Autonomous
Intelligence
Runtime.
Not a chatbot. Not a copilot. Not an assistant you have to babysit. Cuttlefish is the execution environment where intelligence lives, reasons, learns, governs itself, and compounds over time — on your machine, under your control.
Mobile Remote controls your desktop. Install the desktop app first, then pair from Settings → Mobile Access.
Download
Under 2 minutes.
Connect
Your API key. Your keychain.
Work
Governed. Signed. Yours.
You stop instructing.
It starts understanding.
Every other tool on the market runs in a loop. You prompt. It responds. You correct. It tries again. You remind it what you said three messages ago. You re-explain the context. You manage the conversation like a job.
Cuttlefish doesn't work that way. It maintains a persistent model of your work — your projects, your people, your decisions, your contradictions, your predictions and whether they came true. It adapts in real time. It adjusts course without being told. When you come back after a week, it knows what changed while you were gone.
No more re-explaining context
A 9-layer context pipeline assembles everything the runtime knows before every response — your identity, your active surface, your memory, your knowledge graph, what you were doing in other tabs five minutes ago, and retrieved knowledge. The top facts about you are always present. You don't repeat yourself.
No more prompt loops
Cuttlefish doesn't run in a generate-check-retry loop. It reasons against a structured world model with real dependencies, predictions, and evidence. When it encounters a contradiction, it flags it. When a prediction is confirmed or invalidated, it learns. The system gets more accurate the longer you use it.
No more babysitting agents
Inline workers spawn with immutable contracts — budgets, tool permissions, success criteria, escalation policies. They execute on their own graphs, report back with verified artifacts and confidence-scored claims, and request approval only when governance requires it. You supervise outcomes, not keystrokes.
No more forgetting
Persistent semantic memory, procedural rules, and episodic summaries survive across every session. Entities — people, projects, topics, documents — are extracted from your conversations automatically and added to a knowledge graph that grows over weeks. The runtime remembers what matters.
12 surfaces. One runtime.
Everything connected.
Cuttlefish isn't a chat window with features bolted on. It's 12 purpose-built surfaces — each one a different lens on the same underlying intelligence. Switch between conversations, missions, knowledge graphs, team orchestrations, voice sessions, browser automations, and more. The runtime carries context across all of them.
The intelligence follows you across surfaces
Start a conversation in the Inbox. Mention a project. The runtime extracts it into your knowledge graph. Switch to the Life tab — the mission cockpit already shows dependencies, predictions, and open decisions related to that project. Open the Graph tab — the entity is there, connected to people, documents, and workflows you've discussed before. Open Quick Input from anywhere on your desktop — the World Model focus summary is right there, scoped to what matters now. Nothing gets lost between tabs.
Work that persists across days, weeks, and months
Define a mission in the Life tab. Give it an objective, success signals, and a wake pattern. The runtime tracks it as a durable unit of work with its own trust level, autonomy grants, decision rail, and story timeline. When it needs your attention, it surfaces. When it's healthy, it stays quiet. Come back in three weeks — the mission cockpit shows what changed, what's blocked, and what decisions are waiting on you. The runtime doesn't forget goals just because the conversation ended.
Agents that work independently, not in a loop
Ask a complex question. The runtime spawns specialized workers — each with its own execution graph, budget, and tool access. They research in parallel, report findings as verified artifacts and confidence-scored claims, request capability leases when they need access to something, and escalate only when governance requires it. You watch them work in real time through status cards. You approve results, not individual steps. When they're done, their output materializes directly in your conversation — charts, dashboards, interactive apps, all rendered inline.
Governed by architecture, not by restriction
Every action passes through a four-layer preflight before it runs. Every decision is cryptographically signed and stored immutably for seven years. But you don't experience governance as friction. You experience it as trust. The runtime earns autonomy through demonstrated competence — four trust levels from "always ask" to "handle all" — and you set the threshold per mission, per capability, per context. The more it proves, the more you can delegate. That's not restriction. That's a system you can actually rely on.
Talk to it on Monday.
It remembers Friday.
The runtime maintains a persistent understanding of your world. Your projects. Your people. The decisions you've made. The contradictions between what you assumed and what actually happened. The predictions that came true and the ones that didn't. It's not context — it's comprehension.
You ask a question at 3pm. It knows what happened at 9am.
You ran a browser automation this morning that extracted pricing data. You discussed the results with a multi-agent team at noon. You updated a mission in the Life tab an hour ago. Now you ask a question in the Inbox. The runtime doesn't need you to explain any of that. It already tracked the dependencies, recorded the team's conclusions, and linked them to the mission. Your question gets an answer that accounts for all of it.
You leave for a week. The runtime keeps working.
Your scheduled automations fire overnight. Your reactive triggers respond to file changes and system events. Your missions stay alive with wake patterns that know when to surface. When you come back, the Runs tab shows a signed, verifiable record of everything that happened while you were gone. The Life tab shows which missions moved, which ones are blocked, and which decisions are waiting on you. Nothing fell through the cracks because nothing was waiting for a prompt.
You say it once. It never asks again.
Tell the runtime you prefer TypeScript with detailed comments. That you're working on a Kubernetes migration. That your infrastructure lead is on vacation until Thursday. These become persistent facts — weighted, scored, and injected into every interaction for as long as they're true. The runtime adapts its communication style, its technical depth, its tone, and its assumptions to who you are. Not because you set a preference toggle. Because it learned.
Speak. Type. Point.
It operates everywhere you do.
Voice that isn't an afterthought
Three modes. Push-to-talk when you want control. Continuous when you want hands-free. Realtime API when you want sub-second bidirectional conversation with native function calling. The runtime governs voice actions the same way it governs everything else. Say "hey cuttlefish" and it wakes up.
Your phone is a first-class interface
Scan a QR code. Pair in 15 seconds. Send commands from your phone through a cryptographically signed relay. High-risk actions still require desktop approval. Offline queue stores commands when you lose signal and sends them when you reconnect. Up to three paired devices.
Browser, desktop, and shell
The runtime sees your screen, controls a headless browser, operates your mouse and keyboard, and executes terminal commands. Confidence scoring determines when actions auto-execute and when they pause for your approval. Human takeover is always one click away.
62+ connectors. 10 providers. Your keychain.
Notion, GitHub, Slack, Spotify, Stripe, MongoDB, Home Assistant, and dozens more. Each connector runs locally with per-tool consent. Your API keys live in your OS keychain — macOS Keychain, Windows Credential Manager, Linux Secret Service. Hardware-backed. Never in flat files.
Ready to take control?
No account required. No signup. Download, add your API key, and start working in under 2 minutes.
Install Cuttlefish
v0.2.144
Desktop
Mobile
Install on iOS
- 1Open app.getcuttlefish.app in Safari
- 2Tap the Share button in the toolbar
- 3Tap Add to Home Screen
macOS Installation
Your download has started
macOS may show "Cuttlefish can't be opened because Apple cannot check it for malicious software". This is expected for newly distributed apps.
- 1Open the .dmg and drag Cuttlefish to Applications
- 2Right-click (or Control-click) the app in Applications
Do NOT double-click — right-click to get the bypass option
- 3Click "Open" from the context menu
- 4Click "Open" again to confirm
macOS will remember your choice