OFFLINE-FIRST PERSONAL AI
Your own AI, running on your machine. Teach it who you are — then hand it down.
New here? Start Here walks through everything in plain language, no jargon.
The Short Version
Cursiv talks to a local AI model (Ollama) that lives on your computer. No internet needed after the first install, no account required to use it, and nothing you type leaves your device unless you choose to add a cloud API key on top.
Anchor a conversation and it's saved to your own private, searchable archive that later answers can draw on. Over time it builds a real record of what you've thought about and worked through — not somebody else's training data.
Plain chat, not a command line. Ask a question, pull in a live web search when you need something current, dictate by voice, or generate an image — all from one window.
Write private, sealed letters for the people who matter to you — opened only with the code you set for them. A family archive that outlasts any single conversation.
Works Without The Internet
Most AI progress right now is happening inside a handful of companies' cloud infrastructure — powerful, but it requires you to stay online and hand over control. Cursiv takes the other path: local models running on your own hardware, so the core system works the same whether you're on gigabit fiber or have no connection at all. That matters most exactly where always-online tools fail — rural areas, remote work sites, or any stretch where the internet is slow, expensive, or just not there.
Track planting schedules, livestock records, yields, and expenses. Plan crop rotation or a breeding program. Build checklists and seasonal plans — entirely offline.
A tutor for homeschooling — lesson plans, explanations, practice problems. Structured study sessions for practical skills like mechanics, carpentry, or first aid.
Bookkeeping, inventory, and customer records. Draft invoices, proposals, or a grant application when the internet is only occasionally available.
Work through a big decision — moving, a career change, family planning — with the council's multiple perspectives. Keep a running record in strand memory.
Plan a shed, a water system, or a solar setup. Draft a local newsletter, a journal, or a community document. Think through a build before you start it.
Winkler-Codex writes scripts for farm automation, data logging, or device control — no cloud service required. Maintain your own knowledge base as you go.
Real-time web search and routing to Grok, ChatGPT, or Claude need a connection — everything else above doesn't. Initial setup (installing Ollama and a local model) does need internet once, and benefits from a reasonably capable machine.
Everything Cursiv holds for you — sealed letters, strand memory, the whole install — lives in ordinary files on your machine, not in an account only one company can lock or delete. Copy that folder to a USB flash drive, a Raspberry Pi, an external drive, or a new computer entirely, and it runs there the same way. No export tool, no migration wizard, no service that might not exist in ten years.
That's what makes it work as a real family legacy: something you can hand to a spouse, your kids, a close friend, or pass on with a small business — on hardware you already own, that keeps working with or without you at the keyboard.
A limited public preview, running right in this page — no install, no login. Ask it something.
The desktop app runs entirely offline on your own machine, with no rate limits and the full feature set below.
This is a lightweight web preview. Download the desktop app for the full offline experience — real streaming, voice, file tools, and every feature below.
Sealed messages for people who matter to you — a spouse, kids, close friends, extended family. Each letter opens only with the exact code set for it.
This is a vessel for care, not a weapon — sealed messages are never a place for threats, harassment, or anything meant to harm.
Unlock letters through the full desktop app, using the code made for you.
No manual required, but here's the quick version — four things worth trying first.
Type a question into the chat box above and see how it responds. No account, no install, no commitment.
The full feature list below covers the 14-agent council, offline operation, provider cascade, family letters, community board, and everything else that ships in the desktop app.
The demo chat on this website is a limited, public preview. The full desktop app is where API keys live — add your own xAI, OpenAI, or Anthropic key directly in its chat interface for larger cloud models, on top of the offline llama3.1 that ships by default. Ollama always runs first; cloud keys are an upgrade, never a requirement.
The community board and sealed letters both live in the desktop app — reachable through plain chat commands (blast, write to <name>). Download it above to unlock all of it, running fully offline on your own machine.
Capabilities
10 advisors deliberate in parallel via ThreadPoolExecutor. 4 synthesizers surface the response. Council memory — Jaccard similarity + 7-day exponential decay.
Ollama + llama3.1. Zero cloud dependency after initial download. Works airgapped. Conversation data never leaves your machine.
xAI → OpenAI → Claude → Ollama. Falls through on failure. Inline banners show the path. Force any model: hey grok, hey claude, hey ollama.
Four-layer active defense against prompt injection and jailbreak attempts. The firewall faces outward — it's not there to watch you.
bcrypt rounds=12. HMAC constant-time comparison. Security-question recovery — 2-of-3 required, bcrypt-hashed before storage. Nothing sent anywhere.
NDJSON streaming from Ollama. Tokens appear as they generate across all providers. Rate limiter with TPM window tracking and scan display.
Offline coding specialist. Explicit invocation only: codex <prompt>. Designed to get you 70% there — and force you to understand the rest.
Launcher + tray menu. GitHub releases API — background thread, no data sent. Shows release notes. In-place install. Ollama untouched. Never auto-updates.
Probes actual API endpoints. GREEN/RED/GOLD per connection. File access, Obsidian, and Guardian reflect real state. Updates after every key change.
Jaccard similarity across past deliberations. Exponential decay (7-day half-life). System surfaces relevant prior context and builds on it.
Every session streamed to your Obsidian vault in real time. Dataview-compatible YAML frontmatter. Auto-detects vault path. Toggle per session.
System Tray (Win32 IPC). Chat UI port 7860. Nexus Panel port 7861. Terminal Chat — full-screen CLI with file tools, Guardian scan, and cascade routing.
See every machine running Cursiv from one window. Live status dots, version, last-seen. Owner-gated — you control who can connect. No surveillance. Your machines, your view.
Worldwide live search built in — no API key required. Fires automatically on queries about latest news, prices, and live facts. Explicit: search: <query>
One install, several logins — a spouse or kids can each set up their own username and password on the same machine, sharing the same underlying data.
Post a thought publicly and see what others have shared — right from the chat panel. blast register, blast login, then blast to post. No separate site or browser tab needed.
Under The Hood
01 — PRIMARY
xAI Grok-3
fastest · frontier
02 — FALLBACK
OpenAI GPT-4.1
cloud · code gen
03 — FALLBACK
Claude
cloud · reasoning
04 — ALWAYS ON
Ollama / llama3.1
local · offline · private
Requirements
| OS | Windows 10 or 11 (64-bit) |
| RAM | 8 GB minimum · 16 GB recommended |
| DISK | ~1 GB free (application + model download) |
| GPU | Optional — significantly improves inference speed on local model |
| INTERNET | Required for initial download only. Fully offline after that. |
| API KEYS | Optional — system runs completely offline without them |
On the Guardrails
Cursiv has a Guardian firewall. A local-only auth system. A hard policy that your data never leaves your machine without your explicit action. Every one of those decisions was made for the same reason: to keep you safe from things that would compromise you — not to compromise you ourselves.
Guardian
The Guardian layer was not built to monitor what you say or restrict what you can do. It was built to detect when something outside this system is trying to manipulate it — prompt injections, jailbreak attempts, unauthorized access. That lock faces away from you. Toward things that would use this system against you.
Auth
Your login is bcrypt-hashed and stored on your machine. It is never sent to a server. The auth system exists because a tool with access to your files, your conversations, and your local AI should have a lock on the front door. That lock belongs to you — not to us, not to anyone else.
Data
No telemetry. No usage analytics. No conversation logging to any server. The update checker reads GitHub's public releases API — one outbound request to check a version number. Nothing about you, your prompts, your files, or your usage is collected. Nothing. There is no business model built on your data because there is no data leaving your machine.
Fleet
The Fleet Dashboard only works if you set it up — you choose the relay, you choose the token, you choose who gets access. Machines only appear in your fleet if they have been explicitly authorized. This is not surveillance of your users. It is a tool for you to see your own machines.
Transparency
Speed
Ollama runs on your CPU and GPU. Fast hardware = fast answers. On a modern gaming PC it's quick. On older hardware it's slower. That is not a limitation of the software.
Accuracy
The model is not infallible. It will occasionally be wrong, confidently. Treat every answer as a starting point. Never use this to replace your own judgment. Ever.
Coding
Designed to get you most of the way and force you to understand the rest. Feed the errors to a larger cloud model if you need to finish it. That's the point — not dependency, but understanding.
Privacy
Nothing you type is sent anywhere unless you add API keys, which you control. No telemetry. No usage logging. The update checker reads GitHub's public releases API — nothing about you is transmitted.
Version History
Every version here represents a different understanding of what this system could be.
Some of them were wrong. That was the point.
v3.14-U14 RECENT
The Bootstrap Actually Runs, the App Actually Opens
Fixed the auto-install bootstrap silently stalling behind unattended "press a key" prompts, fixed the terminal chat (and the background services) failing to find their own scripts in the installed bundle, and fixed Create Account getting stuck on Login after a reinstall.
v3.14-U08
Zero-Setup First Run — Works Without Keys or Config
Fixed a crash on first install caused by missing data files not being bundled. No API keys, no secrets.bat, no manual setup needed to launch.
v3.14-U04 — Live Demo, Family Welcome & Stability Fixes
Live demo chat on the website. Family member welcome dialog. Terminal chat crash fix. Civilization Decks section. Babel Letters guide.
v3.14-U03
Fleet Dashboard & Substrate Layer
Fleet Dashboard — see every Cursiv install online from one window, live status dots, owner-gated access management. Launcher rebuilt with auto-sizing and Ollama detection. Unified login — one password across all interfaces. Password reset and account recovery from the login screen.
v3.14-U02
Group Discovery, Web Search & Babel
Real-time worldwide web search (DuckDuckGo + Brave). Babel Agent — any language to English via UTF-8 binary. Group Discovery — four providers build on each other's answers toward a structured synthesis. FunForge bounded creative spike timer. Offline routing fixed. Winkler-Codex offline code council.
v3.14.0
Ollama Ready — Offline Edition
The first truly offline release. Ollama + llama3.1 integrated as the primary local runtime. Council deliberation parallelised via ThreadPoolExecutor. System prompt condensed from 12,000 to 4,400 tokens without losing a single functional instruction. Binary-fragment authentication. PyInstaller one-dir bundle.
v2.1.5 LEGACY
Complete Rebuild
Complete ground-up reimagination. The 14-agent council took real shape here — each agent producing genuine LLM responses, 10 advising internally, 4 synthesizing outward. Identity drift abort at 3% deviation. The system started having a real set of guiding rules, not just a config file. Streamlit UI.
v2.x LEGACY
The Council Takes Shape
Early multi-agent experiments. Agents existed but didn't yet deliberate in any real sense — responses were sequential, not parallel. The idea of a synthesizing layer above the advisors began forming here. Lots of things were broken. That was still useful.
v1.x LEGACY
First Contact
The original concept — a personal AI workspace that didn't report to a cloud. Single agent, no council, no constitution. Just the question: what would it look like if this ran on your machine, followed your rules, and grew with your thinking? Everything since has been the answer to that question.
Human Frontier Academy
8 hand-built decks. 200+ cards. One guide per deck. A structured walk through 5,000 years of human civilization, built into Cursiv as reference material you can draw on and build from.
DECK 01
Mesopotamia
~3500 BCE – 500 BCE
The first cities. The first writing. The first laws. The Tigris and Euphrates gave birth to a model of civilization that everything since has been built on — or reacting against.
DECK 02
Egypt
~3100 BCE – 30 BCE
A civilization built on the certainty of death and the refusal to accept it. The Nile as spine. The pyramid as statement. Over 3,000 years of unbroken continuity that most modern nations can't come close to.
DECK 03
Rome
~753 BCE – 476 CE
How a city on seven hills became the model for every republic and empire since. Its fall is more instructive than its rise — a lesson in what happens when institutions hollow out from the inside.
DECK 04
China
~2100 BCE – 1912 CE
The oldest continuous civilization on earth. Four thousand years of dynasties, philosophy, invention, and a scale of social organization the West didn't encounter anything like until much later.
DECK 05
Renaissance
~1300 – 1600 CE
Europe wakes up. The rediscovery of Greek and Roman thought, married to new money and a changed church, produced an explosion of art, science, and selfhood that we still live inside.
DECK 06
Industrial Revolution
~1760 – 1840 CE
The moment human civilization went from muscle power to machine power. Everything accelerated. Not all of it in the right direction. The gains were enormous; the cost was paid by people who had no say in the matter.
DECK 07
Tech Age
~1970 – 2020 CE
The transistor gave us computing. Computing gave us the internet. The internet gave everyone a voice — and then a microphone, an audience, an algorithm. We are still figuring out what we made.
DECK 08
The Future
2020 CE → ?
Not a prediction. A framework. What are the forces already in motion? What does a civilization that has learned from the previous seven actually look like? This is the deck we are building in real time.
Scroll → to see all 8 decks