Local AI,
right on your desktop.

A library of prompts that live as files.

Every prompt is a markdown file with YAML frontmatter in your configured folder — git-friendly, shareable, editable in any editor. Five colored favorites for personal buckets, flat tag filters, full-text search across name, body, and tags. Type / in the composer to invoke one inline.

28 built-ins ship with the app, including Album Deep Dive, Brainstorm, Cliff Notes, Cover Letter Writer, Devil's Advocate, Lateral Thinking Puzzles, Log Triage, Murder Mystery Interrogation, Notes Synthesizer, Resume Coach, Sensitive Doc Q&A, Text Adventure, Tone Reframer, Translate (ES/FR/DE), and more.

Press ⌘⇧Space. Ask anything. Never lose focus.

A transparent spotlight-style panel that pops up over any app and quietly answers — without stealing focus from your editor, your terminal, or your browser. Inline model picker, prompt picker, and streaming responses. Auto-hides on blur or esc.

Keeps its own model preference, so you can run a heavy reasoning model in the main window and a tiny fast one in the overlay. Click Send to main to escalate any overlay exchange into a full multi-turn chat in the main window.

Transparent vibrancy window · ~5ms cold-start · never steals focus

Every conversation, findable in 150ms.

Multi-tab chat with independent histories per tab. Tabs with attached files show a paperclip + count so you can see context coverage at a glance. Every word of every reply is BM25-ranked and full-text searchable via SQLite FTS5, indexed live by triggers — no batch jobs, no stale indexes.

Multi-word queries AND together, prefix matching is automatic, and clicking a result opens the parent chat with the query terms highlighted inline.

Click the lock. Leaves no trace.

Click the lock icon beside New chat in the sidebar to open an ephemeral session. Messages, attachments, file saves — none of it touches the SQLite database. The conversation lives only in memory and disappears when you close the tab.

The composer hides the paperclip and folder buttons (attachments would persist to disk), the export kebab is hidden (nothing on disk to export from), and a lock glyph appears on the tab and as a banner above the composer so you're never visually confused about which mode you're in.

Two or three models. One prompt.

Click the columns button in the sidebar, pick 2 or 3 installed models, and the chat splits into adjacent columns. Send one prompt — every model streams its response in parallel, so you can watch them think side-by-side instead of swapping tabs to re-run the same question.

When one answer is clearly better, click Keep this beneath it. The losing columns collapse, the chat transitions back to a normal single-model conversation with the kept model, and the next message goes only to it — no wasted tokens, no cluttered history.

Press ⌘⇧1. Drag a region. Ask.

Hit the screenshot hotkey from any app — Ekorbia invokes macOS's native region selector (the same crosshair you already know from ⌘⇧4). Drag a region, press Space for window mode, or esc to cancel.

The captured PNG opens in a fresh chat tab with the image already attached as a vision attachment. If your current model can't see images, Ekorbia auto-switches to one that can — and tells you which one via a toast.

Ambient background work, while you sleep.

Point Ekorbia at a folder, an RSS feed, or any public URL. It polls on its own cadence, summarises only what's changed, and appends each summary to a notes file you can chat with later. Diff-mode URL watches send only the added/removed lines to the model — ideal for changelogs and release notes.

Optional native OS notifications (Notification Center on macOS) — coalesced per poll cycle, so five new files become one banner, not five.

Attach a folder. Ask a question. Get cited answers.

Attach .md, .txt, .pdf files, images, or whole directory trees. Small files are inlined verbatim; large ones are chunked, embedded locally with nomic-embed-text, and the top-6 most relevant chunks per query are retrieved at send time.

Replies come back with inline [N] citations that map to a Sources footer of clickable chips. Folder chips expand to show which sub-files matched and their relevance scores. Shift-click any chip to reveal in Finder. Re-indexing is incremental — change one file in a 500-file folder, click ↻, only that file re-embeds.

Let the model write files, safely.

Tool-capable models (gemma3, llama 3.1+, qwen 2.5+, …) can call write_file during a chat to save any output — HTML, CSS, Python, configs, scripts — directly to a per-chat output directory you choose. First save asks permission; the choice sticks per chat.

Every write goes through a server-side sandbox that rejects .., absolute paths, NUL bytes, and symlink escapes — with unit tests for every rejection path. Writes are atomic (temp + rename), so a crash mid-write leaves the previous version intact.

Models without tool support get per-fenced-block Save buttons instead — same sandbox, same flow.

One markdown file. Every chat.

A single markdown file Ekorbia injects as system context on every chat send. Useful for habitual context you'd otherwise re-type — facts about you, project conventions, writing style, things the model needs to know but you shouldn't have to repeat.

Read-only from the model's perspective — the write_file tool can't touch it. The memory is your shared context with the model, not something it gets to rewrite on its own. Edit it in any text editor, any time.

Streaming, edit-retry, export, status.

Replies stream as they're generated; a Stop button freezes what's been written so far and marks the partial reply as Stopped so you see exactly where you cut it off. Click the pencil on any past user message to edit and resend; click retry on the last assistant reply to regenerate.

Export any chat to Markdown (clean role-tagged sections) or JSON (full row preservation with tool calls and timestamps). And the status bar distinguishes Ollama not running, model not pulled, cold, warming, loaded — no more wondering why the first token took 8 seconds.

Five themes. Pick your weather.

One Dark and One Light for the editor-theme crowd. Ayu Dark, Ayu Mirage, and Ayu Light for the warmer palette folks. Switch in Settings — every panel, chip, badge, modal, and syntax-highlighted code block follows.

Background · accent · link · keyword