Local‑first coding agent for Mac
Stop paying flagship prices
for a typo fix.
Exody reads every task before it runs it, and sends it to the model that actually fits — cheap and fast for the mechanical work, strong only when a problem earns it. Same agent, same terminal. Save up to 65–68% versus using the most expensive model for everything.
macOS Universal (Intel + Apple Silicon) · BYOK — Anthropic, Google, OpenAI, Kimi, xAI, DeepSeek · no subscription markup
Drag to rotate
The router picks between
Click a task — watch it route
Where the code actually runs
Your Mac. Your terminal. Nothing leased.
No cloud IDE in the loop. The agent opens a real shell on your machine, and every routing decision surfaces right where the work happens.
Why it's built this way
Four decisions, one honest bill.
01
Local‑first, not cloud‑leased
The agent runs on your Mac. Your project, terminal, and secrets never leave the machine for someone else's cloud IDE.
02
A router that pays for itself
Auto mode sends simple work to a fast model, escalates to strong on real complexity, and Auto-Ultra detects genuinely extreme tasks and picks the frontier model — with learned suppression so it never runs away. Versus sending every prompt to the most expensive model, that mix saves up to 65–68%.
03
BYOK with an honest meter
Bring keys for Anthropic, Google, OpenAI, Kimi, xAI, or DeepSeek. Session and monthly spend are tracked locally — cache discounts included, no markup hidden in the number.
04
A phone that shares the run
Pair once with a QR scan. Approve tools, send follow‑ups mid‑turn, talk over Voice Live — the Mac stays the execution authority. Now with transparent terminal and native Ask Question panels that stay in the chat history.
Personal assistant
A second brain for your day.
Not a coding agent wearing a hat. A real assistant with its own chat — mail, calendar, documents, errands, and the small research that keeps them moving.
Mail & Calendar
Connect Gmail or Outlook via MCP. The assistant reads, drafts, and schedules — but never sends without your eyes on it. You set the red line: ask every time, whitelist trusted contacts, or lock it to read-only.
Errands & Research
Book flights, compare prices, fill forms, track deliveries. The assistant has a real browser (Chrome control), web search, and persistent memory — so it remembers how you like to fly, who matters, and how you sign off.
Documents & Memory
Write docs, update spreadsheets, create canvases. Everything is saved locally in your .exody folder. The assistant learns what matters and writes it to memory automatically — so it never starts from zero.
Four rules: Summarise never relay. Propose, do not act silently. Show the thing. Remember without being asked.
Design studio
Design websites. Not just code them.
A full visual design surface inside Exody. Prompt a landing page, pick a design system, iterate with the agent — then export clean HTML/CSS or hand off to your codebase. Same local-first engine, same smart router, zero cloud IDE lock-in.
🎨 Visual Editor + Inspector
Multi-file projects with a live preview, element inspector, and visual edit mode. Select anything on the page and edit it directly — the agent writes the CSS, you see the result instantly.
📐 Design System Presets
Start from Apple, Material, or custom tokens. The agent generates a full design system with colors, typography, spacing, and components — then builds pages that stay consistent.
🖼️ Render & Animate
Generate real images for mockups and product shots. Preview CSS animations frame-by-frame with real screenshots. No placeholders — every pixel is real.
📤 Export Ready
One ZIP with clean HTML, CSS, and assets. Or keep iterating — the design project lives in .exody/designs/ alongside your code.
What's new
Shipped in the latest build.
🎯 Auto-Ultra
The router now detects genuinely extreme tasks and auto-escalates to the frontier model — with learned suppression so it never runs away. You can still pick Ultra explicitly; the agent just got smarter at knowing when you would have.
❓ Ask Question v2
Multi-question panels with up to 4 questions, free-text answers, and a persistent Q&A record that stays in the chat history. The panel no longer vanishes after you answer.
📝 Changes Card
Every file edit is shown as a diff card with author, timestamp, and undo. No more invisible writes — you see exactly what changed and can roll it back.
🖥️ Transparent Terminal
The xterm sidebar now has real transparency and vibrancy, matching the macOS aesthetic. Dark/light theme aware, with proper backdrop blur.
🗓️ Plan Mode
Read-only planning first. The agent lays out the approach, you approve it, then it implements. No surprise rewrites. Toggle with ⇧Tab.
🔌 Universal Binary
One DMG runs on both Apple Silicon and Intel Macs. Ad-hoc signed and verified, so gatekeeper warnings are manageable with a single xattr command.
Inside the flagship feature
Three tiers. One rule: earn the upgrade.
The router scores every message for real signal — architecture words, cross‑file scope, message length, UI‑implementation intent — before a single token is spent on a model. Nothing escalates by accident.
Fast
Haiku 4.5 · Luna · Gemini 3.7 Flash · Kimi K2.7 Code · Grok Build · DeepSeek V4 Flash
- Trivial edits — rename, typo, one‑line copy fix
- Explore & explain — "where is…", "what does this do"
- Short targeted fixes — under ~800 characters, no architecture terms
- Also the model for background explore sub‑agents mapping a repo
Strong
Sonnet 5 · Terra · Gemini 3.1 Pro · Kimi K3 · Grok 4.6
- Architecture & refactors — "root cause", "migrate", multi‑file scope
- Non‑trivial UI work — layout/animation changes with real implementation intent
- Provider failure fallback — a fast‑tier call that stumbles escalates automatically, never silently downgrades again
- Sticky for one short follow‑up if the thread is still clearly complex
Ultra
Fable · Sol · Opus 5 · Kimi K3 · Grok 4.6
- Auto-Ultra — escalates on genuinely extreme work, with learned suppression so it never runs away
- You can still type an explicit ultra request anytime
- Reserved for system design from scratch, deep multi‑day debugging, stack-wide rewrites
- Falls back to Strong automatically if no ultra model is configured
Versus the field
Built to cost less — not just to look nice.
Cursor is an IDE. Claude Code is a terminal agent. Both send every prompt to one model you pick by hand. Exody routes by task complexity automatically, runs local‑first, and adds a real mobile companion.
| Capability | Cursor | Claude Code | Exody |
|---|
Comparison based on publicly known capabilities of Cursor and Claude Code at time of writing; provider capabilities may change.
Beyond the desk
Pair once. The run follows you.
Scan a QR code from Settings → Mobile. From there your phone pulls the same session history, lets you approve tool calls, send follow‑ups mid‑turn, and talk to the agent over Voice Live — while the Mac keeps doing the actual work.
- → LAN pairing on the same Wi‑Fi, or a public tunnel via Settings
- → Voice Live conversational audio, not just dictation
- → The Mac stays the execution authority — nothing runs "in the cloud"
Pricing
Pay models. Not a cloud desk.
Exody is local‑first. Token spend goes straight to your providers, and the router keeps most of it on cheap models.
Bring your keys
Run Exody on your Mac with Anthropic, Google, OpenAI, Kimi, xAI, or DeepSeek credentials. You pay providers directly, and the router keeps that bill small.
- Local‑first agent and a real terminal
- Live session and monthly ledger
- Auto router: cheap by default, strong on merit
- Mobile bridge pairing
Pro
A product layer on top of BYOK: soft budget guardrails and the full companion stack.
- Everything in BYOK
- Budget alerts with a configurable ceiling
- Priority Autopilot and scheduled tasks
- Voice Live on mobile, Chrome control
Enterprise
Fleet installs, policy, and support for teams that want local agents without a cloud IDE tax.
- Custom MCP and policy hooks
- Audit‑friendly spend exports
- Dedicated onboarding, SLA options
Good to know