Cedar & the AI team
The assistant layer: Cedar, its specialists, the propose-and-confirm rail, and the channels it lives on.
TL;DRAI summary
Cypress uses an AI assistant, Cedar, to help you manage your fleet. Cedar can answer questions, drive the app for you, and consult specialist assistants, but it always asks you to confirm any changes before making them.
For: Everyone
Cypress ships with an assistant layer you'll meet on every page. This chapter names the cast properly, explains the one safety rule that governs all of it, and shows what working with Cedar actually looks like.
The cast
- Cedar — the fleet advisor and orchestrator. The omnibox in the header (⌘J) and the Copilot (⌘K) are both Cedar. It answers questions, reads your current screen, drives the app on your behalf, and consults the specialists.
- Sequoia — market intelligence and asset planning: comparables, trends, the market position and the quarter plan. For a fleet with a live plan, Sequoia's verdict is the one sell-or-hold voice across Sell (the plan in Sell).
- Sakura — sales: listing readiness and channel advice on your side; the vehicle Q&A assistant on the buyer's side.
- Sapele — closing compliance: reads deal documents and flags problems before they cost money.
Orbis and ARVIS are engines, not agents — the valuation computer and the photo-inspection pipeline. You never chat with an engine; the agents and the app pull from them.
These are Cypress's own assistants. Agents your organization connects are different: an app you approved adds them to your team as members with the Agent role, and they work through the API under the rules for agents (see Apps, agents and MCP).
The one rule: propose → confirm
Cedar never writes without you. Reading — summarizing your fleet, finding stale listings, explaining an offer — happens freely. Any change — a price, a status, a launch, an offer decision — becomes a proposal card showing exactly what would happen, and executes only when you confirm. Declines cost nothing. Proposals expire if ignored. And your role still gates everything: Cedar proposing a launch to someone who can't launch goes nowhere — the permission is yours, not the agent's.
Every confirmed action is audited as automation with you as the confirming human — the Activity trail shows both.
Copilot mode
⌘K turns Cedar into a copilot that works your current screen — and rearranges the screen to work alongside you:
- The Cedar Bar docks with the conversation and the orb — Cedar's status in one glance: breathing when idle, pulsing when listening, swirling while it thinks, pumping while it speaks. Drag it to whichever edge suits you; voice starts automatically if you've left it on.
- The Work Surface opens on the right with three lanes: proposals awaiting your confirmation, the live step trace (each tool call as it runs — you watch Cedar work, not wait on a spinner), and result cards — a valuation card, a fleet summary, an inspection readout — generated to fit the answer.
- Screen awareness — on the Seller Hub it reads the grid you're looking at; with a vehicle panel open it knows which vehicle without being told (it literally cannot target the wrong one); on the Buyer Hub the grid it reads is offers. Moving rooms re-briefs it — land on Post-Sale and it opens with what's waiting on you there.
- It drives, you watch: navigate, filter, select rows, run the bulk bar, export the view — "show me live vans over 30 days on market, sort by offers" is a command, not a search query. Esc hands the wheel back instantly.
Beyond the browser
Cedar answers on Telegram too — link your account once and the same assistant (same permissions, same propose-confirm) rides in your pocket. Photo-of-VIN works: snap a windshield, get the vehicle. Confirmations are explicit there as well — a proposal answered "YES" in chat is the same gate as the card in the app.
What to actually ask
- "Which of my listed vans have no offers after 14 days?" — instant answer with rows.
- "Hold the two oldest" — proposal card, confirm, done, audited.
- "Why is this valued at $31,400?" — Sequoia's reasoning, in context.
- "Build me an automation that pauses anything 45 days stale" — Cedar drafts an automation you review on the canvas.
- On a vehicle: "draft a fair counter for this offer" — a suggested number with the reasoning, yours to send or ignore.
Worked example
Maria asks Cedar which listed vehicles have no offers in two weeks — three rows, including a box truck at $41,000. "Drop it to $39,500." A proposal card shows the change, the new Sequoia delta, and the re-armed bands; Maria confirms. The Activity tab reads automation (confirmed by Maria) · listing price $41,000 → $39,500, the Updates feed tells the story to anyone watching, and the bands now answer at the new numbers. Total elapsed: forty seconds.
Business rules
- Cedar caps its own autonomy: a bounded number of steps per request, and anything touching money or legal state always confirms.
- What Cedar knows is what your seat knows — it reads with your permissions, never around them.
- If AI services are degraded, Cedar says so and falls back to simpler reads — it doesn't guess silently.
Under the hood
One Cedar core serves every channel (browser, Telegram, voice) with capability gates per channel; proposals persist server-side and resolve through one path regardless of where you confirm. Specialists are separate agents Cedar consults; engines are services. The automations chapter covers Cedar-as-automation-author.