AI features

Every AI feature below runs through whichever backend you selected, and each one can be turned off individually. With the default in-app backend, no prompt or email ever leaves your machine.

Choosing a backend

Settings → AI Backend & Models controls where inference happens:

• In-app — an embedded llama.cpp runtime. Nothing to install and no daemon; once the model is downloaded, answering makes no network traffic. This is the default. It uses your GPU automatically where there is one — Metal on Apple Silicon, Vulkan on Windows and Linux — and the CPU where there is not. On a Mac it requires Apple Silicon (M1 or newer); on an Intel Mac it stays unavailable.
• Ollama — an Ollama server you already run at http://localhost:11434. Useful if you keep a shared model library. Note that on an Intel Mac it gets no GPU acceleration either, so it will be slow.
• OpenRouter (remote) — a paid cloud API. Requires an API key, supports a monthly budget cap, and sends email content to a third party — so it stays off unless you enable it. Its panel shows what the current period has spent against that cap, with Start a new period to reset the count.

The model catalog

The in-app backend downloads models from a curated catalog, each pinned to a verified checksum:

The right-hand column is the memory EmailOps asks for before it offers a model — a deliberately generous margin, not what the model uses. Measured peak while answering, at the context size a 16 GB Mac gets: about 3.7 GB for Qwen 3.5 4B, 4.3 GB for its 8-bit build, 5.6 GB for Qwen 3.5 9B and 7.1 GB for Gemma 4 12B. Which memory it has to fit in depends on your hardware:

• Apple Silicon — unified memory, shared between CPU and GPU, reached through Metal. Compare the figure against your Mac’s total memory.
• A GPU on Windows or Linux — the card’s VRAM, not your system RAM, reached through Vulkan. An 8 GB card runs the 8 GB row and nothing above it, however much RAM the machine has.
• No GPU — system RAM, on the CPU. It works; it is just slower.

One model carries a Recommended badge, chosen for the machine you are on: EmailOps looks at your system memory and, if you have a discrete graphics card, at its memory too, then suggests the largest model that fits comfortably. A laptop and a workstation will therefore see different suggestions. Bigger models answer better and run slower, so the badge is a starting point rather than a rule. Full requirements are in Installation.

Performance knobs

• Keep model loaded — how long the model stays resident between turns (default 30 minutes). Higher values skip the slow reload; 0 evicts it immediately and frees the memory for other apps.
• Context window — how many tokens the model can attend to per turn. Larger fits more retrieved email, and costs more memory — this is the knob to turn down first when a model only just fits.
• Thinking mode — chain-of-thought reasoning on supported models. Slower, more accurate, and you can show or hide the reasoning trace.
• Limit AI processing — caps what embedding and classification cover: every email of an account up to an email limit (1000 by default), and for larger accounts only mail newer than a day limit (365 days by default).

Chat with your mailbox

Ask questions in natural language — “what did the lawyer say about the contract?”, “summarise this thread”, “who still owes me an answer?” — and get an answer with the source emails cited. Answers stream in as they are generated. Show in email list, under an answer, puts exactly the emails it cites into the email list, so you can open and work through them.

Chat lives in a resizable panel docked to the right of the inbox, so you can keep reading while you ask; there is also a full-page view for longer sessions. With an email open the panel offers that thread as context via a removable chip: questions about that email are answered from the thread, while a question about the rest of your mailbox (“what came in today?”) still searches it. That context applies to a single question and is never saved onto the conversation, so you can move between emails inside one chat.

Chat searches one account at a time, and a picker names which one — so an answer is never silently drawn from the wrong mailbox. Each account keeps its own conversation for as long as the app is open, so switching accounts returns you to where you left off rather than to a blank chat.

Type /clear and press Enter to start a new conversation. It is handled in the app itself — nothing is sent to the model, so it costs no time and no tokens.

Chat also answers questions about EmailOps itself — “how do I connect Ollama?”, “where is my data stored?”, “what does the Tag Board show?” — from these guides, in your language, without searching your mailbox. The answer links the guide section it used, and following the link opens the matching setting or view. With an email open as context, chat answers from that thread only, so remove the chip to ask about the app. Turn this off with Answer questions about EmailOps in Settings → AI Backend & Models; chat then only knows your mailbox.

Every answer has a Show reasoning panel that lists what happened, in order: which route the question took and what decided it, the query planner, the mailbox search, the guide sections used, each model call with its timing, and each tool call with its arguments and result.

Research mode is for questions that need every matching email rather than the few a normal answer reads — “list every invoice from this year”, “how many clients asked for a quote?”. It first estimates how many emails it would read and how long that takes, and asks before a large run. The estimate also shows the search it ran — sender, dates, words — so you can tell whether it makes sense, even when it found nothing. It then reads them in batches, and lists and counts come out exact, with a link to each matching conversation. Cancel research stops a run.

Ask for something the app has a form for — “create a Lens that tracks supplier invoices with amount and date” — and the real form opens with the fields filled in, for you to review and save. Chat never creates anything by itself.

Stop generating ends an answer while it is being written; what was already shown is kept.

Under the hood, chat combines retrieval (semantic search over your embedded mail) with tool calls (direct lookups against the database). The routing mode is configurable:

• Always RAG first — the default; retrieve context, then answer.
• Auto — a heuristic decides per question whether to retrieve first.
• Always tools first — skip retrieval and start from structured lookups.

In every mode the tools stay available; the mode only decides whether retrieval runs before the answer.

Advanced users can edit the system prompt and the retrieval prompts (query rewriting, reranking) in Settings → AI Backend & Models → Chat prompts.

AI drafts

An AI Draft button next to Reply All writes a reply grounded in the thread you are looking at. Configure a persona (one sentence on who the AI writes as) and a writing style — or replace the whole prompt template. Drafts land in the composer for you to review before anything is sent.

A draft reads the thread up to the message you are answering, never the replies that came after it. You can tell the AI what to say before it writes, and Regenerate writes it again.

Classification

Every incoming email is tagged along three axes — priority, intent and topic — so the inbox effectively sorts itself and smart filters have something to filter on. The chat reads them too: asking for urgent mail lists only mail tagged urgent, and says so when there is none.

Classification works in two layers:

1. Rules match on sender or subject patterns (*@*.beehiiv.com, *invoice*) and assign tags instantly, with no model call.
2. The model handles everything the rules do not, using an instruction prompt you can edit.

You control which Gmail categories are classified, can reclassify everything after changing the prompt, and can catch up on unclassified mail on demand.

Tag Board

The Tag Board (under Views in the sidebar, next to the inbox) turns those tags into a board. Pick one dimension — Company, Priority, Intent or Topic — and every tag value becomes a block listing its threads; in All accounts you get one block per account and tag. A thread sits in exactly one block, under the tag of its most recent classified message.

Blocks are ordered by how much attention a tag actually gets — how often you reply to and read its threads, weighted towards recent activity — with promotions and notifications ranked last. The smart filters in the sidebar follow the same order. Drag blocks to reorder them (the order is remembered per dimension), hide a tag from its ⋮ menu — the next tag moves up to take its place, and the filter leaves the sidebar too — and bring hidden tags back with the Show hidden tags link.

The toolbar narrows the board by time (Today, Yesterday, Last 7 days, or a custom date range), by Gmail category, by tag name, and with the same Hide junk messages switch as the inbox; two icons set the block width. Clicking a card opens the thread in the reading pane, its ⋮ menu offers the same actions as an inbox row, and the chat icon in the reading pane starts a conversation with that thread as context.

The board needs classification: it is empty until mail has been tagged, and it is not shown while AI features are off.

Semantic search

Emails are embedded so search can match meaning, not just keywords — describe what you remember and EmailOps finds it. This also powers the retrieval step in chat. Pick which categories get embedded, and rebuild the index from scratch after changing the embedding model, in Settings → AI Search.

With the two local backends, embedding happens on your machine. With OpenRouter it happens only if you pick an Embedding Model in its panel: the text of every email that gets indexed, and of every search and chat question, is then sent to OpenRouter and counts against the monthly budget. With none picked, semantic search is off and search is keyword-only. The model is checked when you save, and one that does not produce the 768-dimension vectors the index needs is refused.

Translation

Translate buttons appear on emails written in another language and in the compose window. The translation prompt is editable like the others.

Tasks

Experimental. EmailOps scans mail for action items, commitments and deadlines and collects them in a Tasks panel. Because real commitments usually live in what you wrote, there is a “learn only from emails I wrote” mode. You can exclude senders and tags (newsletters are excluded by default), cap tasks per email, limit how far back extraction goes, and backfill older mail on demand.

Turn it on in Settings → AI Tasks with Extract tasks from emails; the tasks it finds then appear under Tasks in the sidebar.

Memory

Experimental. Facts the assistant learns about your contacts, domains and projects are stored as long-term context so chat does not start from zero every time. Candidate facts are scored and promoted past a threshold; low-scoring ones expire. Everything it has learned is inspectable, and the whole subsystem has a master off switch.

Turn it on in Settings → AI Memory with Let the assistant remember facts; what it has learned is listed under Memory in the sidebar.

Lenses

Schema-typed views over your mailbox — saved, AI-extracted structured projections (think “all invoices with amount and due date”) that you create and run from the sidebar. A row you exclude stays out of the view; Show excluded rows brings them back into sight so you can include one again.

Lenses are on by default: create and run individual lenses from the Lenses entry in the sidebar. Turn them off in Settings → AI Lenses.

A Lens can be limited to chosen Folders of an account, custom IMAP folders included.

Skills

Experimental. A skill is a saved procedure the chat follows for one kind of request — how you want a weekly summary laid out, how to answer a supplier. Each skill is a folder in the app’s data folder, skills/<name>/SKILL.md: a short header with a name (the folder’s name) and a description of what it does and when to use it, then the instructions in Markdown. Only the name and description stay in the chat’s prompt; the instructions are read on the turns that use the skill. Nothing in a skill folder is ever run.

Skills are off by default. Turn them on in Settings → AI Skills with Enable skills; a Skills entry then appears in the sidebar.

The Skills view lists every skill with its own switch and opens the selected one’s SKILL.md in an editor. New creates a skill from a template. Save writes it back, and refuses text the chat could not load or a file another editor changed since you opened it; changing the name in the file renames the skill. Delete moves it to skills/.deleted, where it can be restored by hand.

The chat applies a skill in three ways: it picks one whose description matches your question, the assistant loads one while answering, or you start the message with / and the skill’s name (/weekly-report last week; /a /b applies both). Typing / lists your enabled skills. The reasoning trace names the skill a turn followed and who chose it.

Turning it all off

Settings → AI Backend & Models → AI Features is a master switch. Turn it off and EmailOps runs as a plain email client: no chat, no classification, no embeddings, no model loaded. Your existing local AI data is preserved in case you switch it back on.