Platform and integration · Coming Soon
SimpleForge

SimpleForge
Zero-retention AI infrastructure for firms whose client data is not allowed to rest anywhere.

For CISOs, Chief Privacy Officers, and Managing Partners at legal, healthcare, financial services, defense, and life sciences firms with hard prohibitions on client data leaving the workstation.

Some organizations restrict where sensitive information may be processed because of contracts, internal policy, or regulatory requirements. Whether a model service is appropriate depends on its configuration, agreements, controls, and the data involved; HIPAA and SOC 2 do not impose a blanket prohibition on all hosted language models. SimpleForge is being built as the substrate underneath that constraint: a workflow engine that holds the data a run touches in memory only, purges it when the run ends, and sweeps anything a crash leaves behind, so there is no disk, database, or cache holding client content. The retention machinery is built and tested. The model that would run under it is not, and until it does, SimpleForge protects a workflow rather than an inference.

Replaces Hosted model services that do not meet a particular organization's approved data requirements Custom on-prem inference rigs Air-gapped self-hosted Llama / Mistral deployments
Web In development Copilot Planned Desktop In development

When you can get SimpleForge

Coming soon

Not open yet. Tell us where to send word when it is.

What you can do now

Get product updates

None of these creates an account, a trial, or access. Each one records what you asked for so a person can answer it.

Where this stands

Availability
Coming soon

What it costs

No published price yet. Pricing is set before this product opens.

Routes that are not open

  • Direct trial with Simple Intelligence Not open yet
  • Direct subscription from Simple Intelligence Not open yet
  • Microsoft Marketplace Not published yet
  • Sponsored subscription for member companies Not open yet

The constraint this is built under.

A firm that holds privileged, regulated, or export-controlled material is not choosing between models. It is choosing between using AI on that material and not using AI on it, and the deciding question is where the data goes and what is left afterwards. Most answers to that question are policies: a retention setting, a contractual promise, a configuration somebody has to keep correct. SimpleForge is being built so the answer is a property of how the thing is constructed instead.

What that means in the engine.

A run holds what it needs in memory and nowhere else. The slot is overwritten before it is released rather than left for the garbage collector, a timer sweeps anything an interrupted run left behind, and clearing the buffer emits its own audit record. What leaves the run is metadata: which run, how long it took, whether it succeeded. There is no database, no cache, and no disk holding customer content, because none was ever added.

And the part that is not built.

A retention guarantee matters when a model is reading the data it protects. No model is served here yet. The only AI integration wired today is an outbound call to a third-party hosted research service, which is the opposite of local inference and is not what this product is for. Until a model runs under the retention rule, SimpleForge is a tested foundation rather than a thing a regulated buyer can use, and that is the reason it is Coming Soon rather than in preview.

What runs today, and what the retention rule is still waiting for.

The retention machinery is real code with tests behind it. The inference it exists to protect is specified and has nothing behind it. Both halves are stated here rather than one of them.

Working today

  • The workflow engine runs a graph of steps and streams progress while a run is in flight
  • Run data is held in a locked in-memory buffer, overwritten before release, and swept on a timer whether or not the run ended cleanly
  • Clearing the buffer emits an audit event carrying run metadata and no run content
  • The desktop client is a working shell that connects to the engine over local loopback
  • Microsoft Sentinel is wired as the audit destination, and Entra managed identity is the credential path

Not yet

  • No language model is served. On-premises and disconnected inference is specified and has no implementation
  • The one AI integration that is wired calls a third-party hosted service over the network, which is not the local-inference design and is not offered as a feature
  • The desktop client is an early shell and has not been submitted to the Microsoft Store or the Mac App Store
  • The MISA designation is a target. It is not held, and nothing on this page should be read as saying it is
  • Nothing is deployed for customers. There is no sign-up, no trial, no price, and no way in

Verified against the product on 2026-08-24.

A run that pauses waits for a person.

The designed seam is a reviewer, not a retry policy. It is described here as designed, because it is.

Nothing on this list is available today. This is the product we intend to build, written down so you can tell us whether it is the right one, and it will change.

A failed run stops

A flow that fails or pauses does not continue on its own. Continuing is a named reviewer's decision, and aborting is the other half of the same decision.

The handoff is a card, not a log line

The design posts a summary to Microsoft Teams when a run completes or stops, so the person who has to decide is told rather than expected to go looking.

Expiry is on a clock

The buffer clears on a time limit whether or not anyone acts. Nobody has to remember to clean up after an abandoned run.

This seam is designed

The reviewer handoff is specified in the architecture and is not running for anyone. It is on this page as a design commitment, not as behaviour available today.

SimpleForge is infrastructure, not one of the operational applications.

The portfolio has a planned operational suite and a platform tier. This is in the platform half of that picture.

Not part of Simple Ops

Simple Ops is the planned grouping of the operational applications. SimpleForge is not one of them and is not planned to be a module inside it.

Not a tile in SimpleOS

SimpleOS is the planned shell for composing products a customer already has. SimpleForge has no browser-embeddable surface, so it is reached directly rather than through a shell.

Adjacent to Simple Atlas

Simple Atlas is the substrate the operational products call when an agent wants to change something in a customer tenant. SimpleForge is the substrate for the case where the data cannot leave the customer at all.

Governed like everything else

Whatever runs here is subject to the same portfolio rules on approval seams, provenance, and evidence that the rest of the products are.

Where each of these claims comes from.

Run data is held in memory only, overwritten before release, and swept on a timer.
In the build The retention buffer takes a lock around every store and clear, writes over the slot before deleting it rather than relying on the garbage collector, and runs a sweeper started at process boot that clears anything past its time limit. The engine's tests cover it. Checked 2026-08-24
No language model is served, and local inference has no implementation.
In the build There is no local model runtime anywhere in the repository. The product's own integration backlog records on-premises inference as having no code, calls the claim aspirational, and states that no model is served today. Checked 2026-08-24
The MISA designation is a target rather than a fact.
Published standard The two Microsoft security integrations a submission rests on are wired: Sentinel receives the audit stream and Entra managed identity is the credential path. No submission has been made and no designation has been granted. Checked 2026-08-24
The desktop client is at an early stage and has been submitted to no store.
In the build The client is at version 0.1.0 across its manifest, its package file, and its Rust crate, and the product's own baseline record puts it at the earliest of its stages. Neither desktop store has a listing. Checked 2026-08-24
Nothing customer-derived is written outside the run.
In the build There is no database, storage account, cache, or persistent disk for customer content anywhere in the deployment definition, and the audit events the engine emits carry run identifiers, durations, and outcomes rather than payloads. Checked 2026-08-24
Storage posture
Zero retention by construction. Customer content is held in memory for the length of a run and nowhere else.
Certification target
MISA, not held today
Distribution
Microsoft Store (planned) · Apple Mac App Store (planned) · Azure Marketplace (planned, control plane)
Status
Coming Soon

How ICM informs SimpleForge

SimpleForge is designed for buyers who cannot let customer data rest anywhere. The workflow below is specified in its architecture note and is not yet running. It is shown here as the designed boundary, not as behaviour available today.

Designed, not yet built

  1. Context in

    A request carrying tenant and project

    The design admits an execution request identified by tenant and project only. Those two values are what the rest of the run is scoped by.

    Planned

  2. Retained or discarded

    Structure is kept, content is not

    The design retains the identifier of the flow, the identifiers and types of its steps, how long each took, and whether it succeeded. Step parameters and step outputs are discarded. Nothing customer-derived is written to disk, and logs are specified to carry metadata only.

    Planned

  3. What the agent proposes

    The graph runs in memory

    Intermediate results are held in a buffer scoped to the run. The prototype overwrites the slot before releasing it and sweeps anything left behind on a timer, rather than relying on the process ending tidily.

    Planned

  4. Human review a person decides

    A Teams card is the handoff

    On completion the design posts a card summarising the run and linking to the output. When a flow fails or pauses, the card changes state and asks a named reviewer to approve or abort.

    Planned

  5. Permitted effect

    A paused run waits

    A paused or failed flow does not continue on its own. Continuing is the reviewer's decision, and aborting is the other half of the same seam.

    Planned

  6. Evidence recorded

    A summary to Sentinel, and nothing else

    What leaves the run is flow-level metadata: which flow, which client, how long, what status. The design routes that to Microsoft Sentinel as the audit surface, and routes no customer content anywhere.

    Planned

  7. Failure and refusal

    Expiry is on a clock, not on goodwill

    The buffer clears on a time limit whether or not the run ended cleanly, and clearing emits its own audit event. A run that crashed halfway is specified to leave the same nothing behind as one that finished.

    Planned

These labels describe each step of this workflow on its own. They are not the availability of the product, which is the status shown at the top of this page.

This describes how the workflow is built. It is not a statement about availability, which is the status shown at the top of this page. ICM informs how context, review and evidence are designed here. It does not replace this product's database, its tenant isolation, its Entra identity, its audit tables, or its authorisation checks.

Methodology note

ICM was developed by Jake Van Clief and David McDermott. Simple Intelligence Group did not originate the methodology.

Simple Intelligence Group applies relevant ICM principles to context, provenance, traceability, human review, and workflow design.

Participation does not imply certification, endorsement, partnership, or an official ICM Architect designation.

The questions a compliance officer asks first.

Can I buy or trial SimpleForge today?

No. It is not deployed for customers, there is no sign-up or trial, and no price has been set. The only thing you can do here is ask to be told when that changes.

Does SimpleForge run a model on my machine right now?

No, and this page previously implied otherwise. No language model is served yet. Local and disconnected inference is specified in the architecture and has no code behind it. When a model runs under the retention rule, this answer changes and not before.

Then what does the zero-retention claim actually cover?

The workflow engine. Data a run touches is held in memory, overwritten before the slot is released, and swept on a timer if the run is interrupted. Audit records carry run metadata and never run content. That behaviour is built and tested. It is a smaller claim than a retention guarantee over inference, and it is the one the code supports.

Is SimpleForge MISA certified?

No. MISA is the certification target and it is not held. The two Microsoft security integrations that a submission would rest on, Sentinel as the audit sink and Entra managed identity as the credential path, are wired. The designation itself has not been applied for.

Is it on the Microsoft Store or the Mac App Store?

No. The desktop client is an early shell and has never been submitted to either store. Both are listed as planned distribution and neither is a channel you can buy through today.

What happens to my data if I evaluate this later?

The design is that customer content never lands on disk, in a database, or in a cache, and that what leaves a run is metadata. When there is something to evaluate, that will be stated as tested behaviour against a build you can run rather than as a design intention, and this page will say which of the two it is.

There is nothing to start here yet.

SimpleForge is not open to customers and will not be until a model runs under the retention rule this page describes. If the constraint it is built for is one you have, tell us and we will come back to you when there is something real to show.

Availability Get product updates