Copilot mode

For: Everyone — this is the agentic platform at full volume

Cypress is agentic-first: the assistant isn't a chat window bolted to the side, it's a co-driver for the screen you're already on. ⌘K engages Copilot mode. This chapter is the full tour.

The stage

When Copilot engages, the app rearranges around the conversation:

• The Cedar Bar docks (drag it to any edge) carrying the conversation and the orb — Cedar's state at a glance: breathing when idle, pulsing while listening, swirling while thinking, pumping while speaking. Voice can auto-start; the orb always tells you when the mic is live.
• The Work Surface opens alongside, with three lanes:
  1. Proposals — every write Cedar wants to make, as cards showing the exact change, waiting for your confirm.
  2. The step trace — each tool call as it runs, live. You watch the work happen instead of trusting a spinner.
  3. Result cards — answers rendered to fit: a valuation card, a fleet summary, an inspection readout, an appraisal.
• The rest of the app dims slightly — sport mode. Esc hands the wheel back instantly, exactly where you were.

What "reading your screen" means

Copilot knows the room you're in and briefs itself on entry: land on Post-Sale and it opens with what's waiting on you; switch to the Seller Hub and its context is your grid — current filter, selection, sort. With a vehicle panel open it binds to that vehicle, so "value this one" or "draft a counter here" cannot hit the wrong car. On the Buyer Hub, the grid it reads is offers.

And it drives: navigation, filters, row selection, the bulk bar, exports — all tool calls you can watch in the trace.

A session, end to end

Maria, Monday morning, Seller Hub open. ⌘K.

"What needs me today?" — Cedar briefs: two offers in review, one counter expiring at 3pm, three vans stale past three weeks.

"Show the stale ones." — the grid filters itself; three rows.

"Ask Sequoia if we should cut prices." — a market card per van: two say hold (segment trending up), one says cut.

"Cut that one 4% and pause the oldest." — two proposal cards appear: a price change with the new bands and Sequoia delta, and a pause. Maria confirms both. The trace shows each apply; the Activity tab records automation (confirmed by Maria).

"Remind me about the counter at 2." — done. Esc. Four minutes, nothing left hanging.

The specialists on the surface

Copilot is Cedar orchestrating; the specialists surface as the cards: Sequoia for anything market ("is this priced right?"), Sakura for selling and buyers ("which channel for this van?"), Sapele when the room is a closing ("any problems in these documents?"). Ask in plain language — routing is Cedar's job, not yours.

Trust, always-on

Copilot mode changes volume, not rules:

• Reads are free; writes are proposals. Nothing mutates without your confirm — money and legal actions always card, even mid-flow.
• Your permissions, not the agent's. Copilot proposing a launch to someone who can't launch goes nowhere.
• Everything lands in the audit trail as automation with you named as the confirming human.
• Esc is sacred — one key, wheel back, no cleanup.

For teams with mixed comfort levels

The copilot is opt-in by design: the whole platform works classically — every button, grid, and form exists without Cedar. A new hire can click through the app on day one and adopt the copilot on day thirty; a power operator can live in ⌘K. Both are first-class. The generational on-ramp is that there isn't one: nothing to learn before the first question, and the first question can simply be "what does this screen do?"

Under the hood

One Cedar brain serves the omnibox (⌘J), Copilot (⌘K), Telegram, and voice — capability-gated per channel, with proposals persisting server-side so a confirmation from any surface resolves the same action. Bounded steps per turn; deliberate refusals over silent guesses.