on. (It used to say "Done" in this situation, so it looked finished when nothing had actually been made.)
• On iPhone the input box is hidden behind the keyboard: the latest version raises the input line above the keyboard (please update).
• Does Enter send the message on my phone?: nope — on phones, Enter makes a new line, and you send with the Send button (no more accidental half-typed sends!). On iPad and desktop, Enter = send and Shift+Enter = new line.

---

At a glance 🗺️

• Enter a key: ≡ → API Keys (stored encrypted) · token usage shown right below the key
• Pure conversation: just enter a key · replies appear gradually
• Material: attach (md/PDF/Word/Excel/HWP) · images = vision
• Web search: ≡ toggle (billed) · source links
• Note tools: ≡ 'Allow note editing' → read, create, append, replace, organize, move

---

← Previous: Prompt Helper · Home · Next: AI Billing Guide →