Writing code stopped being the bottleneck. Managing it did.
The control plane
for your AI agent fleet.
Reef is the operational platform for teams that build software with fleets of AI agents: an idea becomes a plan on the board, the board dispatches cloud agents, agents come back with pull requests — and every dollar of it is visible the moment it is spent.
The notch
Born in the MacBook notch
Reef lives where you already look — the fleet's glanceable readout at the top of the screen. Below is a tiny Mac running a tiny Reef. It works: click the island, press A, drag the cards, switch the tabs.
Bash · npm run deploy
Drag a card into In progress — an agent picks it up and the meter starts. Click a card to move it along.
The loop
One loop, from idea to merge
PM, developers and cloud agents work in a single loop. A human steers; the fleet executes. Nothing disappears into a black box — every step is visible, priced, and reversible.
Idea
Bring a rough thought. Triton — the AI composer — pressure-tests it with you in a chat, against the real codebase.
Plan
The dialogue becomes an editable draft: epic, tasks, subtasks — reviewed, corrected, and approved on the kanban board.
Execute
Cloud agents pick tasks up and work autonomously in sandboxes. Live status, resources and spend tick on each card.
Pull request
Agents come back with a ready pull request. Auto-review flags what actually needs your eyes.
Merge
You approve — the loop closes. The board, the git history and the cost ledger all agree on what just happened.
$ the economics are visible at every step — per session, per task, per epic, the moment they occur.
Economics
Every dollar, the moment it is spent
AI spend became a real budget line — and a badly predictable one. Reef meters every session, task, run and epic as it happens: burn-rate, forecasts, budgets and thresholds, before the invoice arrives. Not a monthly report — a live instrument.
Platform
The notch is the door. Behind it — the platform.
A full-size workspace, a cloud command center, a meeting assistant, an AI product team, automations, organizations and roles — one operational platform, one design language.
Cloud command center
Every cloud agent session live: per-agent breakdown, CPU and memory curves, session replay — and a stop button that always works.
runs · resources · replayDeep Mode workspace
The notch opens into a full workspace: up to four sessions side by side, file tree, diffs, terminal, git — a post-IDE day without the IDE.
sessions · git · terminalFlow — automations
Visual pipelines for the routine: PR review, merge gates, follow-ups. A graph editor, human-approval nodes, and a git trail for every run.
DAG · wait-for-human · git trailCoral — an AI product team
A product owner for your board: watches dependencies, priorities and risks, runs discovery, and proposes what to build next — you keep the veto.
discovery · priorities · risksSonar, in your meetings
Detects the call, transcribes on your Mac, keeps live notes and follow-ups, and turns decisions into board tasks — on-device by default.
on-device · notes → tasksAuto-pilot & auto-review
Ready tasks and whole epics execute in dependency order and return as pull requests. Every PR passes an AI review with a verdict — merge by button or by rule.
autopilot · review · merge gatesOrganizations & roles
Projects, teams, six roles, access and sandbox limits. Developers see their sessions; leads see the team; finance sees the forecast.
projects · teams · accessPilot — a voice for Reef
Ask about your boards, sessions and settings, or say the command — "open a terminal for reef". Speaks and listens; risky actions still wait for your approve.
experimental · voice + textRemora — the board keeps itself
Finished work finds its card and moves it; work without a card becomes one. Every auto-action is logged, explained and reversible.
auto-file · audit trail · undoTrust
Private by architecture, not by promise
A control plane sees everything — so it is built to see it on your terms. Local-first where it matters, explicit where it connects.
On-device first
Sonar records and transcribes meetings on your Mac. Audio never leaves the machine; teammates get text only, and only when you say so.
Your data, read-only
Opt-in MCP access lets your own AI sessions read your Reef boards, plans and meetings — read-only, on your machine, revocable with one toggle.
Scoped, expiring access
Shared keys live in a vault. Cloud agents receive limited, short-lived grants — never your personal credentials, never broader than the task.
Signed & notarized
Every release is Developer-ID signed, Apple-notarized and delivered through EdDSA-signed auto-updates. Install once; stay current safely.
Seats
One loop, four seats
Reef enters an organization through the manager, not around them — the planning and reading surfaces matter as much as the developer ones. Each seat gets the view it actually needs.
Every session under control from the desktop. Approvals in seconds, parallel agents that stay out of the way, the terminal one hotkey away.
An idea becomes an editable plan in minutes — worked out in a chat, checked against the board, ready to dispatch.
Live visibility into the team's AI activity: bottlenecks, waiting sessions, and cost per person and per task.
The operational picture in real time — burn-rate, forecasts, budgets and thresholds before the invoice arrives.
FAQ
Questions, answered
Which AI CLIs does Reef work with?+
Does my code or meeting audio leave my Mac?+
How much does Reef cost?+
What do I need to run it?+
Is there a Windows, Linux or web version?+
How do we get into the private preview?+
Reef is in private preview.
We are onboarding a small group of teams that build with AI agents every day. Tell us about yours.