HEADING OS

The operating system an executive runs their company from — research, communications, CRM, content, and operations. Claude Code is the foundation; HEADING OS is the value built on top: sovereign, security-first, your data kept private.

HEADING OS runs on one guarantee: your data cannot ship with the code. The engine is a shareable, public repository; your data is a private one the engine never contains and never leaks. That separation is enforced by six mechanical layers and a security suite that runs on every commit, not just intended. On top of it sits the workspace an executive actually runs their work from: research, communications, CRM, content, and operations, with Claude Code as the agent.

It is named after its operating philosophy: the Navigation Principle — you set a heading and hold it, correcting course as conditions change, rather than steering toward a fixed point and hoping. The same idea runs through the system: durable state over one-shot prompts, verified completion over hopeful timeouts, operational states over rigid targets. New to the house vocabulary (heading, drift, operational state, engine vs data)? The glossary defines every term in a line.

The two-repository design

Most agent setups keep code and data in one place. HEADING OS splits them on purpose.

flowchart LR
  subgraph ENGINE["ENGINE — .heading-os (this repo)"]
    direction TB
    E1["skills · scripts · rules · hooks · tests"]
    E2["shareable · public"]
  end
  subgraph DATA["DATA — .heading-os-data (private, yours)"]
    direction TB
    D1["crm · knowledge · outputs · threads · context"]
    D2["private forever"]
  end
  ENGINE -->|"get_data_root()"| DATA

• ENGINE (this repo) — skills, scripts, rules, hooks, and tests. No real data, no secrets, no personal information. Shareable, and intended to be public.
• DATA (a separate private repo, yours) — CRM, knowledge, generated outputs, operational threads, and context. The engine resolves it at runtime through a single seam (get_data_root()), as a sibling directory or via the HEADING_OS_DATA environment variable.

You clone the engine; you create your own private data repository (one command); you wire them together. The engine carries the logic, your data stays with you.

What's inside

• Skills — slash-command workflows for research, communications, content, CRM, strategy, and operations, routed from natural language by a single router rule.
• Hooks — PreToolUse / PostToolUse / SessionStart guards that enforce the rules below before a write ever lands.
• Daemons — optional always-on background services (a loopback dashboard, mail/calendar sync) that are driven from the CLI, never required through a browser.
• A security model with teeth — not policy prose alone:
  • Engine ⟂ data separation is proven by six enforcement layers (a bypass guard, a leak guard, a data-path redirect, a build partition, a runtime tree-clean check, and an unbypassable push-time wall in pure code), so the engine clone cannot carry private data regardless of how a file was written — and the data cannot leave on the push, on any path.
  • Outbound send is always human-gated — the lethal-trifecta control. An agent can draft and queue a message; a human clicks before anything leaves.
  • Secrets never reach a remote — a content scan on the sanctioned push path is pure code with no skip flag, behind a bypassable commit-time hook.
  • No "hope-based" waiting — every must-complete step (every push) runs under a progress watchdog that declares a hang only on real inactivity and verifies the postcondition, never trusting a wall-clock timeout or a bare exit code.
• Console-first — every capability is operable from the terminal and from Claude Code chat. The dashboard is a convenience layer, never a dependency.

By the numbers

Every figure here is produced by CI, not asserted by hand. Counts as of v0.7.0, kept honest by scripts/dev/check-readme-numbers.py.

• 231 security tests, run on every commit by the security-tests job over tests/security/. This is the suite that proves the guarantees below.
• 6 enforcement layers hold the engine and data apart: a bypass guard, a leak guard, a data-path redirect hook, a build partition, a runtime tree-clean check, and an unbypassable push-time wall.
• Router accuracy: pending. The nightly router-accuracy trend (F-6.2) needs a week of data before a number is published here.

Quickstart

The full documentation — prerequisites & install, the architecture overview, the security model, daemons & scheduled tasks, the skill/MCP/plugin catalog, the AI model integrations (local Ollama embeddings, the Council models), the service integrations (Exchange email, Telegram, Google, OSINT), workspace personalization, a guide to extending the engine, the memory systems & ODIN, and the data-overlay structure — is published as a browsable site at mishahanin.github.io/heading-os.

The zero-to-running walk-through — WSL2, toolchain, prerequisites, Claude Code, your private data repo, and the engine wired to it — is in docs/DEPLOYMENT.md (with docs/QUICKSTART.md for the short version).

Just want a taste inside your existing Claude Code, with no clone? Install the sovereignty core as a plugin from the marketplace:

/plugin marketplace add mishahanin/heading-os-marketplace
/plugin install heading-core@heading-os-marketplace

The marketplace repo is mishahanin/heading-os-marketplace; docs/PLUGINS.md covers the bundles, how updates work, and when to prefer a plugin over a full clone.

Three focused setup guides cover everything beyond the core install:

• docs/MODELS-SETUP.md — the AI models: installing Ollama for the local bge-m3 embedder behind /recall, and wiring Gemini, Grok, and Kimi as the /council voices.
• docs/INTEGRATIONS-SETUP.md — the services: Exchange email, Telegram, Google contacts, and the OSINT / web-research APIs, with where to get each key and how to verify.
• docs/MAKE-IT-YOURS.md — personalizing a clone: /setup-wizard to generate your voice, business, and personal docs, set your identity, and adapt the house terminology to your own.

The short version, once the prerequisites are in place:

# 1. Clone the engine
git clone https://github.com/mishahanin/heading-os.git .heading-os
cd .heading-os

# 2. Install dependencies (Python 3.11, managed by uv)
uv sync

# 3. Create your own private data repository (one command)
uv run python scripts/create-data-repo.py

# 4. Wire secrets and arm the commit gate
cp .env.example .env        # fill in what you use
pre-commit install

# 5. Verify, then start
uv run python scripts/workspace-health.py
claude       # then /prime

Repository layout

Security

Security is treated as a first-class concern, not an afterthought. The model is walked through end to end in docs/SECURITY-MODEL.md (the lethal-trifecta control, the engine/data layers, the send-gate, the secret gates); the reporting policy and posture summary are in SECURITY.md; the engine ⟂ data guarantee is specified in docs/engine-data-segregation-contract.md.

If you find a vulnerability, please report it privately (see SECURITY.md) rather than opening a public issue.

Status

v0.7.0. The architecture, the security model, and the data seam are stable and load-bearing. Skills and daemons evolve. Interfaces may change between minor versions while the project is pre-1.0. See ROADMAP.md for direction and CHANGELOG.md for what has changed.

Contributing

Issues — bug reports, questions, and ideas — are welcome. Pull requests are accepted by invitation: please open an issue to discuss a change before sending code, so the work fits the direction. See CONTRIBUTING.md and the Code of Conduct.

License

Apache License 2.0 — see LICENSE and NOTICE. You may use, modify, and distribute the engine with attribution; the patent grant and trademark terms are in the license. ODUN.ONE, TrustONE, and the 31 Concept marks are trademarks of 31 Concept, referenced here only as example vocabulary.

Ownership & disclaimer

HEADING OS is a personal project created and maintained by Misha Hanin. Misha is the Founder & CEO of 31 Concept (31C) and uses HEADING OS in his own work and life, but HEADING OS is his personal project. 31 Concept is not involved in HEADING OS, does not maintain, sponsor, or endorse it, and bears no responsibility or liability for it. Any vocabulary that resembles 31C's (ODUN.ONE, TrustONE, DPI+, Tribe, and similar) appears only as illustrative example data and does not imply 31C involvement.

Author

Built by Misha Hanin as a personal project (misha.hanin@odinix.com).