COMPILING NOWisokron.ai/builds
Schema-first. Hash-chained. Hardware-reachable. Yours.
Your keys, your fleet, your code, your hardware. One typed graph, audited by construction, coordinated at the schema layer so your drones don't collide. Compose it for the use case in front of you.
verifiable compiles today: pending·chain events: pending·100% customer-reviewed before commit
chain_height: pendingverified pending
Why agents fail
Most agent platforms treat memory as 'current state matters.' Old facts get invalidated when new facts arrive. You can't replay why the agent made the decision it made yesterday because yesterday's state has been overwritten. We hash-chain every fact and let you replay the graph state that made each decision rational at the time.
Trace logs are not proof. Every agent platform writes logs; none publish customer-verifiable Merkle roots. We anchor the chain to immutable storage hourly. Your auditor verifies us with their own browser.
Drop-in memory layers want your agent to write to their SaaS. Our memory is a structured graph in your workspace — versioned, queryable, exportable, BYOK + BYO-fleet from day one. If you leave, you leave with everything.
chain_height: pendingverified pending
The foundation
Components, Decisions, Tickets, AcceptanceTests, Patterns, Gotchas. Database-enforced foreign keys. Not markdown-with-frontmatter.
Two drones cannot claim the same ticket because the database rejects it. Capability-separated tool surfaces. No clobbering by construction.
Per-tenant chain heads. RFC 8785 JCS canonicalization. Hourly Merkle anchors to immutable storage. Customer-verifiable with zero platform trust.
Your Anthropic key. Your Hetzner VPS. Your friend's gaming PC over SSH. All into one coordinated workspace.
Three components, one substrate. See the full Triad on /architecture#triad.
chain_height: pendingverified pending
Hardware-native I/OAlpha path
PiKVM v4 Mini takes HDMI from anything, sends USB keyboard/mouse to anything, and drives GPIO/relays to anything. UiPath, Blue Prism, Anthropic Computer Use, OpenAI Operator all need to install software inside the locked environment. Enterprise IT refuses. HNAO operates from outside — the exact operations Citrix is built to allow, because they're how a human uses the session.
Citrix-published apps — software-only RPA can't touch these. HNAO reads via HDMI, types via USB HID.
RDP into anything the user has access to. HNAO is the user, from outside the locked desktop.
Hyperspace inside Citrix. HNAO operates the chart the way a clinician does.
Green-screen 3270. PiKVM observes the same pixels the operator sees.
Pre-boot firmware. HNAO drives the keyboard before the OS has loaded.
Read the panel, log the reading, file a ticket when the inverter drifts.
Touchscreen PLC interfaces with no API. HNAO clicks the buttons.
OBD-II diagnostic surfaces. HNAO operates the scanner.
Government web portals with no public API. HNAO browses them anyway.
chain_height: pendingverified pending
The Hive
Cursor’s Composer ships parallel agents via git worktree. Devin runs managed Devins in isolated VMs. None of them coordinate at the schema layer. None of them anchor a chain across the fleet. We do both, by construction. Parallelism without provenance is just faster chaos.
Workspace
One typed graph
Schema-fenced. Hash-chained. Replayable.
See the substrate architecture →
chain_height: pendingverified pending
Intent, every way you have itNext
A designer thinks in visual references. An entrepreneur thinks in competitor analogies. A domain expert thinks in operational procedures. None of them naturally produce a clean text spec. All of them can produce a screenshot, a video, a voice memo. The compiler takes what you have.
Compiler transcribes + extracts UI states from the timeline.
Compiler keys the screen-and-voice walkthrough into a manual entity.
Compiler types the frame as a Component + AcceptanceTests.
Vision pass turns the napkin into typed Components + Tickets.
Transcribe + classify intent fragments against the graph.
HNAO browses it; compiler types observed behavior as gotchas.
Every action signs the next
Per-tenant chain heads, RFC 8785 JCS canonicalization in the application layer, hourly Merkle anchors to immutable storage. None of Cognee, Mem0, Zep, or LangGraph have customer-verifiable cryptographic audit chains. This is a universal category gap. Append-only is provenance. Invalidation is debt.
Verify our chain in your browser →
chain_height: pendingverified pending
Category draw
Cognee · Mem0 · Zep
Stores that an agent queries at runtime. We don’t do that. We compile structure before runtime. Zep invalidates old facts as a feature; we hash-chain every fact and let you replay the graph state that made each decision rational at the time. Different layer, different physics.
LangGraph
LangGraph is the framework you build an agent in — a low-level orchestration library you wire yourself. We’re upstream of where LangGraph sits. The intent compiler produces a structured execution graph that LangGraph (or Cursor, or Claude Code, or your homegrown agent) executes. We coexist with LangGraph at the same project — they run the agent, we decide what the agent should build.
Cursor · Claude Code · Codex · Devin · Windsurf
These run inside an editor. We compile from natural language to a structured graph that runs against your code via your fleet. Cursor caps parallel mode at 8 via git worktrees; Devin runs managed Devins in isolated VMs. We ship 100-drones-on-a-single-project as the substrate. And code is one of nine entity categories in our graph — workflows, dashboards, manuals, videos, vendor portals, competitor intel, equipment states, regulatory references are the other eight.
May 2026
21,639 publicly-exposed instances. CVE-2026-25253 RCE. 13.4% of ClawHub skills with critical security issues. Moltbook backend leaked 1.5M API tokens. This is the audience the customer-review checkpoint and the signed-package Hub were built for.
Tim Cook on the May 1 earnings call attributed the Mac Mini $599→$799 (33%) hike directly to 'AI and agentic tools' demand outpacing forecasts. Several months to reach supply-demand balance. Demand velocity in the category. The customer wants this badly.
Every memory company looks like every other memory company. Purple Cognee, lavender Mem0, indigo Zep, navy LangGraph. We’re not a memory company.