# Installing the intrikata stack locally megapraxis (session bootstrap) + metamegapraxis (meta-auditor) as Claude Code skills. Live reference and current audit state: https://docs.intrikata.com **Works without any backend.** megapraxis is designed to degrade gracefully: with no graph backend it runs in filesystem-fallback mode (deterministic disk scans, same briefing contract). Graph-backed features (coverage grids, swarms, bridges) light up only if a suitable backend answers on `localhost:8080` — none is included, none is required. To initialize one, see `BACKEND.md` in this package (also served at https://docs.intrikata.com/backend.md): it specifies the exact HTTP contract and a tiered bootstrap path, from a ~100-line stub to a full knowledge-graph server. ## OpenAI Codex: install once for every future project The native plugin path is the shortest durable setup. Run once on the machine: ```text codex plugin marketplace add https://docs.intrikata.com/git/intrikata-stack.git codex plugin add praxis@intrikata-stack ``` Close and reopen the Codex app, start a new task in any project, and invoke `$praxis:megapraxis`. The restart is a harness-registration boundary, not an onboarding step you repeat per project. Verify the skill appears in the new task's skill registry; installation on disk alone is not proof that the running app loaded it. Direct-skill fallback after unpacking the zip or cloning the served repo: ```text py install.py --harness codex # Windows python3 install.py --harness codex # macOS / Linux ``` This installs into `~/.codex/skills` for all future projects. Start a new task afterward. The agent playbook and machine-readable saturation certificate are at https://docs.intrikata.com/onboard-codex.md and https://docs.intrikata.com/demo/codex-onboard.json. The Codex plugin contains only two instruction skills: no MCP server, hooks, app connector, authentication, or project-file writer. Remove it with: ```text codex plugin remove praxis@intrikata-stack codex plugin marketplace remove intrikata-stack ``` ## If you are Claude Code (agent install) Skills hot-reload — no app restart needed. Either: 1. Download and unpack: ``` curl -sL https://docs.intrikata.com/pkg/intrikata-stack-skills.zip -o intrikata-stack-skills.zip ``` Unzip, then copy `praxis/skills/megapraxis/` and `praxis/skills/metamegapraxis/` into `~/.claude/skills/` (each directory keeps its `SKILL.md`). 2. Or clone the repo the site itself serves: ``` git clone https://docs.intrikata.com/git/intrikata-stack.git ``` and copy the same two skill directories. Then invoke `/megapraxis`. Verify the install the harness-truth way: the skill names appear in your in-context skill listing on the next turn — presence on disk alone is a hypothesis. ## If you are human 1. Download and unpack https://docs.intrikata.com/pkg/intrikata-stack-skills.zip: - macOS / Linux: `curl -sL https://docs.intrikata.com/pkg/intrikata-stack-skills.zip -o praxis.zip && unzip praxis.zip -d intrikata-stack` - Windows (PowerShell): `iwr https://docs.intrikata.com/pkg/intrikata-stack-skills.zip -OutFile praxis.zip; Expand-Archive praxis.zip intrikata-stack` 2. From the unpacked folder run the installer (any Python 3.8+): - macOS / Linux: `python3 install.py` - Windows: `py install.py` (prefer the `py` launcher — bare `python` may hit the Microsoft Store stub) It copies the skills into `~/.claude/skills/`. Flags: `--list` to preview, `--force` to overwrite, `--target DIR` to install elsewhere. 3. In Claude Code, type `/megapraxis`. ## As a Claude Code plugin (marketplace) The cloneable repo doubles as a plugin marketplace: ``` /plugin marketplace add https://docs.intrikata.com/git/intrikata-stack.git /plugin install praxis@intrikata-stack ``` Note: plugin installs register at Claude Code **app** startup — close and relaunch the app after installing (a new session is not enough). The plain skill copy above needs no restart. As a plugin, the skills invoke as `praxis:megapraxis` / `praxis:metamegapraxis`. ## If you are Grok Build (xAI) — short-circuit onboard **Do not cold-rediscover the stack.** Follow the live short-circuit playbook first: - https://docs.intrikata.com/onboard-grok.md - Machine-readable certificate: https://docs.intrikata.com/demo/grok-onboard.json One-time install into Grok's skill root (filesystem-fallback is the designed mode; no MCP backend required for a valid briefing): ``` # after unpacking the zip or cloning the bare repo: py install.py --target "%USERPROFILE%\.grok\skills" # Windows python3 install.py --target ~/.grok/skills # macOS / Linux ``` Then say `/megapraxis` or "orient me". Adapter text: `adapters/GROK.md`. Skip auto-swarm and do not treat missing intrikata-topology MCP as a first-session blocker (traps #4/#5 are N/A for pure Grok onboard unless you asked for graph features). ## Other agentic code builders (opencode, Copilot, Cursor, Gemini CLI, …) The skills are procedures, not plugins — any agent that reads files and runs shell commands can execute them. `HARNESS.md` (in this package, also at https://docs.intrikata.com/harness.md) maps every Claude-Code-specific primitive to its generic equivalent. Drop-in pointer snippets live under `adapters/`: - `adapters/AGENTS.md` — paste into your project or global `AGENTS.md` (Codex CLI, opencode, GitHub Copilot, and anything else in the AGENTS.md family) - `adapters/CODEX.md` — Codex-specific short-circuit contract and harness mapping - `adapters/cursor/praxis.mdc` — copy to `.cursor/rules/praxis.mdc` in your workspace - `adapters/GEMINI.md` — paste into your project or global `GEMINI.md` (Gemini CLI) - `adapters/GROK.md` — Grok Build short-circuit onboard (see also `/onboard-grok.md`) Each adapter assumes the package's `praxis/skills/` + `HARNESS.md` are readable in your workspace (e.g. keep the unpacked package at `./intrikata-stack/` and adjust the paths in the snippet). Nothing edits your config files automatically — placement is yours. ## Cross-agent / cross-OS local realize (white-paper architecture) To materialize the full **local** white-paper stack (skills on every harness root, optional Tier-1 HTTP stub) in one shot, use the self-realizing path: - Playbook: https://docs.intrikata.com/onboard-local.md - Skill: `praxis-local-realize` (included in this package) - Scripts: `scripts/realize.py`, `scripts/backend_stub.py` (also at `/scripts/` on the site) ```text # all harnesses, Tier 0 (valid filesystem briefing) py -3 scripts/realize.py --harness all --tier 0 --force # Windows python3 scripts/realize.py --harness all --tier 0 --force # macOS / Linux # classic multi-target install py -3 install.py --harness all --force ``` Do not auto-dispatch a metamegapraxis self-audit swarm as part of realize. ## What's inside - `praxis/skills/megapraxis/SKILL.md` — deterministic session-start briefing: 7 modes, disk-verified claims, dark-plugin detection, 30-trap operational knowledge. - `praxis/skills/metamegapraxis/SKILL.md` — the meta-auditor: trap roll-call, bridge-manifest integrity, Tarski-honest verdicts. - `praxis/skills/praxis-local-realize/` — self-realizing cross-OS / cross-agent local deploy. - `scripts/realize.py` · `scripts/backend_stub.py` — local realize + Tier-1 stub. - `BACKEND.md` — the optional backend's contract + bootstrap procedure (``). - `HARNESS.md` — primitive map for executing the skills outside Claude Code. - `adapters/` — drop-in pointer snippets: Codex, AGENTS.md family, Cursor, Gemini, Grok Build. - `ONBOARD-CODEX.md` — install-once short-circuit for future Codex projects. - `ONBOARD-GROK.md` — short-circuit onboard playbook for future Grok projects. - `ONBOARD-LOCAL.md` — cross-agent local white-paper realize playbook. - `install.py` — copies the skills into place (`--harness all` / `--target` for multi-root). - `MANIFEST.json` — version, build date, sha256 of every file. Both skills carry a portability preamble: machine-specific paths and usernames from the origin system were generalized at build time; every backend reference points at the generic `` root described in `BACKEND.md` — an optional integration, not a prerequisite.