Starter
$12/month
50 GB cloud storage
Choose Starter
RELEASE CANDIDATE 0.6.5
37 MB COORDINATOR BASELINE
LOCAL GPU & AGENT STUDIO
Local AI without the bloat.
Release Candidate 0.6.5. Monotonic fencing.
The sovereign, local-first AI desktop harness and agent studio. Orchestrate local models (Ollama, llama.cpp, vLLM), multi-agent durable task boards, and sandboxed tool execution on your own workstation with orders of magnitude less memory than competing harnesses.
Linux x86-64
macOS Apple Silicon
Bubblewrap / tmux Isolation
Local GPU Acceleration
Apache 2.0 Core
ordivyr --mode local --sandbox bwrap
● ILLUSTRATIVE · COMMERCIAL BUILD 0.6.x
$ ordivyr run code-refactor.json [RUNTIME] Initializing ORDIVYR 0.6.5-rc coordinator... [MEM] Coordinator Peak RSS: 37.2 MB (Hermes: 975MB, OpenClaw: 828MB) [SANDBOX] Linux bwrap isolation active: • read_only: ["/workspace", "/usr/lib"] • denied: ["/etc/shadow", "~/.ssh", ".env"] [FENCE] Monotonic SQLite lease acquired (fence_id: 89401) [AGENTS] Supervising 3 local agents: • Worker #1 (Gemma-4-12B QAT · Local GPU): RUNNING (6.57 t/min) • Worker #2 (DeepSeek-Coder · Ollama): CLAIMED • Worker #3 (Deterministic Lint): SUCCEEDED --------------------------------------------------------------- STATE: DURABLE CHECKPOINT RECORDED · Zero duplicate execution
MEASURED REPOSITORY BENCHMARKS [M]
Empirical Scaling & Hardware Benchmarks
Verified evaluation derived from test runs in docs/scalability/RESULTS.md, ORDIVYR_SCALING_AUDIT.md, and 0.6.5 release test logs.
| Agent Harness Platform | Coordinator Peak RAM | CPU-Seconds / Task | Durable Fencing Architecture | Measured Execution Efficiency |
|---|---|---|---|---|
| ORDIVYR (ELYRAON) | 37 MB | 0.15 CPU-s | Monotonic fencing tokens + SQLite WAL | 1.0x (BASELINE) |
| Hermes Agent | 975 MB | 1.42 CPU-s | In-memory process state | 26.3x LARGER |
| OpenClaw DevOps | 828 MB | 1.18 CPU-s | Ephemeral Redis cache | 22.4x LARGER |
Dedicated Local GPU Benchmark
Measured Gemma-4-12B QAT execution under local GPU acceleration: 16/16 validated agent tasks completed with 0 duplicate commits, sustained throughput of 6.57 tasks/min, and compact local accelerator memory footprint.
Verified Scaling & Fencing Boundaries
Synthetic lease testing demonstrated 5,000 to 10,000 concurrent leases. Scaling audit qualifies that single-file SQLite coordinators saturate at ~4–6 multi-step tasks/s due to disk write barriers (~25–35 fsyncs/task), informing our sharded multi-coordinator architecture.
PLATFORM CAPABILITIES FOUR SOVEREIGN ENGINES
Engineered for full local workstation autonomy.
Zero telemetry. Zero lock-in.
Model Hub & Local Hardware Dispatch
Seamlessly route prompt tasks to local GGUF models, Ollama instances, vLLM endpoints, or dedicated local GPU accelerators. Automatically manages quantized weight caching and thread allocation.
Local Hardware Execution
Run models up to 14B parameters natively on local GPU hardware. Zero token invoices, zero latency jitter from public API throttling.
Multi-Backend Routing
Direct drivers for Ollama, vLLM, llama.cpp, and local OpenAI-compatible endpoints with automated fallback upon local memory constraints.
Visual Agent Studio & Multi-Worker Coordination
Availability: Agent Studio is part of the commercial ORDIVYR desktop build. The public open-source release (v0.4.0a0) is CLI-only and does not include it.
Compose complex multi-agent coding workflows. Assign specialized roles (researcher, implementer, reviewer, verifier) with strict hierarchical supervisors and human-in-the-loop consequential approval gates.
Consequential Action Gates
Destructive terminal commands, production deployments, and secret transmissions are halted for human approval before execution.
Real-Time Supervision
Inspect live tool invocations, token throughput, and task execution logs in a unified graphical dashboard.
Durable Task Graphs & Monotonic Fencing
Task state is preserved in an append-only SQLite transaction ledger. If a worker process crashes or reboots mid-task, monotonic fence tokens prevent duplicate work and resume exactly where execution stopped.
Zero Duplicate Execution
Every worker lease carries a monotonic token. Expired leases fail-closed, guaranteeing no split-brain git commits or repeated database writes.
Crash-Resilient State
Task checkpoints, intermediate tool outputs, and execution logs persist in WAL mode across sudden host power losses.
Isolated Local Bubblewrap (bwrap) Sandbox
Agent actions execute in an unprivileged Linux bubblewrap container. Denies access to sensitive host directories (~/.ssh, /etc/shadow, .env files) and restricts network egress to allowlisted endpoints.
Filesystem Boundary
Strict path traversal confinement limits agent write permissions exclusively to the designated active project workspace root.
Credential Defense
Prohibits outbound network transmissions containing detected API keys, passwords, or private environment variables.
INTERFACE PREVIEW DESKTOP WORKSTATION
Real screenshots from the visual QA suite.
Built for operators who value clarity.
ORDIVYR Main Desktop Workstation — Local model selection, agent team overview, and durable task status.
Agent Execution Monitor — Real-time telemetry, tool invocation logs, and lease fence tokens.
Multi-Agent Supervision — Concurrent local agent execution with human-in-the-loop review.
FREE LOCAL HARNESS
OPTIONAL CLOUD STORAGE
Local Workstation Autonomy.
Optional Cloud Room Storage.
The public core ORDIVYR orchestrator is open-source under the Apache-2.0 license (v0.4.0a0 Alpha). Anyone can inspect the source code, run on their local workstation, and coordinate local models without license fees.
Distinguishing Verified Local vs Unverified Cloud Execution: Local workstation execution (via local SQLite WAL, monotonic fencing, and Bubblewrap/tmux isolation) is verified end-to-end (29/29 test suites, 354 passed). In contrast, public cloud-hosted coordinators and distributed multi-tenant cloud execution are experimental. The Cloud Room entitlement bridge to real Cloudflare R2 storage has been verified across 15/15 tests, and paid storage plans launch with commercial ORDIVYR 0.6.6.
Licensing & Release State Distinction:
• Public Open-Source Core: Version 0.4.0a0 (Alpha) released under Apache-2.0 on GitHub (colt2822/ordivyr). Standard library only, Python >= 3.10.
• Commercial ORDIVYR Development: Version 0.6.6 includes Cloud Room encrypted storage subscriptions, packaged desktop application (.deb wrapper), and hosted storage synchronization. Commercial desktop installers are under final owner review and not established as publicly released source. Paid storage plans and the founding offer are described in the pricing section.
OPTIONAL PAID STORAGE
Local first. Storage when you want it.
Plans start at $12/month for 50 GB. No hosted agent compute, no mandatory account to run agents locally. The first 250 paying customers also qualify for the founding offer.
See pricing Dedicated Cloud Room Showcase →PRICING ORDIVYR 0.6.6
Run agents locally.
Pay only for the storage you choose.
ORDIVYR executes agents on your own machine with the models and API keys you control. There is no hosted agent compute. Paid plans add optional encrypted cloud storage for your workspaces, and nothing in the product requires it.
Pro
$20/month
100 GB cloud storage
Choose ProStorage Plus
$39/month
250 GB cloud storage
Choose Storage PlusEvery paid plan is a monthly subscription for optional storage (decimal GB). The Apache-2.0 local engine is free. All three plans qualify for the founding offer below. Already have a storage plan? Open Customer Dashboard →
LIMITED · FIRST 250 PAYING CUSTOMERS
Your First 16 Million AI Tokens Are on Us.
Be Among ORDIVYR’s First 250 Customers.
Launch your AI workspace with ORDIVYR. The first 250 paying customers qualify for promotional DREX credit worth up to 16 million AI tokens on eligible models.
Claim Your Founding Customer OfferOffer details & conditions
- What it is: a $1.00 promotional credit on a DREX account, one per eligible customer. The “16 million” figure is the input-token capacity of that $1: at the published
drex-v2.0rate of $0.06 per million input tokens, $1 covers about 16.7 million input tokens. DREX does not bill output tokens. Rates as published by DREX on 2026-11-06 and subject to change. - Eligible usage: DREX decision inference (
POST /v1/systemone) only. DREX is a decision model that returns calibrated probabilities, not a general chat model. Drex Search, browser runs and document (Perception) jobs are billed at different rates, so $1 buys fewer tokens there. - Who qualifies: the first 250 customers whose first paid ORDIVYR storage subscription payment is confirmed. One allocation per customer; upgrading or changing plans does not create a second one. Free local use does not qualify.
- When it is issued: only after your payment is verified. Your allocation shows as pending until it has actually been delivered to your DREX account, and we will email redemption instructions at that point. Refunded or failed payments forfeit the allocation.
- Expiry: promotional credit is subject to DREX’s own terms and may expire. Credit has no cash value and is non-transferable.
- Third-party service: DREX is operated by Nace.AI, a separate company. ELYRAON does not own, operate or speak for DREX, and this offer is not sponsored or endorsed by Nace.AI. Model usage has its own terms and costs. ORDIVYR runs agents locally with the models and API keys you choose.
- Availability: limited to 250 allocations in total and may end earlier if fulfillment cannot be completed.
After you purchase
- Your payment is verified. No credit is promised before this step.
- If a founding slot is available, your allocation is recorded as pending.
- We email you step-by-step redemption instructions once delivery to your DREX account succeeds, and the allocation changes to delivered. If delivery fails, you will hear from us rather than be left waiting.
GET STARTED PUBLIC OPEN-SOURCE CORE (v0.4.0a0 ALPHA)
Build ORDIVYR Core from Apache-2.0 Source
Public open-source repository: github.com/colt2822/ordivyr. Requires Python >= 3.10 on Linux. Standard library only with zero external runtime dependencies. Tested with Ollama, LM Studio, and local hardware dispatch.
Terminal · Clone & Run ORDIVYR Core
● APACHE 2.0 REPOSITORY
# Public release v0.4.0a0 is CLI-only. The desktop app and Agent Studio are part of the commercial build and are not in this repository. # 1. Clone the canonical public open-source repository git clone https://github.com/colt2822/ordivyr.git && cd ordivyr # 2. Create isolated virtual environment python3 -m venv .venv && source .venv/bin/activate # 3. Install core package (standard library only, no external wheels required) pip install . # 4. Run environment diagnostics and inspect local hardware ordivyr doctor ordivyr hw # 5. Check hardware-aware team templates and launch an agent board ordivyr templates ordivyr new local-dev-duo -o board.json ordivyr run board.json --mission "audit dependencies"