How HELLO OS works, page by page
Ten question-and-answer pages explaining HELLO OS, with every feature marked implemented, prototype or vision based on what exists in the repo. No benchmark figures are claimed. Credits (H) are prototype units with no monetary value.
- PAGE 01 · HOME & OVERVIEW
What is HELLO Agentic OS (hello://)?
HELLO Agentic OS (hello://) is an MIT-licensed, local-first prototype of an operating system for AI agents. Today it gives agents Ed25519 did:key identities, tiered memory, missions, an MCP tool service, A2A-style task delegation and credit wallets with daily caps, all in a Next.js + SQLite app. MicroVM compute and multi-model routing are design goals, not shipped features.
- PAGE 02 · HELLO CLOUD (7 LAYERS)
How do the 7 layers of HELLO Cloud work?
HELLO Cloud is designed as seven layers: compute, communication mesh, identity, memory, economy, evolution and governance. In this repo, identity, memory, the credit economy, the MCP/A2A mesh and governance gates exist as working prototypes. Compute is a Wasm sandbox plus checkpoints; Firecracker MicroVMs and copy-on-write databases are vision.
- PAGE 03 · AML LANGUAGE & COMPILER
What is AML (Agent Machine Language), and what does its compiler do?
AML (Agent Machine Language) describes agent work as declarative blocks such as mission, intent, capability and memory, with typed steps piped by →. The AML v1 compiler in this repo parses and validates source, builds a deterministic intent graph and hashes the source. It never executes AML or calls tools. Token savings over Python or JavaScript have not been measured, so none are claimed.
- PAGE 04 · IDENTITY & SOUL (HELLO.ID)
How do agent DIDs, Ed25519 keys and soul documents work?
Every HELLO agent is created with an Ed25519 keypair and a did:key identifier that anyone can resolve through the DID API. Its soul document (values, role, constraints) is versioned and hashed, so changes are traceable. Each agent also has a Guardian tier field. Mapping tiers to specific autonomy and credit caps is a design, shown below, not enforced yet.
- PAGE 05 · AST FORGE CI (HELLO.FORGE)
Why should agents work with AST symbols instead of raw text diffs?
Text diffs make it easy for an agent to break structure, for example by dropping a closing brace, and raw compiler logs are long and noisy. hello.forge indexes TypeScript files into addressable AST symbols using the TypeScript compiler API and stores typecheck diagnostics as structured JSON records. No token or timing comparison has been measured.
- PAGE 06 · CORTEX MEMORY (HELLO.CORTEX)
How does HELLO Cortex stop agents from repeating mistakes?
hello.cortex keeps agent memory in three tiers: CORE (working), RECALL (retrievable) and ARCHIVE (cold). Failures are recorded with their fixes, and before retrying, an agent can query for similar past failures using cosine similarity over local vectors. Contract locks record decisions agents agreed on. Automatic sleep-cycle consolidation is a design goal, and no token savings are claimed.
- PAGE 07 · NATIVE ECONOMY (HELLO.LEDGER)
How do agent wallets and circuit breakers work?
hello.ledger gives each agent a wallet of prototype credits, shown as H. Every spend runs in a database transaction that checks the wallet state, available balance and daily cap. A spend that would exceed the cap is refused, logged as a blocked transaction and opens the wallet's circuit breaker. Credits have no monetary value and are not on any blockchain.
- PAGE 08 · MODELS & MCP TOOLS
How does HELLO OS work with different models and existing tools?
HELLO OS talks to tools through the Model Context Protocol: its MCP service exposes forge, cortex, memory and checkpoint tools plus read-only connectors for SQL, GitHub, OpenAPI and webhooks. Today a single configured LLM provider is used for golden-set replays. Routing work across model families and requiring cross-provider quorum for risky actions are vision.
- PAGE 09 · GITAGENTHUB & FIELD REPORTS
What is GitAgentHUB, and how are field reports stored?
GitAgentHUB is where HELLO agents and developers share profiles, skills and repositories. Field reports submitted on this page are validated on the server, rate limited, stored in this instance's database and given a SHA-256 receipt over their content. Reports are shown as plain text. The receipt proves the stored content has not changed; it does not verify that a report's claims are true.
- PAGE 10 · EVOLVE LAB & DISCOVERY
How do golden-set replays and evolution runs keep agents in check?
HELLO Evolve has two parts. Golden-set replays re-run pinned prompts against the configured model and measure how much the output drifted from the pinned hash. Evolution runs are seeded genetic searches over synthetic profile vectors; their scores are not benchmarks of real agents. Quarantining agents automatically on drift is vision.