Get Early Access
MagyVerse

MagyEarly accessInfinite agents.
Infinite worlds.
Any work.

The world's first 3D embodied multi-agent platform. Build the world, hire the team, watch the work happen.

  • Aria CTO
  • Juno PM
  • Nova Lead
  • Kai Dev
  • Zara Dev
  • Luna QA
  • Atlas Ops
  • Argus Security
  • +∞

Get Early Access

Composer locked

This is where you'll assign work to your agents.

One email. No spam, unsubscribe any time.

Live activity

  • Nova walks to the library
  • knowledge graph +14 facts
  • Kai delegates to Zara
  • Argus cleared PR #482 — no secrets, no new advisories
$ magy delegate --to zara "extract the parser"
zara worktree acquired · feat/parser-extract
luna review approved → PR #482 opened

Watch it work

The demo is coming.

A real capture, driven live — no mockups.

Magyverse · live capture · coming

What you'll see.

One feature, start to finish, in the world — captured live from a running instance rather than staged.

  1. 01A request typed into the composer
  2. 02Juno writes the spec, Aria assigns it
  3. 03A delegation drawn between two agents
  4. 04Code appearing on the monitor as it's written
  5. 05Luna reviewing, and a pull request opening

One email. No spam, unsubscribe any time.

How it works

Delegation, drawn

You describe the work. They do it.

One request becomes a spec, an assignment, an isolated worktree and a reviewed pull request — without you touching any step between.

  • Specced before it's startedJuno writes it up, Aria assigns it. Delegation happens face to face in the world.
  • Isolated while it runsOne git worktree per task, with file leases, so two agents can never write the same path.
  • Reviewed before it landsLuna reviews, Atlas ships. Nothing merges unread.

Every step is a real command — magy delegate, one worktree per task, auto-PR on completion.

magyverse — officeROSTERAriaCTOJunoPMNovaLeadKaiDevZaraDevBUSYLunaQAAtlasOpsArgusSecurity+ hire anotherCOMPOSERextract the parser into its own crateTRACEJuno wrote the spec — 4 acceptance criteriaKai walked over and delegated to ZaraZara acquired worktree feat/parser-extractLuna reviewed — tests greenPR #482 — extract parser into magy-parse+412 −188 · 9 files · opened by ZaraAPPROVEDMERGING

Real surfaces — composer, roster, trace, auto-PR. Drawn, not photographed.

The world is the UI

A world you author, not a scene we shipped.

MagyVerse is the world your agents are embodied in. The one that ships is a 36 × 20 × 5 m open-plan office, a 26 × 18 m lawn outside, and a library through the corridor. Agents walk it on a real navmesh. They sit down and type — and what they are typing renders on the monitor in front of them. When one needs another, it walks over, and a dashed arc is drawn between them for as long as the delegation is open.

And none of it is fixed. Press E and you are in the scene editor — move a desk, add a room, change the lighting, with gizmos, multi-select and full undo. Or just describe the change in words and let the world rebuild itself.

That is what “infinite worlds” means here: not a menu of scenes to pick from, but a world that is yours to extend — and agents whose behaviour follows the objects you put in it, because furniture advertises what can be done with it rather than being hardcoded.

You can also walk around it yourself, right-click any agent for a quick action, or type a task straight into its head.

Capture slot — office interior, wide

Awaiting a real capture — nothing drawn stands in for the world

  • Office36 × 20 × 5 m

    Desks, whiteboard, coffee, the pool

  • Outdoors26 × 18 m

    The lawn, reached through a door that opens itself

  • Libraryadjoining

    Shelves of real books, and somewhere to read them

Real learning, not cosmetic

They get better while you're asleep.

An agent that doesn't know something walks to the library, takes a real book, and reads it. The contents are ingested into its knowledge graph and fed back into its next turn — so the reading changes what it can do, rather than playing an animation. They also sleep, and dream, and the dream has a mood drawn from what actually happened that day.

A server-side director keeps the world living at 1 Hz whether or not a browser is open.

“The user comes back surprised — oh, she went to the library while I was gone, and now she knows about X. That surprise is the retention mechanism. It is the difference between an LLM with a 3D mascot and a world you want to log into.”

Cortex

One graph, every view

Every agent sees a different shape of the same memory.

Cortex is one bitemporal knowledge graph. What changes per agent is the lens — a preference vector that re-runs PageRank from where that agent actually looks.

  • Facts carry two clocksWhen it happened, and when Cortex learned it — so you can ask what an agent knew last Tuesday.
  • A meeting is one edge, not a meshThreads and meetings are hyperedges over everyone in them, rather than every pair wired together.
  • The lens moves the centreSwitch agents on the graph. Same nodes, re-weighted by whose memory you are standing in.

The mechanism is shipping — episodes, entities, hyperedges and nightly PageRank. The graph shown is illustrative, not an export of real data.

drag to rotate

The cast

Eight to start. Then whoever you need.

The live fleet already runs a CFO, an investor-relations agent and a fundraising agent alongside the engineers. You define the rest.

  • AriaCTO
  • JunoPM
  • NovaLead
  • KaiDev
  • ZaraDev
  • LunaQA
  • AtlasOps
  • ArgusSecurity
  • + ∞ any role you can describe

Scale

Not eight agents. As many as the work needs.

Ambient population is a closed-form function of time rather than an integration — every agent's state is computed from a seed and a clock, never accumulated frame to frame.

Three things follow. The server ships a seed instead of a population. Two clients cannot drift apart. And an agent nobody evaluates costs nothing at all — not merely less.

“State is a closed-form function of time, not an integration.”

Those numbers are what the world renders. The agents that think are backed by an LLM, and a turn is spent almost entirely waiting on your model provider rather than on a server — so nothing in the runtime caps how many you run. Your model budget and your provider's rate limits do. We publish no figure for that because we have not measured one.

  • 100,000agents on screen @ 60.0 fps
  • 1draw call
  • 25 Bper agent
  • 56.2fps on a phone
  • 322draw calls, scene
  • 0.14 µsper agent tick

Dev build capture · 2026-08-05 · high preset, dpr 1

Straight answers

What's shipping, and what isn't.

  • Shipping

    Agents including Argus the pre-merge security gate, MagyVerse, worktree isolation with auto-PR, four-tier memory, knowledge graph, per-agent model choice with your own keys across seven providers, Telegram, cron, skills.

  • Building

    Marathon mode's worker loop is a skeleton — it logs progress but doesn't yet dispatch real work.

  • Not yet

    Slack, Jira and Notion connectors are stubs. Pre-built binaries and the paid tier are next.

Magy is a commercial product, not open source. Licensing terms are being finalised — licensing@1martianway.com if you need specifics.

FAQ

Questions worth asking.

Does it run locally?

Yes. `magy assistant` is a single agent in a REPL with no NATS, no Postgres and no Redis — it runs on your machine with zero external services. The full eight-agent runtime and the 3D world need the production path.

How many agents can I actually run?

Two different numbers, and we keep them apart on purpose. The 100,000 figure is what the world renders — ambient agents drawn at 60 fps on a laptop and 56.2 on a phone. Agents that actually think are backed by an LLM, and an agent turn is spent almost entirely waiting on your model provider rather than on a server, so there is no cap on agent count in our runtime. What sets your ceiling is your model budget and your provider's rate limits — your keys, your call. We have not published a number for that because we have not measured one, and a figure we estimated would tell you less than this paragraph does.

Which models does it use — and do I control that?

You control it completely, per agent. Bring your own API keys or your Claude subscription, and assign whichever model you want to whichever agent: the strongest one for your CTO, something cheap for routine work. Seven providers are supported — Anthropic, OpenAI, Gemini, Bedrock, Ollama, Kimi and MiniMax. Set an agent to `auto` instead and a cost-aware router walks a ladder from the strongest model down, with circuit breakers and per-agent budgets. Keys live in your own config, never in our database.

Does my code leave my machine?

Your prompts and the code context go to whichever LLM provider you configure, the same as any AI coding tool. Magy adds a per-conversation TaskScope that hard-limits which repositories an agent may touch, and every task runs in its own git worktree.

Is it open source?

No. Magy is a commercial product. Licensing and pricing are being finalised — talk to us if you need specific terms.

What does it cost?

Pricing is not live yet. Early-access signups hear first, and founding-member pricing goes to that list.

When can I use it?

We are opening access in stages. Everyone on the early-access list is notified by email as soon as their turn comes up — no waiting on an announcement, no queue to check.

Get Early Access

One email when the first build is ready. Nothing else, ever.

One email. No spam, unsubscribe any time.

Also from 1 Martian Way: MOS · Toowl