AOSP
AI-Operable Software Protocol

Shadow-execute first. Verify the diff.
Commit to reality.

Scroll

Git for Real‑World
AI Operations

Preview every action

Shadow execution runs each intent in a sandbox. Zero side effects until verified.

Verify every diff

Policy engine assesses risk. Blast radius analysis traces transitive impact.

Audit every trace

Merkle-chain logging. SHA-256 hash linkage. Tamper-evident by design.

Ten steps from intent
to audited reality

Every action flows through a structured pipeline. Each step can halt execution.

Fork reality.

Explore every possible future in parallel sandboxes. Score across six dimensions. Commit only the winner.

Declarative safety.
Quantified risk.

Define policies with a human-readable DSL. Five built-in factors catch dangerous patterns automatically.

Policy DSL
// Block destructive operations
DENY WHEN action MATCHES "delete_*"

// Require approval for high risk
REQUIRE_APPROVAL WHEN risk.score GT 75

// Warn on production surfaces
WARN WHEN surface.name CONTAINS "prod"

// Allow trusted actors
ALLOW WHEN actor.trust == "full"

Financial

pay, charge, refund, transfer

Critical · 81–100

Destructive

delete, remove, drop, destroy

High · 51–80

Secrets Exposure

password, token, key, credential

High · 51–80

Production Target

surface name contains "prod"

High · 51–80

External Side Effects

send, email, notify, publish

Medium · 21–50

Six modules.
One protocol.

Dependency-injected. Adapter-driven. Fully composable.

Start in five lines.

Minimal dependencies. Maximum safety.

import { createAosp } from 'aosp';

const { kernel, policy, temporal, trace } = createAosp();

// Register a target system
kernel.registerSurface({ id: 'my-fs', type: 'filesystem' });

// Execute with full 10-step lifecycle
const result = await kernel.execute({
  action: 'write_file',
  surface: 'my-fs',
  params: { path: 'hello.txt', content: 'Hello, AOSP' }
});
0
Lines of TypeScript
0
Source Files
0
Dependencies
MIT
License