AI Integration (Bring Your Own Key)

Connect your own AI-provider account to speed up three workflows: parsing rental-house quotes, importing external call sheets as day briefs, and reading expense receipts. The feature is off by default — until you enable it in Settings, nothing is ever sent to any AI provider and no AI buttons appear anywhere.

You may not need an API key at all. If you use the app's MCP server with an AI client, your assistant can read receipts, invoices, call sheets, and rental quotes into the app directly — it does the reading, so no key and no on-device model is involved. See MCP_SERVER.md § Make changes. Everything below describes the in-app path, which uses your own key and works without an AI client.

How it works

You bring your own API key from Anthropic (Claude), OpenAI (GPT), or Google Gemini. Usage is billed by your provider to your key — Cine Power Planner adds no charge and has no access to your provider account.
Your key is encrypted in your browser with your account password before it is stored. The server only ever stores ciphertext it cannot decrypt; the encrypted key syncs across your devices like your other account data.
Requests go directly from your browser to your provider (for OpenAI, through a minimal authenticated relay, because OpenAI blocks direct browser calls — the relay never stores or logs your key or content).

Setup

Open Settings → Account → AI Integration (sign-in required).
Read and tick the consent note, then press Enable AI features.
Add a key for at least one provider (Add key → paste the API key + your account password, which encrypts it). The connection is tested automatically.
Pick the active provider and optionally override the model (defaults: claude-sonnet-5, gpt-4o-mini, gemini-2.5-flash).
After a reload or on a new device, unlock your keys once per session with your account password (Unlock keys for this session, or directly in the quote-parsing dialog).

One-click quote parsing (in the project Rental tab)

Open a project → Rental tab → Quote Comparison (collapsible section below the Rental Workflow panel). The workflow ("Parse Quote with AI") ships with a highlighted "Parse directly with …" box when the feature is enabled: paste the raw quote text (from the PDF or email), press Parse with AI, and the extracted items land in the comparison — validated exactly like the manual copy-paste flow, which stays available without any stored key. After parsing you can compare the quote against the project's gear list, validate items against your device library, and add missing items to the project; the parsed quote is saved with the project.

External call sheet import (in the project Call Sheets tab)

Open a project → Call Sheets tab → Import external. Paste the text of a call sheet you received from another production (or pick the PDF — the text layer is extracted locally; scanned PDFs are pointed to the OCR tool first), press Create day brief, and the AI turns it into a compact Day Brief: timeline, locations and hotels with Google-Maps links, contacts, weather-based gear warnings per location (rubber boots, winter jacket, sunscreen, …), drive times between sets, an indoor-day note, your own custom notes, a live "where should I be now" panel on the shoot day, and a Send brief to team email button. Gated by the separate External call sheet import sub-toggle (on by default once AI is enabled). Full details: CALL_SHEETS.md. Works with all cloud providers and the on-device Gemma model (text input only, so on-device is fully sufficient).

AI receipt reading

Enable AI receipt reading separately in the settings panel (off by default — each scan sends the receipt to your provider and uses your key). When on, the expense scanner (camera scan, upload, and batch drop) sends the compressed receipt image to your provider and pre-fills vendor, amount, VAT, date, and more, with the usual "needs review" badges. If the AI call fails for any reason (offline, quota, invalid key), the built-in offline extraction runs instead — scanning never breaks.

Ask Mr Watts (Help → Assistant)

The assistant is the camera mascot: the heading and the in-project sparkle button both read Ask Mr Watts, and Ask Miss Watts in the pink theme. The Help page's Assistant tab answers questions about the app, about keyboard shortcuts, and about device specifications — grounded exclusively in the app's own knowledge base, the user guides in this Documentation section, the shortcut list, and the device catalog, so it cites its sources instead of guessing. It is the one AI surface that is useful without any AI configured: with AI off it shows the matching help sections, guide sections, shortcuts, and device records directly.

Conversation: follow-up questions carry the subject forward ("…and the weight?"), the transcript survives switching Help tabs (e.g. opening a cited topic) for as long as the app is open, and is never saved to disk or synced. Copy answer and Ask again live on each turn.
Show me: cited topics that have a guided tour offer a one-click Show me button that starts the tour.
It reads the full user guides, not just the help topics. The help topics orient you on a screen; these guides are where the detail lives — permission matrices, the offload wizard's checksum pass, what a field lock does when a tab closes. All 69 of them are searched alongside the help topics and ranked on the same scale, so a guide section can outrank a help topic when it is the better answer, and the citation names the guide and the exact section it came from. The guides are written in English; the answer is still written in your language.
Device specs are shown in full. A cited device record lists every mount, connector, sensor mode and codec it has — nothing is shortened to a …. Only the copy the model reads is budgeted, and that budget stops at the model. Cameras additionally show a Data rates group: each codec's bitrate in Mbit/s, with the raster and frame rate it was quoted at ("Quoted at 4K UHD 3840×2160 @ 24 fps"). Codecs with no sourced rate are omitted rather than guessed.
It knows which screen you are on. "How do I add one here?" and "what is this tab for?" resolve to the surface in front of you: the route, the workspace section, and the name of the open project, named with the app's own breadcrumb and tab labels. The help topic for that screen is always in reach, so a question about the current screen is answerable even when it shares no words with any topic. Opening Help from elsewhere keeps referring to the screen you came from — /help itself is never recorded. A turn that used it says so underneath the answer. What travels is the screen's name, its purpose and the project's title; nothing that is on the screen — no gear, crew, rates or attachments.
It stops offering to take you where you already are. The "Take me there" button is hidden when the answer's topic is the screen on display.
Troubleshooting questions (something stuck, failing, lost) include a few session details — app version, platform, connection state — so the answer can be concrete; the turn discloses when that happened. Your project contents are never included — the open project's name is, as part of the screen you are on.
Escalation: "This didn't help — report it" opens the bug-report form pre-filled with your question and the answer.
With a cloud provider configured the assistant reads a larger slice of the knowledge base per question than with an on-device model, whose small context window gets the compact scale.

On-device model (Gemma) — Mac & iOS apps

In the Mac desktop app (and the iOS app once its native module ships), you can skip cloud providers entirely: Settings → Account → AI Integration → On-device model (Gemma) lets you download a Gemma model and select Use this model. The catalog is headlined by Gemma 4 (E2B ~3 GB for 8 GB machines, E4B ~5 GB — recommended, 12B ~7 GB for best quality) with Gemma 3 1B (~0.8 GB) as the smallest option for older hardware. Everything then runs entirely on your machine — no API key, no per-call cost, and nothing ever leaves the device. You can switch back to a cloud provider at any time; downloaded models can be deleted to free disk space. Models are provided under the Gemma Terms of Use. In the web app this section shows a hint instead — on-device models need the native apps.

Privacy

Content you submit (quote text, receipt images) goes to the provider you chose, under that provider's terms.
Keys are never stored or logged in plaintext anywhere; delete them anytime (Remove), which also deletes the server-side ciphertext.
Disabling the feature removes every AI affordance immediately; a factory reset wipes the local encrypted key cache.

Last Updated: 2026-09-08 Version: 0.790.2