Capability native
Paths, ports, names and identities are not ambient authority. Rights can be delegated narrowly and revoked.
Clean-slate operating system · active research & development
MRML is a capability-native operating system designed around memory-safe software, isolated services and VMs, a semantic interface for sighted and blind users, and AI that can help without becoming root.
principal AgentSession #18
scope BrowserTab #7
ALLOW Semantic.Inspect
ALLOW Semantic.Interact
ASK Media.Describe
DENY Secret.Export
DENY RawHID.Inject
DENY OtherTabs.Inspect
▮
“Power should be explicit, limited, inspectable, and revocable.”
Architecture
Names identify things; they do not authorize them. Applications, services, devices, VMs, accessibility tools and AI operate through explicit, attenuable authority.
Paths, ports, names and identities are not ambient authority. Rights can be delegated narrowly and revoked.
Services are designed to fail, clean up, and restart without turning one compromise into a machine-wide compromise.
Unsafe and legacy software belongs behind explicit virtual-machine boundaries whose devices project only delegated capabilities.
Accessibility is architecture
MRML is designed so visual presentation, Braille, speech, keyboard control and text-only AI can consume the same underlying semantic state rather than reverse-engineering pixels.
⠠⠎⠊⠛⠝⠀⠊⠝“Sign in, button”ACTIVATE node #21A transport-neutral Braille model, routing and input are planned as native semantic interactions, with virtual devices for CI and later physical-device validation.
Secrets are redacted before semantic state reaches GUI, Braille, speech, automation or AI. Accessibility is not a bypass.
Accessibility APIs are not meant to stabilize before practical testing with disabled users and monitor-off workflows.
Security model
Revocation, mapping invalidation, cleanup-gated retirement and bounded restart are core mechanisms—not optional hardening.
An AI can inspect or act only through capabilities actually delegated to its session. Prompt injection does not manufacture OS authority.
The architecture aims for owner-controlled trust roots, scoped attestation and security mechanisms that do not become permanent publisher authority over the machine.
MRML distinguishes goals, implemented mechanisms, tested behavior, measurements, audits and proofs. Virtualized evidence is not called physical-hardware proof.
mrml-rustc
MRML is developing an original, dependency-free Rust compiler aimed at a no_std, no-global-alloc environment and contemporary, forward-looking Rust semantics. Rust 2024 is the floor, not the destination.
The compiler remains incomplete. Ownership/CFG analysis, broad trait solving, full language coverage, linking and package orchestration remain future milestones.
Current progress
The current roadmap emphasizes service isolation, persistent configuration/storage and the virtual-device stack before the full desktop.
Blocking IPC, typed peer death, revocable shared memory, cleanup-gated retirement and bounded restart have equivalent signed KVM/WHP paths.
Capability-native object identities, COW publication/recovery machinery and configuration checkpointing are moving toward authenticated durable storage.
Block-device work is establishing common queue, reset, interrupt and restart semantics intended to extend toward PCI/VirtIO-compatible devices.
The compiler's place/access preanalysis is groundwork for initialization, moves, borrows, loans and Polonius-style flow-sensitive analysis.
Roadmap
Capability lifecycle, IPC, shared memory, cleanup and restart.
Configuration, storage, networking, firewall, block/device transport and QEMU differential testing.
Memory-safe application model, microVM projection, HiDPI graphics and Semantic Scene Graph.
No-JIT browser, structured semantic web, Braille/speech parity, capability-scoped local/remote AI.
Capability provenance, trusted consent, SecretObject use/export separation and visible AI principals.
Rollback-resistant boot, owner trust roots, physical IOMMU evidence, fuzzing, audit and external review.
MRML in four lines
Names identify. Capabilities authorize.
Power should be explicit, limited, inspectable, and revocable.
One compromise should remain one compromise.
Intelligence does not imply authority.
Research system · work in progress
MRML is not production-ready. Its architecture is being implemented and tested incrementally, with unsupported and unproven properties kept explicit.