Syzygy
Syzygy builds Mach Studio, a macOS desktop app that runs open large language models entirely on Apple Silicon. Its agent connects to Slack, Gmail, Notion and Linear, drives the browser, and schedules recurring work locally.
What is Syzygy?
Syzygy, Inc. is a United States company building tools that run large language models on personal hardware instead of renting them from a cloud provider. Its About page describes a group of mathematics and physics researchers whose stated mission is to make intelligence something people own the way they own an operating system, and to kill the billable token. The work is framed around four bottlenecks: capability per byte of memory, capability per unit of computation, computation per watt, and useful work per dollar of total system cost.
Four pieces make up the product line. Mach Studio is the flagship macOS desktop app: it runs open models on-device, connects to Slack, Gmail, Calendar, Notion and Linear, drives the browser by scrolling, typing and clicking, works alongside the user in shared documents and sheets, and schedules recurring tasks. Mach Engine is the local inference engine, whose fused Metal kernels decode the packed low-bit weights inside the matrix multiplication itself, so no dequantised copy of the model ever exists in memory. Mach Agents is a TypeScript agents SDK built for small local models, with permission policies that reach into tool arguments and hard ceilings on tokens, dollars and wall-clock time. Syzygy Gateway routes requests to OpenAI, Anthropic, xAI, Google and OpenRouter through a single OpenAI-compatible endpoint, with automatic fallbacks and unified observability.
The flagship model is Mach-1 Additive 35B, also called Mach-1 Small: a 35B-A3B mixture of experts, 256 experts with top-8 routing, compressed to roughly 1.7 bits per weight and shipped as an 8.1 GB pack that serves on Apple Silicon Macs with 16 GB or more of unified memory. Syzygy reports it retains 96.3% of its own BF16 teacher's score across twelve benchmarks, against 93.6% for Ternary Bonsai 27B and 85.6% for Gemma 4 Q2_K_XL. A multimodal variant adds a vision tower. Mach-1 XS for mobile and Mach-1 Medium for higher-memory laptops are announced without dates, and the product is in beta for Apple Silicon.
What it does
- Run an open 35B model entirely on an Apple Silicon Mac with 16 GB or more of unified memory, with no account, API key or cloud connection
- Read, draft and act across connected apps: Slack, Gmail, Calendar, Notion and Linear
- Drive the browser on the user's behalf, scrolling, typing and clicking through a task
- Point Claude Code, Codex, OpenCode, Hermes Agent or Cursor at the same local model
- Schedule recurring agent tasks that run on the Mac by themselves, such as daily digests, inbox triage or status reports
- Route requests to OpenAI, Anthropic, xAI, Google or OpenRouter through one OpenAI-compatible endpoint with automatic fallbacks
- Build TypeScript agents with per-argument permission policies and hard budgets on tokens, dollars and wall-clock time
When to use Syzygy / When not to
A quick filter to help you decide if Syzygy is the right fit.
When to use Syzygy
- Owners of an Apple Silicon Mac with 16 GB or more of unified memory who want a capable assistant that runs without an account, an API key or a cloud connection
- Developers and engineering teams who want Claude Code, Codex, OpenCode, Hermes Agent or Cursor pointed at a local model instead of a metered cloud endpoint
- People handling confidential material, for whom the default behaviour matters: chats, files and agent memory stay in a local SQLite database on the machine
- Teams trying to cap what they spend per token, which is the explicit thesis of the vendor: own the intelligence rather than rent it
- Builders of agents who need deterministic guardrails, with permission policies that inspect tool arguments and hard ceilings on tokens, dollars and wall-clock time
When not to use Syzygy
- Windows and Linux users: the app requires an Apple Silicon Mac (M1 or later) with Metal, and there is no CPU or CUDA path
- Anyone on a Mac with less than 16 GB of unified memory, or looking for a phone or tablet app, since the mobile-targeted Mach-1 XS model is announced without a date
- Buyers who need contractual compliance guarantees: the site publishes no DPA, no GDPR statement, no Article 27 representative and no certification such as SOC 2 or ISO
- Organisations that require a negotiated enterprise agreement, a service-level commitment or a named account contact, none of which are published
- Users under 13, the minimum age set by the terms of service, and anyone who cannot accept a product that is still explicitly in beta
How to use Syzygy
A typical end-to-end flow, from setup to results.
- Check the requirements first: an Apple Silicon Mac (M1 or later) running macOS with Metal, and at least 16 GB of unified memory for the 35B pack
- Try it without installing anything through the in-browser WebGPU demo, which streams the 7.0 GB Mach-1 Small weights into the GPU with no account and no API key
- Download Mach Studio for Mac from the site and install it; the free plan requires no account
- Pick a model from the curated library and switch between speed and depth in a click, or let Mach Studio manage the engine, choose the serving mode and show throughput, time-to-first-token and memory
- Connect the apps that matter, Slack, Gmail, Calendar, Notion or Linear, authorising each through the provider by OAuth, revocable at any time
- Set every tool to Allow, Ask or Block, configure computer access per app, then review the tool calls in the conversation
- Schedule recurring agent tasks such as a daily digest, inbox triage or a status report
- For a terminal workflow, run pip install mach-engine with Python 3.12 or later, then mach-serve to expose an OpenAI-compatible endpoint on port 8080; mach-check reports whether a checkpoint is servable and fits in memory before any weights are loaded
- To attach a coding harness, run mach launch claude, or start OpenCode, Codex, Hermes Agent or Cursor inside Mach Studio
- For agents or hosted inference, install @syzygy-ai/agents from npm and point it at localhost, or get a key from the account dashboard and target the Syzygy Gateway endpoint
Pros & Cons
Pros
- Privacy by construction: when a local model is selected, inference and chat content stay on the Mac, and chats, files and agent memory are stored locally by default
- Works offline and asks for nothing: no account, no API key and no cloud connection for local models
- A permanent free plan at zero cost with no account, and a predictable Pro plan at USD 20 per month that includes USD 20 of inference credits
- Quality claims published with their methodology, protocol and stated limitations on a dedicated benchmarks page, rather than as a bare marketing number
- The inference engine is released as open source under Apache-2.0 and the model packs are downloadable from Hugging Face
- A named list of subprocessors is published on the Security & Data page, which is unusual at this stage of maturity, and the vendor states it does not train on local conversations
- Fine-grained control for agent builders: permission policies evaluated inside tool arguments, durable approval checkpoints and hard budget ceilings
Cons
- macOS on Apple Silicon only, with Metal required and no CPU or CUDA path, and a floor of 16 GB of unified memory for the 35B pack
- No mobile app at all; the mobile-targeted Mach-1 XS model is announced as coming soon, with no date
- A very young vendor still in beta: the domain was registered on 21/07/2026 and the first Wayback capture dates from 13/08/2026
- No postal address, no named executives and no identifiable beneficial owner anywhere on the site, beyond the legal name Syzygy, Inc. and Delaware governing law
- A single generic address, support@withsyzygy.com, serves support, legal, privacy and vulnerability reports alike
- No GDPR mention, no DPA, no Article 27 representative and no published certification, and no interface or processing languages are declared
- The Gemma comparison figures are taken from a third party's publication rather than measured in-house, and no sample-level reproducibility package is public: artifact hashes, run IDs, sample counts, prompt hashes and raw outputs are not released
Pricing & Plans
A permanent free plan is available at USD 0 and requires no account: everything that runs on the Mac is free, including local models and the user's own keys and integrations. The lowest paid entry point is the Pro plan at USD 20 per month, which includes USD 20 of hosted inference credits, more than 150 hosted integrations, remote control from any browser and automations. A Usage Based plan sits between the two and charges hosted inference per token. Paid plans are billed in advance on a recurring basis through Stripe and renew automatically until cancelled; cancellation takes effect at the end of the current billing period and, except where the law requires otherwise, fees are non-refundable. No time-limited free trial and no enterprise pricing are published.
- local models through Mach
- MLX and llama.cpp
- plus the ability to bring your own keys and integrations. Managed cloud inference
- hosted integrations
- remote control and automations are not included
- the same local capabilities
- with managed cloud inference charged per token. Hosted integrations
- remote control and automations are not included
- local capabilities
- USD 20 of managed cloud inference included each month
- hosted integrations across more than 150 apps
- remote control from any browser and automations
Data, GDPR & hosting
A consolidated view of how Syzygy handles your data.
GDPR overview
There is no mention of the GDPR anywhere on the site: neither the privacy policy, the terms of service nor the Security & Data page names the regulation, cites a legal basis, designates an Article 27 representative or appoints a data protection officer. What the privacy policy does say is generic and conditional: depending on where the user lives, they may have the right to access, correct, export or delete their personal information and to object to or restrict processing, with a single channel, support@withsyzygy.com, and a promise to respond as required by applicable law. Syzygy states it is based in the United States and that data handled by its hosted services may be transferred to the U.S. and other countries under unspecified appropriate safeguards. One European signal is concrete: no Mach Studio product analytics are collected from Europe at all.
Who owns the data?
The terms are unambiguous: the user retains all rights to the prompts, files and outputs they create, which Syzygy, Inc. calls Your Content. The company grants itself only the limited rights needed to operate the services and carry out the tasks the user asks for. Syzygy owns the services themselves and the surrounding intellectual property, with two carve-outs: the open-source components, governed by their own licences, and the user's content. In practice most of that content never leaves the machine, because chats, prompts, generated artifacts and agent memory are stored locally in a SQLite database inside the user's data directory.
Reuse rights
Nothing in the terms restricts what the user may do with the prompts, files and outputs produced through the services: the rights stay with the user, so the material can be reused, published or sold without asking permission. The counterpart is a responsibility clause. Syzygy warns that AI output can be inaccurate, incomplete or offensive, that the same prompt may produce different results, and that the user must review and verify anything before relying on it, including any action taken on that basis. The services are explicitly not to be relied on for professional advice, legal, medical or financial. On the vendor's side, the privacy policy states that the contents of local conversations are not used to train its models; data leaves the device only for hosted inference, connected integrations and other network-backed actions the user invokes.
Data retention & training
Hosting summary
By default nothing is hosted at all. Chats, prompts, generated artifacts and agent memory live in a SQLite database inside the user's data directory on the Mac, protected by the operating system's account and disk-level protections. Data leaves the device only when a feature requires a network: hosted inference, connected integrations, web-backed tools, account services, updates and permitted analytics. Syzygy states it is based in the United States and that information processed by its hosted services may be transferred to and handled in the U.S. and other countries under appropriate safeguards it does not detail. No data centre, cloud region or residency option is named. Traffic is encrypted in transit with TLS, while persistent databases and object storage rely on the protections of their hosting providers. The named providers are Supabase for authentication, account data and release distribution, Vercel for the public site and dashboard, Fly.io for the Gateway and the integration, browser-session and remote-control APIs, and Upstash for queues, caches and rate limits. No Mach Studio product analytics are collected from Europe.
Things to keep in mind
Risks and trade-offs to weigh before adopting Syzygy.
- The local-first promise only covers the local mode: hosted inference, connected integrations, remote control and out-of-Europe analytics all send data off the machine, and it is easy to assume the guarantee is broader than it is
- The agent can drive the browser, use the computer per app and act inside Slack, Gmail, Calendar, Notion and Linear; the permission scope granted deserves a deliberate review before deployment, because an over-permissive setup delegates real-world actions to a model
- Syzygy is a very young vendor still in beta, with no postal address, no named executives and no identifiable beneficial owner, which makes it difficult to know who is accountable if something goes wrong
- No GDPR mention, no data processing agreement, no Article 27 representative and no certification: an organisation with regulatory obligations cannot document this tool from the vendor's public pages alone
- Support, legal questions, data rights and vulnerability reports all go to a single generic address, so there is no dedicated escalation path when one is needed
- The terms disclaim all warranties, make fees non-refundable except where the law requires otherwise, and cap total liability at the greater of twelve months of fees or USD 100
- Running an assistant that is always available, private and free of metering makes it easy to delegate more thinking than intended; the vendor itself warns that output can be inaccurate and must be verified before being relied on, and that the services are not a source of professional advice
Setup & Integrations
Technical difficulty
Low for the desktop route, moderate for the rest. Mach Studio installs from a direct download, needs no account, and manages the engine itself: it installs the runtime, picks the serving mode, spawns the server and shows throughput, time-to-first-token and memory, so no terminal is required. The real barrier is hardware: an Apple Silicon Mac with Metal and at least 16 GB of unified memory. The command-line engine needs Python 3.12 or later, the agents SDK needs npm. Expect a configuration step to connect apps by OAuth and set each tool to Allow, Ask or Block.
Deployment
Integrations
Behind Syzygy
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Syzygy.
Frequently asked questions
What hardware and operating system does Syzygy require?
Do I need an account or an internet connection?
How much does it cost?
Is there a free trial?
Are my conversations used to train the models?
Where is my data stored?
Which tools and harnesses can I connect?
Is there an API?
What is the minimum age, and how mature is the product?
Should you pick Syzygy?
Syzygy is one of the more coherent statements of the local-AI thesis currently on the market. The company does not simply repackage someone else's open weights: it publishes its own compressed model at roughly 1.7 bits per weight, an inference engine released under Apache-2.0, an agents SDK designed for the failure modes of small models, and a gateway that routes to the major hosted providers when local is not the right answer. That vertical coverage is rare, and the technical transparency is genuine: the benchmarks page states its protocol, names its harness and hardware, and admits what it has not published.
The constraints are equally clear. This is a macOS Apple Silicon product with a 16 GB floor, in beta, from a company whose domain was registered in July 2026 and whose site first appeared in the Wayback Machine in August 2026. Corporate transparency runs in the opposite direction to the technical kind: no postal address, no named leadership, no disclosed funding, and one generic email for support, legal and privacy alike. Compliance buyers will find nothing to hold onto — no GDPR statement, no data processing agreement, no certification.
The privacy promise is real but bounded, and worth reading precisely: it holds when a local model is selected. Hosted inference, connected integrations and remote control all send data off the machine, and an agent authorised to drive a browser and act inside Slack, Gmail and Notion deserves a deliberate review of what it is allowed to touch.
For a developer or a technical team with the right Mac, the free plan makes evaluation cost nothing, and the in-browser demo costs even less. For anyone who needs contractual guarantees, a known counterparty or a platform other than macOS, this is not yet the tool.
- Choosing a selection results in a full page refresh.
- Opens in a new window.