Skip to guide
ASL AI Hub Master Suite

ASL AI Hub Master Suite help center

Support: support@autosecurelogin.com. Include the product version, the exact warning and its time, the last successful step, and a short description of what you expected. Never send passwords, viewer keys, dispatcher keys, private keys, provider tokens, whole private-state folders or customer files without reviewing the request and the material first. The app does not automatically upload a support bundle.

Start here: Master Suite in plain words

In plain words: Master Suite is a toolbox for people who use more than one computer, or who want one computer to work harder and more safely. It can check a computer's health and explain problems in ordinary language, let your computers share work between them, connect AI services, and plan multi-step jobs that run on the computers you choose and deliver the results where you want them.

Nothing happens by surprise. Opening the app starts no scan, no connection and no job. Every action that changes your computer or lets another computer use it asks you first, shows exactly what it will do, and can be stopped.

How to use this guide: each numbered section starts with a short In plain words explanation, followed by the exact technical details for whoever sets things up. You can read only the plain explanations to understand what something does, and come back to the details when you need them. Section 18 explains the words this guide uses; section 17 answers the most common questions; section 19 says where to start when something does not work.

A typical first week:

  1. Open PC health and run a scan to see how your computer is doing (section 9).
  2. If you have a second computer, join both to a fleet (section 3) and run the small practice job in section 11.
  3. When that works, describe your own work as a workflow (section 4), approve it and check the delivered result.

New in Master Suite 0.4.19 (rBroker 1.5.2 inside Resource hosting): send a complete Blender, Android or Windows project to another rBroker computer, run one fixed kind of task there and receive a verified result ZIP. Blender can use its own Python support and either CPU-only work or an owned GPU lease. Android can build, use one named emulator, install built APKs on one explicitly enabled local physical ADB device, or use a separately enabled direct-only ASL Tunnel device.

Windows can build ordinary, MSIX or Microsoft Store-ready candidates. Each receiving computer detects its own tools once and tells you what is missing; its device permission, paths, GPU configuration and signing identity stay there. The embedded rBroker workspace and its complete guide remain available in 14 languages. Section 12 has the setup and safety details.

1. What the four products do

In plain words: Master Suite is one Windows app that contains four tools. PC Steward checks your computer's health and explains what it finds. rBroker lets several of your computers work together and shares out their processors and graphics cards fairly. AI Hub Companion connects AI services and lets you plan work that runs on your computers. Master Suite puts all of them in one window, so you do not need to buy or install the others separately.

PC Steward inspects the computer, explains findings with available evidence and timing, and offers narrowly scoped, reviewed Windows actions. A health scan is not a repair. It is useful without a fleet or any other ASL purchase.

rBroker coordinates resource ownership and runs the local broker, canonical parallel worker, node reporting and approved execution components. It can also host a fleet coordinator. It is more than a monitor; hosting and remote execution are explicit controls, not consequences of opening a status page.

AI Hub Companion connects supported AI providers and lets you construct, approve and advance described fleet workflows. It does not grant resources on a remote PC itself. The remote node makes the final admission decision.

Master Suite embeds their shared native workspaces in one Windows application. It bundles its own engines and approved-action helpers and does not open the three other desktop programs. It remains useful by itself. You do not have to purchase or install the separate apps to use Suite's included functionality. Standalone apps remain independently usable and are not unlocked by running Suite.

Each product keeps its own local profiles and history. Suite does not silently copy a standalone app's credentials or take over a running node. A single machine must not run two competing authorities for the same GPU or parallel-worker port. An “already managed” warning means inspect the existing owner and deliberately select the appropriate connection or hosting components, not kill its process.

2. First launch and choosing a workspace

In plain words: Opening Master Suite never starts anything by itself: no scan, no connection and no jobs. Each area on the left is a separate workspace you open when you need it. Status cards show what was true when they were last read, so always check how old a reading is.

Opening Suite does not start a health scan, connect a provider, join a fleet, execute a task, change cooling, install a service or authorize cleanup.

Status cards are observations, not allocations. Read their sample time and stale indicator. An old green snapshot does not mean a machine is still online. Optional monitor modes refresh while enabled; they are not required for background services or a coordinator to function, and are not AI conversations consuming tokens.

3. Connect computers on different networks

In plain words: To link computers in different places, they all need one shared meeting point on the internet with a secure web address: the fleet coordinator. Each computer contacts that meeting point itself, so you do not need to open ports on your router. Joining a fleet only introduces a computer; it does not let anyone run work on it.

Use a fleet coordinator reachable by every participating computer over HTTPS with a valid public certificate. The server address is configurable; it is not restricted to autosecurelogin.com. A Linux server, a Windows server with a suitable public HTTPS front end, or your own managed host can provide this endpoint.

The coordinator holds membership, scoped job state and transfer receipts. Nodes make outbound connections, so the participating PCs do not need public GPU-broker ports or router forwarding to their local worker services. Do not expose the loopback GPU or parallel-worker ports to the internet. The coordinator and its HTTPS front end are separate from local resource admission.

A hub of your own, and reaching it from the internet. In plain words: on Resource hosting, *Host a fleet for my LAN* makes this computer the meeting point for the computers around it. It mints its own certificate, listens on a LAN port and shows a one-line invitation; another computer pastes the line and joins, trusting the hub only after its certificate matched the invitation.

To let a computer anywhere join, *Reach this hub from the internet: ASL Tunnel* takes a pairing code from the ASL Tunnel control plane and gives the hub a public address (https://tunnel.autosecurelogin.com/t/<route>) through one outbound connection from this computer: no port forwarding, only the hub's own port forwarded, nothing else on this computer exposed, and every request still needs the fleet's keys.

The invitation then carries that address after the network ones. The Addresses card lists this computer's network addresses, the hub's and, on request, the public address the internet sees. This hub route is separate from the direct-only Android device route in section 12; turning on one does not turn on the other.

Technically: the hub is fleet coordinator 0.19.0 with its own ECDSA certificate on a LAN HTTPS port beside its loopback port; the invitation carries the hub's addresses, the fleet id, the join key and the certificate's SHA-256 fingerprint, and joining computers pin that certificate per address.

The tunnel's public address is an origin with a short path prefix; such a *relayed* origin is verified by platform trust (the tunnel's public name has a certificate from a public authority) and is never pinned, in every client: the embedded RBroker.Core, the bundled Python and Node, and the phone app.

The agent runs from the Suite's package (fleet/tunnel-runtime, the tunnel repository's own code) and forwards only the hub's loopback port; the gateway's forwarded caller address is honoured, so /v1/address tells an internet caller its public address. Jobs from outside the network travel through the hub's store, as from phones; direct peer transfer is preferred where the two authorized computers can reach each other.

On each participating Windows PC, open Resource hosting, review its name and local configuration, and connect to the intended fleet. Join fleet sends the public enrollment record; JSON export/import remains available for an administrator who prefers that workflow. A node's private signing key stays on the node.

Membership means the fleet recognizes this computer. It does not authorize remote execution. Review the permitted dispatcher, handlers, resource limits and authorization lifetime on that computer before starting the executor. Viewer, enrollment, management and dispatcher credentials have different roles. A viewer key cannot be used to execute work, and a node's signing key is not a viewer key.

If enrollment fails, retain the existing identity and retry the intended join. Do not generate a replacement identity to make a warning disappear. A changed fingerprint is a different node until explicitly reviewed and reconciled.

4. Build a workflow that people can understand

In plain words: A workflow is a written plan of one or more jobs, for example: prepare the data on the big computer, then send the result to my laptop. You describe each step in ordinary words, choose which computer does it and how much of that computer it may use, check the plan and approve it. Nothing runs until you approve the exact plan.

Connect the workflow workspace to the correct customer fleet using its dispatcher credential. Give the overall task and each step a clear title, project, purpose, stage and expected result. For example, “Prepare forest-scene data for episode 1” explains more than a session number. IDs remain available for exact correlation, but are not used as a substitute for meaningful descriptions.

Choose the actual node and supported handler for each step, the CPU-worker count, system-memory grant and any GPU-memory requirement. Add dependencies so a later step does not begin until the earlier result is verified. Select the intended input files and final output destination explicitly. Read and approve the exact saved plan before advancing it; editing a plan is not approval of different work.

The currently reviewed handler set includes program bundles, bounded JSON work, hashing, the supported shape-animation/media workflow and typed Blender, Android and Windows project jobs. A handler is a locally registered capability, not arbitrary remote shell access. The program does not install an AI model, purchase provider access or turn an unrestricted prompt into arbitrary code execution on every joined PC. Supported media work needs a reviewed local encoder and appropriate hardware/driver configuration. An encoder found on the development PC is not automatically redistributed or licensed for customers.

A practical distributed example is preparing scene data on a CPU/RAM-rich node, running supported media work on the GPU node, and verifying/delivering the result to the frequently used laptop. The workflow engine coordinates dependencies and receipts; it does not promise artistic quality or support for unregistered creative tools. Review the actual supported capabilities before assigning production work.

Choose Run a program with its data for general CPU computation. Select a ZIP containing a trusted Python program or Windows executable and its data. The compressed bundle limit is 16 GiB, the expanded-input limit is 32 GiB, and declared returned files total at most 16 GiB, plus bounded logs. Actual free space on the sending computer, worker and destination can impose a lower limit. The small JSON description is separate from the file bytes; it is not a one-megabyte file limit.

Enter the relative entry point, one argument per line, and the relative files to return. Review the bundle size and SHA-256, target computer, RAM, time limit and Maximum returned data in MiB (1–16384) before approving. The return limit defaults to 2048 MiB; increase it deliberately when your expected result is larger. One bundle uses one worker; independent steps may target different computers. Program results are ZIP archives and cannot be used as JSON dependencies. General program bundles are CPU-only; selecting a GPU does not make an arbitrary program or Blender project a distributed GPU render.

Each receiving computer separately allows program work and grants the named dispatcher access. Old JSON consent and upgrades do not enable it. Programs run with the worker Windows account's access, so send only trusted code. CPU/RAM limits and process-tree shutdown are not an untrusted-code security sandbox. Bundled Python is included; other executable dependencies belong in the ZIP. Keep free space for input chunks, verified copies, expanded files and results. Interrupted transfers resume on the original workflow; a bounded client operation may pause after five minutes, and Run / reconcile one pass continues it without reposting accepted work.

5. System RAM and GPU RAM are different budgets

In plain words: A computer has two separate memories: the main memory every program shares (system RAM) and the graphics card's own memory (GPU RAM). They are separate budgets; spare main memory cannot make up for a full graphics card. Jobs that use the graphics card must book its memory first, like reserving a room.

System RAM is the computer's main memory used by CPU workers and applications. GPU RAM is the graphics card's dedicated memory. A free amount in one pool cannot be substituted for a shortage in the other. The parallel worker enforces its declared system-memory/worker commitments; GPU workloads additionally need an owned GPU lease from the local broker.

A lease is a time-limited reservation for a specific owner and attempt, not proof that the full reservation is currently allocated by the driver. Conversely, an unleased GPU process can consume real graphics memory. Compare leased commitments, observed processes and telemetry timestamps; do not add Windows allocation references as if they were all exclusive dedicated VRAM.

Local policy sets the system allowance and managed ceiling. The OMEN development configuration uses one 5,120 MiB Windows allowance, not two reserves. Customers must size their own policy for their hardware. A 28 GB lease is only possible on a node whose verified capacity and policy admit it; it is not promised on every graphics card. Missing or inconsistent telemetry is “unknown,” never free memory.

Heartbeats keep an admitted attempt alive only until the server-confirmed expiry. Authentication rejection is not a temporary networking stall and is not permission to continue using the GPU. Timeout-aware retry must fit inside the remaining confirmed deadline. The workload must stop safely when ownership is lost; it must not obtain another session's lease or keep running under an invented extension.

6. Files, dependencies and final destinations

In plain words: Files for a job travel directly between the two computers, encrypted, like a private courier. The fleet server only handles the paperwork, never the files. If the connection breaks, the transfer picks up where it stopped. A file only counts as delivered once it has been checked and placed in the folder you chose.

Transfers are scoped to the approved fleet, workflow/job and artifact namespace. The sender sends encrypted chunks directly to the receiving computer. Keep both apps open. The coordinator keeps metadata only and never provides a bulk-data fallback. The receiver verifies the completed content before publishing it to the chosen destination. A completed render is not the same as a file delivered to your laptop. Check the final delivery receipt.

If the connection breaks, resume the original transfer. Already confirmed chunks are reused and only missing chunks are sent. A lost acknowledgement is uncertain until reconciled with the original transfer, not an instruction to render again. A partial destination remains partial and is not labeled as the finished output. Existing unrelated destination files must not be silently overwritten.

Files travel directly between the two authorized computers over an encrypted peer connection. The coordinator retains permissions, job state, manifests and connection messages; it does not store or relay new input or result bytes. Keep both applications open and both computers online until delivery is verified. A service-hosted CPU node has its own separately approved lifecycle; closing the sending application can still prevent file delivery.

Keep enough local storage for verified chunks, the bundle snapshot, expanded inputs and results. A 2 GiB ZIP can require considerably more than 2 GiB of free worker space. Files use 256 KiB resumable chunks and smaller encrypted network frames. Previously accepted server files remain readable for recovery; new transfers never fall back to uploading content to the coordinator.

Different networks can work when the computers establish a direct connection. Restrictive NAT or firewall rules can prevent this even when HTTPS fleet status works. A connection-discovery service exchanges network addresses, not file contents. Leave the original workflow and partial data intact, restore permitted connectivity and use Run / reconcile one pass. Do not expose local GPU/CPU ports, disable protection, or create a duplicate job to bypass a blocked connection.

7. Recovery without duplicate jobs

In plain words: If something is interrupted, reopen the same workflow instead of starting a new one. Master Suite then works out what already happened, so the same job is never accidentally run twice. Unknown means not confirmed yet; it does not mean failed.

After an interruption, reopen the same saved profile and original workflow. Keep its node fingerprint, plan hash, job/attempt IDs and receipts. Unknown means the final outcome has not been confirmed, not failed and safe to repeat.

An offer whose acknowledgement was lost is reconciled against the original job. The explicit exact-offer retry sends the same saved request; it does not create new work just because a status call was temporarily unavailable. A changed node identity, changed scope, damaged encrypted store or missing initialized ledger requires recovery. None is treated as an empty successful installation.

Cancellation is a request. Claimed ownership remains until the original node reports the stopped outcome. Do not stop shared brokers or kill a worker simply because one screen has stale information. Review the original owner and process creation time, not the PID alone.

Finished workflows can be archived explicitly when their requested destinations are verified. Archive history preserves readable plans and receipts in protected storage, prevents reuse of archived workflow IDs and does not delete your output files. Pending or uncertain work cannot be hidden by archiving it. See the bundled FLEET-WORKFLOW-ARCHIVES.md for retention and recovery details.

8. Always-running service versus an open desktop app

In plain words: Normally work happens while Master Suite is open. If a computer should keep working when the app is closed or nobody is signed in, you can install an optional background service, which needs Windows administrator approval. Before updating that service, let its current work finish.

The optional Windows service is installed through the reviewed service controls and Windows administrator approval. It runs the selected resource/coordinator/ reporting components independently of the desktop window. It is not installed merely by downloading the Store package. The desktop remains usable without it.

Service storage and identity are separate from the interactive user's profile. Select only the components this machine should own. Service-hosted execution has its own explicit CPU-only consent and resource/dispatcher boundaries; it is not automatic permission to execute GPU media or renew an expired authorization. Interactive sensor and cooling controls are not moved into Windows Session 0.

Before upgrading a service, drain accepted work and stop only the owned components. The reviewed upgrade transaction retains the original identity and configuration, checks the stopped process identity, and keeps the replacement stopped until its payload/configuration are verified. It does not silently enable new components. If an upgrade stops partway through, retain the pending record and resume the intended upgrade; do not delete its evidence or point an older writer at newer data.

Closing an app requests safe shutdown of what that app owns. Switching tabs does not cancel work. An active service can continue without the desktop, while a desktop-owned executor must drain safely. An unconfirmed shutdown keeps the orderly-close guard visible rather than claiming that every job stopped.

9. Diagnostics, temperatures, fans and cleanup

In plain words: PC health looks at your computer and explains what it finds, with dates, so you can tell an old problem from a current one. Temperatures and fan speeds only appear when your hardware reports them. Nothing is repaired, cleaned up or changed until you have reviewed it and said yes.

Expand a health issue to see what was observed, available timing, affected target and the recommended next step. “Last checked,” “last ran” and “last failed” mean different things. Historical event evidence is not proof that a fault exists now.

The health score reflects current checks from its scan. Retained crash/update history and recovered, stale or unconfirmed evidence remain visible but do not deduct points. Many old incidents cannot alone turn a healthy current sample into an immediate emergency. This does not prove recovery; review exact dates and verify current behavior. Current resource pressure and actionable findings still affect the score. The score is not a performance guarantee or certification.

Temperature, hotspot, power and RPM values depend on the hardware/driver exposing them. Missing is not zero. A fan percentage is not a physical RPM measurement, and one GPU-wide RPM reading does not identify every physical fan. Cooling control is only offered when the documented adapter reports support.

A temporary cooling boost is bounded and restores the reviewed policy. NVIDIA automatic control and AMD's saved curve/Zero RPM settings are different restoration paths. Do not run competing fan tools or force-stop a restoring helper. The app does not bundle or install a replacement display driver; see AMD_COOLING.md.

Optional AMD component terms: in PC health → Temperatures & Cooling, choose Review AMD component terms. Read the full versioned text and use Save terms for a local copy. Saving or closing does not accept. Choose Accept AMD terms only if you agree and have authority to do so, or Decline AMD adapter to keep it disabled. No acceptance is preselected or inferred from purchase/installation. The full text is also bundled as Help/AMD-COMPONENT-TERMS.md, version 2026-09-13.1. These terms concern only the AMD adapter, not your own work or the products' unrelated features. The app does not install the SDK or a display driver.

Your decision, time, terms version and fingerprint are kept in Suite's own local health profile. Standalone Steward keeps a separate record; neither app copies the other's consent. Missing, corrupted or outdated consent blocks new AMD reads and boosts. Review new terms explicitly; acceptance does not launch a scan or cooling action. Refresh deliberately after deciding. Unavailable sensors remain unknown, not zero. NVIDIA, Windows diagnostics, AI and resource hosting are usable if AMD is declined. No consent or sensor content is sent to a fleet server.

You can decline later through the same dialog. This does not stop restoration already underway, and the consent check does not block an explicit restore of previously saved AMD settings. Restoration still needs a supported exact device, a valid protected saved profile and normal Windows approval. It cannot start a new boost or invent a default profile. If recovery cannot be confirmed, preserve the evidence, use the manufacturer's controls and contact support.

Cleanup is off by default. Previewed deletion is permanent, not the Recycle Bin. The bounded categories, age limits and exact preview are shown before removal. The optional daily in-app schedule is separately authorized and stops when the app closes; it is not an invisible system-wide cleanup service. Project/model directories are not automatically chosen because a drive has low space.

Damaged action history is visible and blocks repetition of Windows actions. It is not skipped or treated as empty. Preserve the original file and reconcile any running helper before repair; see ACTION-HISTORY-RECOVERY.md.

10. Platform, privacy and purchasing expectations

In plain words: Master Suite runs on 64-bit Windows 11. Your purchase covers the app, not outside services such as AI providers or a server you run yourself. Only computers you deliberately set up receive work. If you see harmful content, Report inappropriate content opens an email draft to support; nothing is sent automatically.

Master Suite and PC Steward require supported 64-bit Windows 11 versions. rBroker and AI Hub Companion have separate Windows compatibility declarations. Linux headless nodes use the portable contracts and outbound HTTPS, not a Windows UI running remotely. Apple scaffolding is not a claim of tested macOS hardware support.

Fleet-server hosting, electricity, third-party models, encoders and provider/API usage may have their own costs or licenses. Purchasing an ASL desktop product does not include unlimited third-party AI usage. The Microsoft Store checkout shows the current desktop price for your region and taxes. Third-party subscriptions and customer-operated hosting are separate.

Only deliberately configured participants receive scoped work and transfers. Readable descriptions should not contain private prompts, credentials or sensitive paths just to improve a label. Review exports and coordinator hosting policy for your organization. Back up private identity, consent and workflow storage securely; copying public enrollment JSON alone is not a private-key backup.

Content safety and reporting: open Help, expand Complete help center, and choose Report inappropriate content. The shared native control opens a draft to support@autosecurelogin.com identifying Master Suite, not a different product. Nothing is sent automatically. It reads or attaches no task, prompt, result, identity, key, file or log. Enter your own brief concern, app version and time; do not send harmful media or confidential material. If no email app is configured, use the displayed address yourself. Opening a draft is not a submitted report.

Use only authorized content and services. Do not use ASL tools for child sexual exploitation, non-consensual intimate content, threats, hateful abuse, fraud, malware or other illegal harm. Do not request or distribute pornography or graphic real-world violence. Respect privacy, intellectual property, provider policies and fleet-owner rules. Review generated material before sharing it. There is no public feed or promise of automatic moderation. Also contact the private fleet's owner; stop automatic advancement and use supported cancellation or consent controls while preserving unresolved ownership and recovery state. ASL reviews concerns involving its products and operated services and can restrict access where it has control. A report cannot silently delete another server operator's retained copies. No response-time guarantee is implied.

Reading diagnostic dates correctly

In This computer, an expanded finding uses the same explanation and timing rules as PC Steward. Recorded Windows events are converted to your computer's local time and show both the time-zone name and UTC offset. The time the scan checked history is separate from the first and last retained failure. A historical crash is not a claim that an app is crashing now; incomplete history is labeled, and a future timestamp tells you to check the clock. The detail includes evidence, confidence, recommended next steps, guided checks and their impact. Reading these details does not run a repair, change drivers or stop a workload.

11. First complete CPU workflow

In plain words: A first practice run with a small, harmless job: join a computer as a worker, allow it to accept work, connect the workflow screen, describe a tiny job, approve it and watch the result arrive. Doing this once teaches you the whole process.

  1. On the worker computer, open rBroker Connect this machine, enter the fleet address and join key, then choose Join as a CPU worker. This starts or connects to the reviewed local worker; it does not approve programs, folders or GPU work.
  2. In the worker's Fleet page, allow the required handler with modest worker/RAM limits and an expiry. The fleet manager separately grants the same named dispatcher the required scope. Enable automatic fleet execution only after both choices are confirmed. A service identity needs its own setup and consent.
  3. In Fleet workflows → Connect, enter that coordinator's HTTPS address and separate dispatcher key. Choose Connect and protect profile and check the intended computer's identity and capabilities.
  4. Under Build a workflow, describe the workflow and step, select Prepare or copy JSON data, enter harmless sample text, request one worker with a modest system-RAM amount, and choose a new final output folder and filename.
  5. Select Add step to plan, then Save private draft for review. Under Saved work, choose Read selected plan and progress. Check the computer, input, handler, resources and destination before Approve reviewed plan.
  6. Choose Run / reconcile one pass. Automatic progress is a separate opt-in for the selected workflow. Follow the stage until the verified result is delivered and the node releases its resources. Keep both apps open through file delivery.
  7. Reopen the same protected profile to inspect or reconcile the saved work. Choose Archive verified finished workflow only after completion and any selected delivery are confirmed. Archiving keeps the receipt and does not erase your output files.

Use a small non-sensitive job first. Approval covers the saved plan, not later edits. An enrolled computer with no execution consent can report its health while refusing every workflow.

12. NVIDIA, AMD, Blender, Android and Windows projects

In plain words: Master Suite can send a whole Blender, Android or Windows project to another computer that has the right tool installed and explicitly enabled. The receiving computer tells you what it needs, keeps its own paths, device permissions and signing identity, and returns a checked ZIP. These are fixed project recipes, not open remote command access. The earlier NVIDIA/AMD shape-animation choices and pictures from a prompt with the owner's own image program remain available separately.

The two GPU choices are Encode NVIDIA shape animation and Encode AMD shape animation. These are bounded shape-animation encoders. On the receiving node, open Host & setup → Local GPU media setup, choose the matching NVIDIA (NVENC) or AMD (AMF) backend and the intended existing GPU authority, then Choose and validate local media tools. Compatible FFmpeg and the graphics driver must already be available. An explicitly selected GPU Session Broker 1.3.2 can remain the authority.

Enable GPU media for that app-owned worker, and separately permit the matching handler in both node consent and dispatcher permission. An already-running worker does not gain a changed media configuration; drain owned work before deliberately restarting that owned worker. Never stop an unrelated broker. The GPU attempt needs its own GPU lease as well as CPU-worker and system-RAM admission. A backend mismatch or unavailable encoder fails visibly; it does not quietly switch to CPU.

Set up a computer that will receive projects

  1. On that computer open Resource hosting → Host & setup → Project task tools and choose Detect installed tools. Empty fields show what must be installed or selected. rBroker itself bundles its own runtimes, but it does not secretly install Blender, Android Studio/SDK, Java or Visual Studio Build Tools.
  2. Enable only the categories this computer should accept. Blender Python, Android emulator, local physical ADB, direct-only ASL Tunnel Android access and Microsoft Store packaging each have their own choice. Local physical ADB and Tunnel are independent: choosing Tunnel does not enable USB/local physical ADB, and choosing physical ADB does not authorize Tunnel. An Android install still targets exactly one intended, authorized device through the selected path.
  3. Review the detected executable/folder fields. A GPU Blender task also needs this computer's existing rBroker GPU-broker configuration. A Store-ready Windows build may use the optional certificate thumbprint already installed here. Passwords, private keys, token arguments and remote signing material are rejected.
  4. Choose Save and verify project setup. This read-only validation does not build, start an emulator, connect ADB, sign, reserve a GPU or execute a project. Then enable the selected project handlers and deliberately restart only this app's idle private worker.
  5. On the Fleet page, give local project consent and have the fleet manager grant the named dispatcher the same typed handlers. Neither permission can override missing tools or local resource admission.

Send a project

On Resource hosting → Fleet, choose the project folder and one of the five included small JSON recipes, refresh ready computers, select a computer or leave automatic nearest-ready selection, then choose Send project and get results. Keep both apps online until the verified result ZIP is delivered. The same dispatcher credential used by the document card is used only after the fleet owner grants it project access. Claude, ChatGPT or another local MCP client can read readiness and guide this setup through the same checks without receiving fleet secrets.

The project folder travels through rBroker's existing signed-job and encrypted, resumable, digest-verified artifact path. Installed tool paths, device authority, signing identity and GPU configuration never come from the transferred bundle. The returned archive is checked against the recipe's paths and size limit before delivery.

13. Diagnose a transfer without repeating the computation

In plain words: When a delivery stalls, these checks tell you which part is stuck, so you can fix that part without running the whole job again.

14. Version and platform checklist

In plain words: Every computer in a fleet needs a recent enough version for files to move directly between them. Check the version actually installed on each computer; reading this guide does not update anything.

The current direct-transfer workflow requires coordinator 0.16.0 or newer, compatible rBroker 1.4.5 or newer nodes, and AI Hub Companion 0.14.10 or newer or Master Suite 0.4.14 or newer. This help revision accompanies Master Suite 0.4.19 with rBroker 1.5.2 and coordinator 0.19.2. Typed Blender, Android and Windows project jobs require rBroker 1.5.0 on both participating computers or this Master Suite release at an endpoint; the separately typed direct-only Tunnel Android target requires rBroker 1.5.2 or this Suite release on the receiving computer and the current ASL Tunnel client.

They do not raise the minimum for older direct file work. Picture jobs additionally need rBroker 1.4.7 or newer on the computer that makes the pictures and fleet coordinator 0.17.1 or newer. Older status-only connections may still work without supporting current execution or file transfers.

Compare the versions actually installed on every participating computer. Check the selected handler and driver on the actual worker, and keep its configured GPU authority unchanged. The Windows desktop and included peer runtime are the qualified path here; older headless Linux examples do not establish current direct-transfer hardware acceptance on Linux or macOS. A package's Windows minimum version is separate from physical testing on your Windows edition.

15. Pictures from a prompt with your own image program

In plain words: A picture job is a short written description, called a prompt (for example: a red bicycle leaning against a white brick wall, morning light), that someone in your fleet sends to a computer. That computer turns the description into one to four pictures and sends them back. Master Suite can now let your computer make these pictures with an image program you installed, on your graphics card, for senders you allow. It does not include an image program or AI model, and it never uses anyone else's computer or image program.

Think of a print shop where you own the printer. You decide whether your computer makes pictures, which image program it uses, how much of the graphics card it may use and who may ask. The sender writes the prompt, chooses the shape and how many pictures, and receives the files. rBroker, inside Master Suite, carries the request, checks it, reserves graphics memory for exactly that job, starts your image program, checks every file and delivers the pictures straight to the sender, encrypted.

What you need

Switch it on, step by step

  1. Open Resource hosting, then Host & setup, and scroll to Pictures from a prompt · your own generator.
  2. Choose Choose and validate my picture generator. Three file windows open in turn: pick the python.exe that belongs to your image program, then the image program's script (.py), then this computer's GPU broker configuration file (the same one GPU media uses). This checks the files; it does not run your program or reserve memory.
  3. When the message says Local paths validated and saved, tick Enable pictures when I next start this app's private Parallel worker, then start the worker (stop it while idle and start it again if it was already running).
  4. On the Fleet page, tick Pictures from a prompt with my own generator, enter at least 14336 for GPU memory and 18432 for system RAM, one job at a time and an expiry, and choose Allow. You can withdraw this at any moment with Block new fleet work locally.
  5. Ask the fleet's manager to give the named sender permission for pictures on this computer. Until they do, no sender can reach it.
  6. Turn on automatic fleet execution when you want jobs picked up without pressing a button for each one.

What happens when a picture job arrives

  1. Your computer checks the job against your permission: the right sender, the right kind of work, within your memory limits and before the expiry.
  2. Prompts asking for sexual or nude pictures, or for children in revealing situations, are refused before anything starts.
  3. If the graphics card is busy, the job waits for room instead of failing. It holds nothing while it waits and gives up after an hour without having used the card.
  4. rBroker reserves exactly the job's graphics memory and starts your image program.
  5. rBroker checks every file your program writes: the right size, the right number, unaltered and belonging to this request.
  6. The pictures travel directly to the sender, encrypted. The fleet server never receives them.
  7. The graphics memory is released the moment your program finishes. Nothing stays reserved between jobs.

Safety, privacy and switching it off

Your image program should run its own safety check on every picture. Pictures it holds back are never delivered; if it holds back every picture, the job ends with nothing delivered. You remain responsible for what your computer produces, for your image program's licence and for the law where you are. Only allow senders you trust.

The fleet server sees the job's readable title, which begins with the first part of the prompt (up to about 160 characters), the sender's name, your computer's name and the job's progress; never the pictures. Your computer sees the whole prompt. A copy of each finished result stays in rBroker's private storage on your computer (Open my private configuration folder, then image and outputs); delete old results when you no longer need them.

To stop straight away, choose Block new fleet work locally, or allow your permission again without the picture box ticked. To stop offering pictures at all, untick Enable pictures and restart the worker while it is idle. A picture already being made finishes first.

Sending picture jobs from Workflows

In plain words: you send a picture job from Workflows like any other step. Pick the computer, choose Make pictures from a prompt (on that computer's own image program), describe the picture, choose the shape and how many, choose the folder, then save, read, approve and run the plan. The pictures arrive in your folder as pictures.zip, which Windows opens like a folder.

  1. Open Workflows, connect with your dispatcher key (section 4) and describe the workflow.
  2. Choose the computer, then the work type Make pictures from a prompt (on that computer's own image program). The fixed resources are filled in: one worker, 18,432 MiB of system RAM and 14,336 MiB of GPU memory.
  3. Type the description in Describe the picture (3 to 1,000 characters), anything to avoid in Leave out (up to 300), the Shape and How many pictures (1 to 4). A prompt that would be refused is explained straight away and never added.
  4. Choose the final output folder (pictures.zip by default; an existing file is never replaced), choose Add step to plan, then Save private draft for review.
  5. In Saved work, read the plan: it shows the exact prompt, shape, count and leave-out text. Approve it, then choose Run / reconcile one pass or turn on Automatic progress, and wait for delivered.

Technical: the step offers media.sdxl-image with inline data of exactly prompt, avoid, shape and count under the item result. The app checks the request with the same rules as rBroker's media-image.mjs before saving (a test keeps the word lists identical), and rBroker's offer validation, the coordinator and the receiving computer check it again. A picture step cannot be the JSON input of a later step. The result is received directly over the encrypted peer connection and verified before it is saved. Programs and websites can also send picture jobs with rBroker's picture tool (fleet/picture-job.mjs in the rBroker source). A sender needs a dispatcher key from the fleet's manager and permission for pictures on the chosen computer.

For whoever installs the image program: the exact rules

This part is for the person who installs or writes the image program; the computer's owner does not need it. rBroker works with any image program that follows these rules, with no tie to one model or company.

For each job rBroker first reserves the graphics memory, then runs: python.exe -I your_script.py oneshot --request REQUEST.json --out FOLDER. Two environment settings tell the program its reservation: ASL_GPU_LEASE_ID and ASL_GPU_GRANTED_MIB. The program must stay within the granted MiB and finish within 900 seconds, or rBroker stops it and the job fails. The -I switch means it does not see the worker's own Python settings, so its own environment must hold everything it needs; keep the model on local disk.

REQUEST.json holds exactly four values, for example: {"prompt": "a red bicycle leaning against a white brick wall, morning light", "avoid": "text, watermark", "shape": "landscape", "count": 2}. The prompt is 3 to 1,000 characters; avoid is 0 to 300 characters; shape is square (1024 x 1024), portrait (832 x 1216) or landscape (1216 x 832), and pictures must be exactly that size; count is 1 to 4.

The program writes two files into FOLDER:

Exit code 0 means at least one picture is in the zip; exit code 3 means your safety check held back every picture (nothing is delivered); any other exit code is a failure. rBroker never falls back to the processor and never delivers partial output.

Check it by hand before allowing jobs: write a REQUEST.json into an empty folder, run the command above yourself with your own python.exe and an empty output folder, confirm result.zip opens and holds stored (not compressed) numbered PNGs and recipe.json, compare each PNG's SHA-256 (for example certutil -hashfile 1.png SHA256) with summary.json, and watch that the graphics memory used stays inside 14,336 MiB. The rBroker help guide, sections 37 and 38, has the same rules with example files.

16. When the internet or the fleet server drops for a moment

In plain words: If your internet connection or the fleet server is unavailable for a moment, your computer's automatic fleet work no longer stops. It keeps all saved work, shows the time it will try again, waits and carries on by itself. Before this release, one failed check stopped automatic work until you switched it on again.

The status then reads Fleet or worker unreachable - saved work kept; trying again at, followed by a time. The first wait is one minute; while the outage lasts each wait doubles (two, four, eight minutes) up to fifteen minutes, so a long outage does not flood the server. The same rule applies to the always-running service.

Technical detail: only transport failures count as an outage: a dropped connection, a request that ran out of time, or a busy or unavailable answer from the server (HTTP 429 or 5xx). Anything else, such as a job whose outcome is uncertain, a refused permission or a damaged record, still pauses automatic work and waits for you to review it, exactly as before. While you are draining (stopping safely), an outage pauses instead of waiting, and saved ownership is always kept. A job whose submission is uncertain still blocks new claims until it is reconciled.

17. Frequently asked questions

Does opening Master Suite change anything on my computer?

No. Opening it starts no scan, connection, job, cleanup or service. Each of those is a separate choice you make, and each shows what it will do first.

Can another computer use my graphics card or processor without asking me?

No. Joining a fleet only introduces your computer. Work runs only when you allow that kind of work on your computer and the fleet's manager gives a named sender permission. You can block new work instantly with Block new fleet work locally.

Does Master Suite include an AI model or an image generator?

No. AI Hub Companion connects AI providers you already use, under their own terms and costs. For pictures from a prompt you install your own image program (section 15).

Do my files go through ASL's servers?

No. Files travel directly between your two computers, encrypted. The fleet server only carries permissions, short descriptions and connection messages (section 6).

What happens if my internet drops while work is running?

Nothing is lost. Transfers continue where they stopped once the connection returns, and automatic fleet work tries again by itself (section 16).

Why does a computer show as online but not take any work?

Being online only means it reports its status. It also needs your permission for that kind of work, the fleet manager's permission for the sender, enough free processor workers and memory, and an unexpired permission (sections 3 and 13).

How do I send a picture job without writing code?

Use Workflows and choose the work type Make pictures from a prompt (section 15, Sending picture jobs from Workflows). The computer that makes the pictures must have its owner's own image program set up and pictures allowed.

Do I need to buy rBroker, PC Steward or AI Hub Companion separately?

No. Master Suite includes their workspaces. The standalone apps are separate products for people who want only one of them.

Is Master Suite's price a subscription?

The Microsoft Store shows the current price for your region. Outside services you choose, such as AI providers or your own server, are paid separately.

18. Words you will see

In plain words: The words this guide and the app use, explained without jargon.

19. Something is not working: where to start

In plain words: Find the line that matches what you see and do the first check. Most problems are a helper that is not started, a key or address that is wrong, a missing permission, or information on screen that is simply old. Please try these before reinstalling anything.

Guide updated October 25, 2026. Generated from the same guide embedded in the native product.

Help / AyudaHelp / Ayuda