Autonomous systems · developer infrastructure · research tooling · experimental technology

I build the layer underneath the products: execution engines, control planes, deterministic tooling and research harnesses that stay honest when the model gets expensive, slow, or wrong.

Six systems carry most of the weight right now. Each one has tests, an architecture, and a failure story — that is the bar, not the feature list.

Supporting engineering systems live in the repository list. Dedicated project-site repositories are intentionally kept private; the deployed sites remain public where useful.

flowchart TB
    subgraph Intent
        U[Human or agent intent]
    end

    subgraph Control["Control plane"]
        POS["Portfolio OS<br/>queue · locks · review · publish"]
    end

    subgraph Execute["Execution"]
        AOS["AgentOS<br/>inspect → plan → execute → verify"]
    end

    subgraph Determinism["Deterministic layer"]
        GI["GrokInstall<br/>capability contracts"]
        GM["GrokMax<br/>route · cache · ledger"]
    end

    subgraph Surfaces["Surfaces"]
        GH["gh0st<br/>encrypted local client"]
        OW["OpenCode Watchdog<br/>degeneration guardrail"]
    end

    U --> POS
    POS --> AOS
    AOS --> GM
    GM -. "cheaper than paid" .-> AOS
    GI -. "installs into" .-> AOS
    AOS --> GH
    AOS --> OW
    OW -. "halts degenerate run" .-> AOS

    POS --> PUB[("sanitized public snapshot")]
    PUB --> SITE[noaerth.com]

    classDef core fill:#0A0C10,stroke:#5EE7D0,color:#ECEEF1
    classDef det fill:#0A0C10,stroke:#7C8CFF,color:#ECEEF1
    class POS,AOS core
    class GI,GM det

The rule that holds it together: no component is allowed to report a number it did not measure. Savings are proxies until they are billed. Capability counts come from the ledger, not from a marketing page.

Everything below this line is regenerated from the GitHub API by .github/workflows/sync-profile.yml. Nothing here is hand-maintained, and nothing here is estimated.

Latest releases

• M4G3LL4N0/gh0st 1.0.0-rc.1
• M4G3LL4N0/agentos 0.5.0
• M4G3LL4N0/grokmax 0.2.0-rc.2
• M4G3LL4N0/grokinstall 0.1.1
• M4G3LL4N0/grokbot-office 0.1.0

Pinned systems

• noaerth-portfolio-os — Stdlib-only control plane for a portfolio of autonomous agents: SQLite work q… · 162 tests
• agentos — Provider-neutral AI agent execution and orchestration engine: objective in, v… · 710 tests
• grokinstall — Install the capability, not the complexity. A Go CLI that inspects a repo and… · 486 tests
• grokmax — Deterministic-first LLM routing: zero-cost executors first, five-layer cache,… · 287 tests
• gh0st — Local-first encrypted AI client for xAI/Grok with verifiable no-retention gua… · 27 tests
• opencode-watchdog — Local circuit breaker for runaway OpenCode sessions. Deterministic detection,… · 70 tests

_{Snapshot generated 2026-10-02T19:18:44+00:00 from the GitHub API.}

Public, buildable, and covered by tests.

• GrokInstall — go install a single binary. The deepest documentation in the portfolio and the easiest thing here to actually try.
• OpenCode Watchdog — the most immediately useful thing to a stranger. Solves a problem every agentic-coding user hits.
• GrokMax — the cost-control reference implementation. Routing, caching, and honest telemetry in one pipeline.
• gh0st — desktop client with local-first encryption.
• AgentOS — the engine the rest of the stack runs on.
• Portfolio OS — the control plane that actually manages this portfolio.

Issues, discussions and good-first-issue labels are live on every public repository.

Contributions are welcome on the open-source spearhead first. Read CONTRIBUTING.md — the short version is: branch, keep it reviewable, prove it with a test, open a PR.

Report vulnerabilities privately. See SECURITY.md.

Profile content and documentation: MIT.

<sub>Some projects reference Grok, ChatGPT, OpenAI and other providers as integration targets inside provider-neutral runtimes. Not affiliated with or endorsed by them.

Still curious how you got here? why are you here?</sub>