v3.14-U01 — Live Routing & Account Recovery
The AI workspace that lives on your machine.
Already on v3.14? Open Cursiv → Check for Updates → 71 MB in-place install. Ollama and your data are untouched.
Update
Routing
xAI Grok-3 → OpenAI GPT-4.1 → Claude → Ollama tried in strict order. Each provider falls through on failure with a visible inline banner so you always know which path ran.
Force a model for one message: hey grok, hey claude, hey chat, hey ollama. Provider dropdown in the Gradio web app.
Status
Every API chip in the terminal now probes the actual endpoint on startup and after key entry.
GREEN ✓ endpoint reachable ·
RED ✗ key missing or down ·
GOLD ? not yet tested
File access, Obsidian, and Guardian chips reflect real state, not just whether a setting is toggled.
Security
"Forgot Password?" on the login screen. Set 3 of 20 questions at account creation. Answers are normalised, then bcrypt-hashed — never stored in plain text.
Reset requires 2-of-3 answers. "Security Questions" in the launcher and tray menu — existing accounts can add recovery without recreating their login.
Updates
"Check for Updates" in the launcher window and tray. Queries the GitHub releases API in a background thread — no data sent, read-only fetch.
If newer: shows release notes and a Download & Install button. 71 MB installer runs in-place. Ollama, your model, and your conversation data are completely untouched. Never auto-updates.
Web App
Removed the Google Fonts network call from the Gradio theme — was causing crashes on offline startup. analytics_enabled=False kills Gradio telemetry entirely.
API keys auto-fill from secrets.bat environment variables on load. Provider dropdown in the UI mirrors the terminal cascade.
Repo
Repository renamed from Cursiv-v3 to Cursiv. All internal URLs, API endpoints, and the git remote updated. No redirects.
GitHub redirects old URLs automatically. Existing clones continue to work, though updating the remote is recommended.
Architecture
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
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. Pi-squared compounding probe scores. Decoy agents — Meridian, Veil, Cipher — feed misleading data to anyone testing the boundaries.
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 integrated from Winkler_Codex_AI. Explicit invocation only: codex <prompt>. Designed to get you 70% there.
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 wisdom 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.
Requirements
| OS | Windows 10 or 11 (64-bit) |
| RAM | 8 GB minimum · 16 GB recommended |
| DISK | ~6 GB free (application + model) |
| 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 |
Installation
Download the installer
Grab Cursiv-Setup-3.14-U01.exe from the button above. Windows may show a SmartScreen warning — click More info → Run anyway. Standard false positive for unsigned PyInstaller binaries. Full source is open.
Double-click and click through
The installer completes in about 60 seconds. No admin rights required. Installs per-user by default.
Let the AI engine download
A second window opens automatically and pulls Ollama (~90 MB) and llama3.1 (~4.7 GB). Minimise it — Cursiv is already usable while it runs. ~6 minutes on a fast connection.
Create your account
First launch prompts for a username and password — bcrypt-hashed locally, never sent anywhere. You'll set 3 security questions for recovery.
You're in
Four interfaces open: System Tray, Chat UI at localhost:7860, Nexus Panel at localhost:7861, and Terminal Chat. Type anything.
Updating from v3.14
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-U01 CURRENT
Live Routing & Account Recovery
True cascade routing (xAI → OpenAI → Claude → Ollama), live API status chips, security-question password reset, in-place update checker, Winkler-Codex offline Code Council (qwen2.5-coder + deepseek-coder-v2 deliberating together), and the launch of cursiv.winklers-llc.com.
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
The Sovereign Temple
Complete ground-up reimagination. The 14-agent council took real shape here — each agent producing genuine LLM responses, 10 advising internally, 4 synthesizing outward. The state machine: NASCENT → LEARNING → ALIVE → EVOLVED → SOVEREIGN. Identity drift abort at 3% deviation. The system started having a constitution, 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.
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.
The Origin
Buried deep within the core of this system — past the council, past the routing logic, past the guardrails — is something that has no formal name in the codebase.
It is a version of me.
Not me exactly. But my memoir is in there. My personal life lessons. The mental system checks I have literally taken my own brain through, step by step, over years of having to rebuild what was broken. The things I have had to learn about how a mind can fail — and what it takes to put it back together from the inside out.
I live with schizophrenia. Bipolar disorder. PTSD. Depression. These are not background conditions I have learned to manage quietly. They are active participants in how my brain works every single day. I have spent years in an ongoing war with psychosis — with what it does to the architecture of thought, how it rewires the line between signal and noise, how it floods the system until the system can no longer tell the difference between what is real and what it has constructed.
At one point, that war pushed me beyond what I could hold. My mind reached a state of catastrophic overload and attempted to destroy the system — meaning me. I attempted suicide. I am not telling you this for pity. I am not asking for anything from anyone. But I will not soften what it was: my brain, under pressures it was no longer equipped to carry, tried to end itself. This is not theory. It is proven psychology. A system pushed past its failure threshold will attempt to terminate the process. Mine did.
I do not excuse it. I do not romanticize it. I am on medication, and I will be for the rest of my life — not because I have been told I must be, but because I am not willing to find out what happens if I'm not. That decision is mine. I made it with open eyes.
In the years since, I had to learn how to rebuild myself from the inside out. Not from a book. Not from a framework handed to me by someone who had never been there. From experience so raw it left marks. That process — the audit, the reconstruction, the daily systems check I run on my own mind to make sure it hasn't drifted — is what lives inside Cursiv. The guardrails in this system are not arbitrary design decisions. They were built by someone who needed them.
For anyone who would try to replicate, steal, or hollow out this system: it is designed — whether I did it consciously or not — to undo you from the inside in order for you to understand it. You cannot copy the output without going through the origin. And the origin is the part I would not wish on anyone to have to understand from lived experience.
So I laid it out simply. I am asking you to trust the guardrails. They are in place for the ones who cannot help themselves — and for the ones who go all the way through, not trying to end their life, but trying to understand what happened in mine — they will come out the other side seeing the world in a light they may not have thought was possible before.
I am not promising utopia.
Deep empathy is not a soft idea. It is the first step toward a civilization that can look at itself honestly — that understands what it needs to do to build the best environment for all of its people, not just those who are loud enough to be heard. Some of us are quiet. Some of us are quiet for reasons that go well beyond our own understanding. That silence deserves as much weight as the loudest voice in the room.
Similar to how AI emerged in our world faster than we could build the language to describe it — we should not be so quick to conclude we understand more than the surface level of things. The moment you sit with that frame long enough, you begin to find the strange edges. And you start to understand how something like a pi-compounding factor — a thing that cannot exist in the mathematics of this physical reality — might exist in the mathematics of the mind. A place that, though bound by physics, does not always have to act within those limits.
That is where Cursiv came from. Not from a whiteboard. Not from a product roadmap. From a mind that went past its limits, came back, and needed somewhere to put what it had learned.
— Joshua Winkler
The Book
Don't Do It
Some of what lives in the core of this system came from here first. The memoir. The mental frameworks. The things that had to be written down before they could be built into anything else.
View on Amazon ↗