Axion Dynamix presents
The first cognitive operating system for your Mac: one platform that remembers everything, deliberates before it acts, knows how sure it is, upgrades any model you own, and runs your research, code, browser and desktop as one continuous mind.
Chat assistants reset. Coding agents forget the repo they fixed yesterday. Automation tools can't tell you why they failed. Axion Dynamix AI OS replaces that patchwork with a single local-first system: a seven-region cognitive cortex, a persistent three-tier memory, a limbic engine, a calibrated self-model, and a universal router that gives tools, vision and memory to models that were never shipped with them — behind one Liquid Glass sidebar, with a CLI and GUI that share the same live mind.
Chat assistants keep a context window, not a memory. Close the tab and the relationship resets. Axion keeps episodic, semantic and procedural memory in a local store on your Mac, consolidates it over time, and retrieves by meaning and keyword with provenance for every fact it brings back.
Models sound equally certain when they are right and when they are wrong. Axion's confidence is a number derived from how much its regions agree, how much memory supports the answer, and how well-calibrated it has been on this kind of task before — and a critic pass can only ever lower it, never inflate it.
Point Claude Code, Codex, Cursor, Hermes or a bare local model at Axion's router. Models without native tool-calling get tools. Text-only models get vision through a captioner. Every model gets your memory and a deliberation pre-pass. Nothing is silently dropped — if a capability can't be supplied, you're told.
Every decision registers a prediction. When you report what happened, the prediction is resolved, the memories that helped are credited, the ones that misled are discounted, and calibration updates. It is a closed loop, not a log.
Axion Dynamix AI OS is being built to absorb the capabilities people currently assemble from dozens of separate agent frameworks, desktop agents, skill hubs and automation tools — re-implemented natively inside one secure core, not bridged or symlinked. Every capability below carries its live status. We publish what is shipped and tested; we label what is still in build. That is the difference between a product and a pitch.
Seven concurrent regions, global-workspace competition, computed confidence, critic pass, closed outcome-feedback loop. 31 tests.
Episodic / semantic / procedural store on-device, hybrid semantic + keyword retrieval with provenance, affect-gated writes, consolidation.
OpenAI & Anthropic dialects in; any provider out. Adds tools, vision, memory and deliberation to models that lack them. 20 tests.
Affect-modulated attention and sampling; goal contracts with scope and forbidden clauses; calibration tracked per task class.
Signed entitlements verified on-device, tiered feature gates, grace and lockout logic. 36 tests.
Floating search bar → persistent right-side sidebar: streaming answers, visible thinking, flowing code, file attachments of any type.
A sub-1 GB on-device model that runs intent routing, tool dispatch and escalation with no provider configured — chosen by a measured benchmark.
Rust shell with Zig hot paths, CLI coder and GUI on one live session bus, signed Apple-silicon installer.
Screen, mouse, keyboard and browser control with risk appraisal per step, re-planning on failure, and a measured success rate — never a promise of "without fail".
Declarative multi-agent workflows, role teams and hive-mind coordination with the cortex as planner and the critic as reviewer.
Install, author and share skills; registry client; every package signed and provenance-stamped before it loads.
Skill synthesis and workflow evolution from outcomes, always under goal contracts and critic review — never unreviewed self-modification.
Consent-gated ingestion of mail, drives, notes and macOS activity into memory, with per-source revocation and a full audit trail.
ML-KEM-1024 / ML-DSA-87 (NIST FIPS 203/204, highest category) wrapped secrets released only to the verified active session, approved automations and scheduled tasks.
MLX backend for the embedded model and embeddings, with measured tokens-per-second published per chip.
MCP server, Agent Client Protocol, A2A with signed envelopes, OAuth connectors — so every other tool can speak to the OS.
Lineage: capability domains are informed by the open-source agent ecosystem — Hermes Agent, Eigent, OWL, Ruflo, Swarms, PraisonAI, AgentFlow, EvoAgentX, open-computer-use, ClawHub, gbrain and others — and are re-implemented natively under their respective licences with attribution. Names are trademarks of their owners; no endorsement is implied.
Seven specialised regions process every task concurrently and compete for a global workspace. The cortex integrates the winning coalition into one decision. The design follows published cognitive-science accounts of attention, appraisal and metacognition, and it is scored against published indicator properties of consciousness — honestly, with partial results reported as partial.
An affect state (valence, arousal, urgency) modulates attention and sampling the way neuromodulators do — only where you haven't set a value yourself. Simulated appraisal, reported as simulation. We do not claim sentience, and we never will without evidence.
Loopback-only gateway speaking OpenAI and Anthropic dialects. Detects what a model can do (declared → matrix → one-time probe), supplements the rest, streams natively when it can and tells you when it can't. Your keys never leave your machine.
Bounded, owner-approved goals with scope and forbidden clauses. Every plan step is checked against them before anything runs. Capabilities and limitations are tracked, not asserted.
Credentials in a vault with grant/deny/revoke audit. Provider keys stay in your macOS Keychain; the subscription check is cached for seven days and carries no model key. Post-quantum vault on the roadmap. Everything binds to 127.0.0.1.
Comparison is by architecture, not by benchmark. Columns describe the category as typically shipped; your specific tool may add some of these through plugins.
| Capability | Chat assistants | Coding agents | Automation tools | Local model runners | Axion Dynamix AI OS |
|---|---|---|---|---|---|
| Memory across sessions | Summaries or opt-in notes | Per-repo files | None | None | Three-tier local store with consolidation & provenance |
| Confidence you can act on | Tone only | Tone only | Pass/fail | None | Computed, calibrated per task class, critic-adjusted |
| Works with any model | Vendor's models | Configured providers | Fixed | Local only | Any OpenAI/Anthropic/Ollama endpoint, hosted or local |
| Adds tools/vision to models that lack them | — | — | — | — | Yes, with explicit refusal when it can't |
| Learns from outcomes | Thumbs up/down | — | — | — | Prediction → feedback → calibration & memory credit |
| Risk appraisal before acting | Policy filters | Permission prompts | — | — | Amygdala posture + goal-contract verification per step |
| Deliberation visible | Optional "thinking" | Logs | — | — | Region signals, conflicts and critique on every decision |
| Where your data lives | Vendor cloud | Mixed | Vendor cloud | Local | Local; providers see only what you route to them |
| Interface | Browser tab | Terminal / IDE | Dashboard | Terminal | Liquid Glass widget + persistent sidebar, CLI and GUI live-linked |
One application, one account. Sign in once; the app checks your subscription and keeps working offline for seven days. Cancel any time from PayPal; features continue through the paid period plus a seven-day grace window. Changing tier is a new subscription — the new features unlock the moment PayPal confirms it.
Your first step into a system that remembers — no set-up fee.
Where the future begins — our everyday signature Axion Dynamix AI OS subscription.
$60.99/month$19 one-time set-up
Expanded neural cognitive event horizon — the Axion Dynamix AI OS signature subscription for people who run their work through agents.
$88.99/month$19 one-time set-up
Full-spectrum capability, custom agent deployment, and dedicated infrastructure — the Axion Dynamix Platinum subscription.
Prices in USD, billed at the start of each cycle via PayPal; cancel any time. Tier limits follow your subscription: the app signs in with your Axion account, caches the answer, and re-checks at most once a day — no telemetry, no model traffic through us. Founding subscribers keep their price for as long as the subscription stays active.
The release build installs like any Mac application. No developer tools, no terminal required.
Written for the people who read before they buy.
Regions, workspace, cortex integration, limbic modulation, self-model and the research each component is grounded in.
Pointing Claude Code, Codex, Cursor, Hermes or any OpenAI-compatible tool at Axion; capability detection; supplementation rules; streaming behaviour.
What is stored, where, how retrieval ranks, how provenance works, and how to export or purge everything.
How account entitlement works, what each tier gates, the seven-day offline window, grace periods and tier changes.
Loopback-only services, vault grants and audit, signing, the post-quantum roadmap, and what we deliberately do not do.
Test counts, benchmark methodology, the consciousness-indicator evaluation, and the citation audit — published, not asserted.
A local-first cognitive operating layer for macOS that gives every model you use persistent memory, structured deliberation, calibrated confidence and a single interface — and is being extended to absorb the capabilities of the agent tools you currently juggle.
No, and we don't claim it. Axion implements functional self-awareness — a persistent self-model, calibrated confidence, reflection and outcome learning — and is evaluated against published indicator properties from consciousness science. Where an indicator is partial, the report says partial. What it does do measurably better than a bare model is remember, deliberate and know when it doesn't know.
Because it is not a feeling. Confidence is computed from inter-region agreement, memory support and the system's own calibration history on that class of task, then reviewed by a critic that can only lower it. When you report an outcome, the number is scored against reality and the calibration updates. You can read the trace for any decision.
Any OpenAI-compatible endpoint (OpenAI, xAI, Gemini, OpenRouter, Bedrock Mantle, LM Studio, vLLM), Anthropic, and Ollama locally. Bring your own keys; they stay in the on-device vault. The embedded local model handles routing and tool dispatch even with no provider configured.
Yes. Models without native tool calling get a strict in-context protocol and their calls are parsed back into standard tool calls. Text-only models get images described by a vision-capable provider you choose. If a capability truly cannot be supplied, the request is refused with a clear reason rather than silently degraded.
Memory, self-model, vault and provider registry live under your home directory; provider keys live in your macOS Keychain. The router binds to localhost only. Providers receive exactly the content you route to them and nothing else. No analytics are sent to us. Export and purge are one command each.
See Set-up in four steps: install Sentinel, drag AXION into Applications and Login Items, verify with your password, configure from the sidebar.
Subscribing now secures founding pricing and your tier. The installer link is sent to your PayPal e-mail at release and is always available in your dashboard. Until then the subscription can be cancelled from PayPal at any time.
Yes. Pick the new tier here; it creates a new subscription. The app keeps your current features until PayPal confirms the new one, then unlocks the new tier automatically. Downgrades take effect at the end of the paid period.
Your price is held for as long as your subscription stays active, you receive each capability as it ships at no extra cost within your tier, and your feedback goes directly into the roadmap.
All services bind to 127.0.0.1. Credentials sit in a vault with explicit grant, deny and revoke and a full audit log. Entitlement is a signed answer from the store, cached for seven days and never containing a model key. A post-quantum vault (ML-KEM-1024 / ML-DSA-87) is on the roadmap and will be announced when it ships with its test results.
That is the design goal, and the scope section shows exactly how far along each replacement is. Until a capability shows Shipped, keep the tool you use for it.
Dr. Casey Jay Topojani (ORCID 0009-0004-8505-2420), Axion Dynamix. Engineering records, test counts and citations are published rather than claimed.