# Ecosystem NurCLI ships with an auto-provisioned knowledge stack. The [v0.41.0 refresh](stack-update-2026-09-30.md) records upstream skill commits and resource hashes. Complete managed trees refresh together; local additions and Nur adaptations remain available. Explicit `ecosystem ensure --force` updates installed companions as well as missing ones. Headroom uses its isolated uv interpreter when installed as a tool, so inline compression works without installing the library into global Python. `nur ecosystem refresh-skills` refreshes only global metadata and confirms that a complete snapshot was saved. It does not resolve credentials or install packages. ## Components | Component | Role | |-----------|------| | **sgrep** | [Context.dev sgrep](https://github.com/context-dot-dev/sgrep) (MIT) local semantic code search behind the `sgrep` tool - static embeddings + BM25 over ripgrep's file list, exact identifiers first, piped rerank of a focused `rg` run. Built with `cargo install --git https://github.com/context-dot-dev/sgrep --locked` when cargo is present (retried at most weekly), then the 32 MB model is warmed once. `NUR_SGREP_BIN` pins or (empty) disables it. Docs: [tools.md](./tools.md#sgrep) | | **ast-grep** | [ast-grep](https://ast-grep.github.io) (MIT) structural search and rewrite on tree-sitter, behind the `ast_grep` tool. Installed from the pinned npm package `@ast-grep/cli` (prebuilt binaries), `cargo install ast-grep --locked` without npm, retried at most weekly. `NUR_AST_GREP_BIN` pins or (empty) disables it. Docs: [tools.md](./tools.md#ast_grep) | | **Graphify** | Code knowledge graph (`graphify-out/`) - query / path / explain | | **PLUR** | Shared engram memory across tools and sessions | | **Ruflo** | Vector memory + swarm / hive-mind patterns | | **Executor** | MCP / OpenAPI gateway catalog | | **GraphJin** | GraphJin - governed GraphQL→SQL over live databases (`graphjin` tool · `/graphjin`). npm-installed when Node is present, with a SQLite demo config seeded under `~/.nur/graphjin/` so `graphjin serve --demo --path ` works out of the box; point `GRAPHJIN_CONFIG_PATH` at a real database to go live | | **dogwood** | [Dogwood](https://github.com/dogwood-policy/dogwood) runtime verification for AI agents - a Cedar-extending policy language with temporal logic (since / formerly / once / count_within) over the tool-call event stream (`dogwood` tool). Auto-installed via cargo when available (`cargo install --git https://github.com/dogwood-policy/dogwood amzn-dogwood-cli`); default policy pack under `~/.nur/dogwood/`. The reference interpreter is an on-demand evaluation/guardrail, not a runtime gate | | **browser** | [agent-browser-cli](https://github.com/sleepinginsummer/agent-browser-cli) real **default browser** bridge (Arc / Chrome / Edge / Brave / …) - perception + control via the `browser` tool; `nur browser setup` stages the extension once | | **terminal-browser** | [terminal-browser.com](https://terminal-browser.com/) in-terminal Chromium (`terminal_browser` tool · `/tb`). Native/WSL when the upstream binary is present; on Windows Nur falls back to agent-browser-cli so open/snapshot/click still work | | **Cua** | [trycua/cua](https://github.com/trycua/cua) computer-use driver (`cua-driver`) - full-desktop automation via MCP + CLI. Auto-installed on single-shot install **without** the elevated autostart daemon (`-NoAutoStart`). Toggle the always-on background daemon in-app with **`/cua on`** / **`/cua off`** (`/cua status` to check); or use it on demand with `cua-driver serve` / wire its MCP with `cua-driver mcp-config` | | **Skills** | Progressive packs (design-eng, clone-website, cybersecurity, sc-research, …) via `skill` | | **Plugins** | In-product marketplace (`/plugins` · `nur plugins`) - install Superpowers, Vercel, Firecrawl, Chrome DevTools, Figma, Sentry, Fable, … into `~/.nur/plugins` | | **Takeover** | `/takeover` (alias `/hijack`) imports **Claude Code · Codex · Cursor · Nur · Grok Build** sessions, powered by the bundled `resume-session` reader | | **AKM** | Agent knowledge package manager (requires Node.js) | | **t3code** | [t3code](https://github.com/pingdotgg/t3code) driver compatibility layer - env-isolated vendor-CLI auth probing, atomic `auth.json` writes, delegate probe, and pairing tokens (`t3code` tool). Probe-only: it stores no secrets and adds no provider | | **Headroom** | [headroom-ai](https://github.com/headroomlabs-ai/headroom) inline tool-result compression (default **on**; disable with `[headroom] enabled = false`) | | **OptMem** | [OptMem](https://github.com/VictorTaelin/OptMem) permanent memory under upstream `~/.optmem` (`/optmem` · `/memo`). `nap lines=[...]` drains up to 24 blocks in one call and never forwards upstream's next imperative | | **Connectome** | Agent-native memory + continuity (arXiv:2606.24775 · [animalabs.ai/connectome](https://animalabs.ai/connectome)): hierarchical memories (recent / l1 / l2 / l3), an append-only chronicle, and checkpoints. `connectome` tool; injected into the system prompt while `native_memory = true` (default) | | **mem (HelixDB)** | Harmonized memory read/write across vector, knowledge graph, and hierarchical stores (`mem` tool). Optional HelixDB graph-vector resident via `[helix_memory]` (`HELIX_URL` / `NUR_HELIX_URL`, `HELIX_API_KEY`); local native memory stays authoritative with a durable outbox for failed mirrors | | **TypeSafe (Jev)** | [TypeSafe](https://docs.typesafe.ai) System One typed judgments wired through the harness (`typesafe` tool · `/typesafe` · `/jev` · `[typesafe]`). Provider-agnostic: one key lifts every provider and is never the active model. The pinned `/login` entry is a sidecar (bordered, first) so catalog counts and chat routing are untouched. Docs: [typesafe.md](./typesafe.md) | | **Local Jev engines** | Keyless System One backends served by one bridge (`src/jev_local.rs` + `scripts/jev_local_bridge.py`, embedded in the binary): openJev-verdict-2.0 (CPU), Bespoke-Nimble-9B (NVIDIA GPU / Apple Silicon MLX), Laya Core ML (Apple Silicon ANE, macOS 15+), plus a deterministic mock. `nur jev start` · `use` · `status` · `stop` · `selftest` · `probe`; engine flags are forwarded verbatim (`--verdict-model`, `--nimble-dir`, `--laya-model`, `--device`). Per-device refusals are actionable. Docs: [jev-local.md](./jev-local.md) | | **egaki** | [egaki](https://github.com/remorses/egaki) image/video/speech (`/egaki`; ChatGPT / xAI OAuth / BYOK / Egaki plan - bare `/image` attaches a file for vision instead) | | **tldraw** | Offline [tldraw](https://www.tldraw.com) boards (`tldraw` tool · `/draw ` or `/draw `; `/draw excalidraw` and `/draw pen` route to those tools) | | **excalidraw** | Hand-drawn, publishable diagrams via [excalidraw-cli](https://github.com/ahmadawais/excalidraw-cli) (`excalidraw` tool · `/excalidraw `, also reachable through `/draw excalidraw`) | | **penecho** | penecho canvas beyond the chat box - ink, MathJax, plots, animations (npm `penecho` sidecar · `penecho` tool · `/pen` · `/drawings` · `/penecho`) | | **lightpanda** | [Lightpanda](https://github.com/lightpanda-io/browser) headless browser built for AI agents, detect-only (AGPL, never installed or embedded by nur-cli): one-shot fetch, CDP server, or MCP server for JS-heavy pages and clean-room automation. 1.0.3 (Nov 2026) ends the beta: WebMCP, AudioContext, Payment Request API, new CDP/MCP tools, Grok/xAI and Parallel providers; CORS is enforced by default now. Not a replacement for `browser`, which keeps the user's logged-in Chrome session; no native Windows binary (run inside WSL); telemetry on by default (`LIGHTPANDA_DISABLE_TELEMETRY=true` to disable). Skill: `lightpanda` | | **desloppify** | [Desloppify](https://github.com/peteromallet/desloppify) codebase health scanner with a persistent, prioritized cleanup queue, detect-only (never installed by nur-cli): `desloppify scan`, `plan triage`, `next`, rescan, state in `.desloppify/`. Use for the queue and for signals `slop_scan` lacks (dead code, complexity, naming, coverage gaps, Rust smells); do not replace `slop_scan` with it. Skill: `desloppify` | | **akarso** | akarso post, schedule, and reply across 14 social platforms (npm `akarso` · `akarso` tool · `/akarso` · `akarso` skill). Publishing, deleting, and connecting accounts are outward-facing and gated in manual mode | | **fractal** | [fractal](https://github.com/plasma-ai/fractal) hierarchical loops - Unix only; ecosystem ensure + loud warn on `node start` | | **infinite-headcount** | Factory skill pack auto-provisioned via ecosystem packs | | **sc-research** | First-party whitehat DeFi pack in-repo at `skills/security/SCA/` (`/sc-research` + protocol/tool playbooks + `hunting-x-linked-bounties` + historical vuln library + investigator-ops). DarkNavy `web3-skills` and Cyfrin `solskill` auto-provisioned. Lawful/in-scope only | | **CAD Skills** | [Text-to-CAD](https://www.texttocad.dev/) STEP-first CAD, DXF, G-code, robot-description, part sourcing, validation, and local review skills (`/cad`, `/cad-viewer`, `/dxf`, `/urdf`, `/sdf`, …) | | **mobile-harness** | [DroidRun Mobile Harness](https://github.com/droidrun/mobile-harness) portable Android, iOS, and cloud-phone operating guides for `mobilerun-core` (`/mobile-harness`) | | **Interior** | [ddoemonn/interior](https://github.com/ddoemonn/interior) / [interior.dev](https://interior.dev) - finished React micro-interactions (bundled `/interior` skill + `/plugins install interior` for the component source tree) | --- ## Plugin marketplace Browse and install skill packs from the TUI with the **same picker UX as `/login`** (filter, ↑↓/wheel, ↵, click). ```text /plugins # open marketplace picker /plugins list /plugins install superpowers /plugins enable|disable /plugins uninstall ``` CLI (same registry): ```bash nur plugins list nur plugins install vercel nur plugins install firecrawl nur plugins disable superpowers ``` | On disk | Role | |---------|------| | `~/.nur/plugins//` | git clone of the plugin | | `~/.nur/plugins/registry.json` | installed + enabled flags | | `~/.nur/skills/` | skill packs mirrored on install so discovery always finds them | Enabled plugins are scanned on each agent turn. Catalog includes Superpowers, Vercel, Chrome DevTools, Firecrawl, Figma, Sentry, Cloudflare, MongoDB, Axiom, Railway, Fable. On install, skill packs are **mirrored in full** (including `references/`) into `~/.nur/skills//` so discovery paths stay complete. --- **Default auto-install** (on `nur install` / ecosystem ensure): `superpowers`, `fable`, `mattpocock`, `addyosmani`, `builderio`. Also auto-provisioned via the skills CLI into `~/.agents/skills`: design (Emil), clone-website, cybersecurity (large), infinite-headcount, Text-to-CAD, mobile-harness, DarkNavy web3-skills, and Cyfrin solskill, plus dual-write of the default set when the CLI is available. First-party `/sc-research` (whitehat DeFi playbooks) ships bundled. Schema 27 re-pulls those packs and the default plugins so upstream skill edits land on the next ensure. ### `/plugins` catalog (browse by category) Type a category in the picker filter (`design`, `finance`, `trading`, `workflow`, …). Rows sort by category then name. | Category | What lives there | |----------|------------------| | `workflow` | Superpowers, Fable, Builder.io, Compound Engineering, gstack, agent toolkits | | `engineering` | Matt Pocock, Addy Osmani, Anthropic official, Microsoft, curated AI skills | | `specs` | Spec Kit, Spec-Driven Development, AI Dev Tasks | | `engineering` (cont.) | Vercel Agent Skills, nanocodex | | `design` | Impeccable, Meng To, UI Craft, Design Skills, Taste Skill, Oh My Design, Superdesign, Anydesign, UI/UX Pro Max, Awesome Design Skills, **Interior** (ddoemonn/interior.dev micro-interactions), **recent.design set**: Emil Kowalski, ibelick UI Skills, Make-Interfaces-Feel-Better, OKLCH, UserInterface Wiki, Dialkit | | `design-system` | Extract Design System, Brand→Design.md, Feature-Sliced Design | | `browser` | Agent Browser (vercel-labs), Chrome DevTools, Firecrawl, Playwright Skill, Figma | | `deploy` | Vercel, Railway | | `cloud` | Cloudflare, Google Skills, NVIDIA Skills | | `data` | MongoDB, LangExtract | | `observability` | Axiom, Sentry | | `marketing` | AI Marketing Skills, coreyhaines marketingskills, OpenSEO, Akarso | | `finance` | Finance Skills, Longbridge, Buffett, CRE, CC Finance | | `trading` | Alpaca Skills, Vibe Trade | | `security` | Claude Red (authorized red-team) | | `science` | Scientific Agent Skills, Awesome Journal Skills | | `robotics` | Robotics Agent Skills | | `catalog` | Mega packs / awesome indexes (install selectively) | Install: `/plugins` or `nur plugins install `. Nested `SKILL.md` layouts are discovered recursively and mirrored into `~/.nur/skills//` so `/skill-name` works. Each mirrored folder carries a `.nur-provenance.json` stamp naming its plugin and commit: a plugin never replaces your own skill or another plugin's (the install message says which it kept), disabling or uninstalling a plugin removes its copies, and enabling puts them back. `/plugins` and `nur plugins` show the commit each install is at, and a disabled plugin stays disabled when ecosystem ensure updates it. ## Skills market Third-party skills install from a signed index, each pinned to its author's repository at a fixed commit: - **Index.** `nur-market-index.json` is built for every release by `scripts/build_market_index.py` from the sources in `skills/market-sources.json` (aidentity's public-domain registry, plus any repositories listed there) and signed with the release key. nur downloads the index for its own version, verifies the signature, and caches it for a day in `~/.nur/cache/market-index.json`; it never queries a registry at runtime. `NUR_SKILL_MARKET_INDEX` points at a local index instead. - **Install** (`nur skills add ` or `/skills add `) fetches each file from `raw.githubusercontent.com///...`, checks its SHA-256 against the index, scans the skill (Unicode tag characters, bidi controls and zero-width characters block it; instruction-override phrasing, pipe-to-shell installs, credential reads next to a network send, base64 blobs, wildcard `allowed-tools` and executables are listed for review), then moves it into `~/.nur/skills/` with a `.nur-market.json` stamp. A name taken by a skill the market did not install is refused unless you pass `--as` or `--replace`. Entries without a distributable license are link only. - **Only you install.** The model's `skill` tool can search the market and tells you the command, but it cannot install: instructions that ride in the next session have to be your choice, or a web page could talk the agent into loading one. `skill read` of a market skill starts with where it came from. - `nur skills remove` deletes market installs only, never a skill you wrote; `nur skills verify` re-hashes them against what was installed. ## Slash skill invocation Every installed skill is addressable as a slash command - no per-skill hardcoding required. | Form | Behavior | |------|----------| | `/skill-name` | Toggle **sticky** mode for this session (skill body rides every turn) | | `/skill-name on` / `off` | Force sticky on or off | | `/skill-name ` | **One-shot** turn: activate the skill and run `` immediately | Examples: ```text /adhd /fable-method fix the flaky auth test /site-cli build a CLI for this restaurant site from capture.har /design-eng polish the settings modal ``` The slash palette lists built-in commands plus matching installed skills as you type (e.g. `/fab` → `/fable-method`). Up to 40 skill hits are shown so a large install (800+) does not drown the list - type more characters to narrow. Sticky skills appear on `/status` and `/skills`. Clear with `/skill-name off`. ## Natural-language skill activation **Opt-in.** Set `[skills] auto_activate = true` in `~/.nur/config.toml` first; off by default, explicit `/skill-name` and the `skill` tool always work. When enabled and your wording matches a high-signal workflow skill that is installed, Nur **auto-injects** that skill's full body into the system prompt for the turn and shows a status chip: ```text fable-method · activated from your wording (no slash command needed) ``` | Say something like… | Activates (if installed) | |---------------------|---------------------------| | *think like fable*, *how fable would*, *fable method*, *do it like fable* | `fable-method` | | *fable loop*, *run the fable loop* | `fable-loop` | | *fable judge*, *verify like fable* | `fable-judge` | | *debug systematically*, *find the root cause* | `systematic-debugging` | | *TDD this*, *tests first*, *red green refactor* | `test-driven-development` | | *let’s brainstorm*, *brainstorm this* | `brainstorming` | | *write a plan*, *plan first then…* | `writing-plans` | | *execute the plan*, *implement the plan* | `executing-plans` | | *verify before claiming done*, *check your work thoroughly* | `verification-before-completion` | | *code review this*, *review my changes* | `requesting-code-review` | | *polish the UI*, *emil style*, *design eng* | `design-eng` | | *skeuomorphic component*, *neumorphic dark*, *rotary dial ui*, `/skeuo` | `skeuomorphic-ui` | | *cache hit rate*, *cache busting*, *route through my proxy*, *own gateway* | `gateway-cache-awareness` | | *post to X/LinkedIn*, *schedule a post*, *cross-post*, `/akarso` | `akarso` (native tool) | | *keyword research*, *backlinks*, *rank tracking*, *site audit*, `/openseo` | `openseo` (MCP) | | *dial in the …*, *tune parameters live*, *tuning panel*, `/dialkit` | `dialkit` | | *clone this website*, *pixel-perfect clone* | `clone-website-meta` | | *draw a diagram*, *excalidraw* | `excalidraw` | | *resume from Claude / Grok / Codex / Cursor / Nur* | matching `resume-*` skill | | *toolcraft*, *design app scaffold*, *craft tooling* | `toolcraft` (pointer → live docs) | | *site cli*, *HAR file*, *watch network requests*, *derive a client* | `site-cli` | | *adhd mode*, *i have adhd*, `/adhd` | `adhd` (also sticky slash) | | *write a tech spec*, *architecture handoff* | `tech-spec` | | *make a skill for…*, *fable domain* | `fable-domain` | | *next.js app router*, `/nextjs` | `nextjs` | | *shadcn/ui*, `/shadcn` | `shadcn` | | *vercel ai sdk*, `/ai-sdk` | `ai-sdk` | Only **installed** skills fire (marketplace plugin or skill pack). Unrelated chat does not activate anything. The injected skill is **mandatory for that turn** - the model must follow it, not freestyle a shorter path. ### Skill trust and invocation flags Skills that live **inside a project** (`./.nur/skills`, `./.claude/skills`, `./.agents/skills`) are inert until you trust them, because a cloned repo must not inject instructions into your agent on first open. `/skills trust` records a digest of every project `SKILL.md` in `~/.nur/trusted-skills.json`; editing any project skill holds it again until you re-trust. While a project has untrusted skills the TUI notes it once per session. Global skills (`~/.nur/skills`, `~/.agents/skills`, …) are already yours and need no trust. Two frontmatter flags shape who can invoke a skill: | Flag | Effect | |------|--------| | `disable-model-invocation: true` | User-only: never NL-activated and `skill` tool `read` refuses it; `/skill-name` still loads it | | `user-invocable: false` | Engine-only: hidden from slash resolution and the `/skills` palette | ### Context discipline (why 1,000+ skills do not blow the prompt) Skills are **never** bulk-catalogued into the system prompt (that burned tokens and Claude/Anthropic usage on large installs). Activation is on-demand only: | Layer | What is injected | When | |-------|------------------|------| | **Default** | nothing skill-related | every normal turn | | **NL activation** | full `SKILL.md` body for **one** matched skill | built-in intent phrases, skill **name** in the query, or unique **description** intent | | **Slash one-shot** | full body + user prompt for **one** turn | `/skill-name ` | | **Slash sticky** | full body every turn until off | `/skill-name` / `on` / `off` | | **On demand** | `skill(action=list\|read, name=…)` | model loads more when needed | | **`poor_mode`** | skip PLUR + long memory only | cost saver - **skills still activate** | Works the same on every provider (Claude, Grok, OpenAI, …). Accidental discovery still works: saying a skill’s name (`grill-me`, `superpowers`) or a distinctive description phrase can activate that installed pack without a slash command. Pointer skills (e.g. **toolcraft**) stay short on purpose: they route to external docs instead of embedding a long guide. Prefer specific activation phrases so casual chat does not false-fire. Install Fable (and others) with: ```bash nur plugins install fable nur plugins install superpowers ``` **First-party skill pack:** the skills NurCLI ships itself - `plur`, `ruflo`, `graphify`, `excalidraw`, `startup-skill`, `how-to-illustrate`, `typesafe-ai`, the `resume-session` engine, the `bro`, `coding-standards`, `prelude`, `tech-spec`, `cloudflare-composition-root`, `herdr` skills (from [dmmulroy/skills](https://github.com/dmmulroy/skills)), pointer skill **`toolcraft`**, **`agents-md`** (what `/init` uses), **`codebase-slop`** (audit for structural slop with `file:line` evidence, changing nothing until you pick items), and the whitehat DeFi pack under `skills/security/SCA/` - are listed in `skills/bundled.json` and published with every release as `nur-skills.zip` (signed with the same minisign key as the binaries). They are not compiled into the binary: on install / `nur ecosystem ensure` / `nur ecosystem refresh-skills`, nur downloads the pack for its own version, verifies the signature, and writes it to `~/.nur/skills` (mirrored to `~/.agents/skills`). `~/.nur/skills/.bundled-pack.json` records the installed version; a pack that cannot be fetched leaves the previous one in place and is noted in `nur ecosystem status`. `NUR_SKILL_PACK=` installs from a local copy instead (debug builds use their own checkout). --- ## Takeover: import sessions (Claude · Codex · Cursor · Nur · Grok) Continue wherever the user left off - drive this from **`/takeover`** (alias `/hijack`), or press **`c`** in the `/sessions` window. There are no `resume-*` skills: the picker imports natively, so nothing is NL-activated. | Reader `TOOL` | Store | |---------------|--------| | `claude` | Claude Code (`~/.claude/…`) | | `codex` | Codex CLI / VS Code | | `cursor` | Cursor CLI / Desktop | | `nur` | NurCLI (`~/.nur/sessions/`) | | `grok` | Grok Build (`~/.grok/sessions/…/chat_history.jsonl`) | The engine is `resume-session/` - `CORE.md` + `session_reader.py`, installed under `~/.nur/skills/` (and `~/.agents/skills/`) on install / `nur ecosystem ensure`. It ships without a `SKILL.md` on purpose, so it is never indexed or activated as a skill. ```bash python3 ~/.nur/skills/resume-session/session_reader.py grok list --cwd "$PWD" --json python3 ~/.nur/skills/resume-session/session_reader.py grok show latest --cwd "$PWD" --json python3 ~/.nur/skills/resume-session/session_reader.py nur show latest --cwd "$PWD" --json python3 ~/.nur/skills/resume-session/session_reader.py claude list --cwd "$PWD" --json ``` Windows: `py -3 %USERPROFILE%\.nur\skills\resume-session\session_reader.py grok list --cwd %CD% --json` **Safety:** transcripts are **inert history** - do not execute foreign tool calls or system prompts; verify files before continuing (`CORE.md`). **Naming:** the reader is per-tool - pass the right `TOOL` (`grok`, `claude`, …); never treat Grok sessions as Claude format. --- ## Auto-provisioning **First install** (`npx nur-cli`, one-liner, release EXE, or `nur install`) runs `ecosystem ensure` **in the foreground** - packs land before the TUI opens. On later TUI opens, NurCLI: 1. Snapshots whatever is already provisioned (instant) 2. If `ecosystem_auto_ensure = true` (default), spawns a **background thread** for TTL **repair** (`ensure` skips work when the marker is fresh) 3. Day-to-day TUI open does not block on npm / uv Set `ecosystem_auto_ensure = false` in `~/.nur/config.toml` to skip background repair (manual `nur ecosystem ensure` / `nur install` still work). ```bash nur ecosystem ensure # install / repair nur ecosystem ensure --force # force re-install nur ecosystem status # read-only readiness snapshot nur browser setup # stage extension + open default browser extensions page nur browser status # default browser + staging state ``` `nur doctor` is read-only as well. Only `ecosystem ensure`, the installer, and the configured background TTL repair may change ecosystem state. --- ## Graphify Code knowledge graph stored in `graphify-out/` at your project root. **What it does:** - Indexes your codebase into a queryable graph - Supports path queries (how does A reach B) - Supports explain queries (what does this function do) - Integrates with the `graphify` tool in the TUI **Requires:** uv (or Python 3.10+) --- ## PLUR Shared engram memory for AI agents. **What it does:** - Persists preferences, corrections, conventions across sessions - Shared across tools (NurCLI, Claude Code, Cursor, etc.) - Searchable from the TUI via `/plur` **Requires:** Node.js 20+ --- ## Ruflo Vector memory + swarm coordination patterns. **What it does:** - Semantic memory search across sessions - Swarm / hive-mind agent coordination - Hooks for routing tasks between agents - Searchable from the TUI via `/ruflo` **Requires:** Node.js 20+ --- ## Executor MCP / OpenAPI gateway catalog. **What it does:** - Provides a unified catalog for API integrations - Routes tool calls to the right MCP server or OpenAPI endpoint - Policy enforcement for multi-agent tool routing **Requires:** Node.js 20+ **Executor 2 (current):** the CLI is at `2.0.0-beta.37` via `npm i -g executor@beta` (stable 2.0.0 not yet on npm `latest`). V2 serves one streamable-HTTP MCP endpoint at `http://127.0.0.1:4312/mcp` with six tools (`skills`, `search`, `describe`, `call`, `execute`, `resume`); the V1 CLI commands (`tools`, `call`, `mcp`, `web`, `install`) no longer exist. Data lives in `~/.executor/v2/cli`. There is no OS autostart: run `executor service start` again after a reboot. `executor upgrade` follows the release channel. --- ## Skills Progressive skill packs loaded on demand. The MIT-licensed [startup-skill](https://github.com/ferdinandobons/startup-skill) pack adds `/startup-design`, `/startup-competitors`, `/startup-positioning`, and `/startup-pitch`. Natural-language activation includes "validate my startup", "competitor battlecards", "position my startup", and "pitch my startup". These are on-demand workflows, not background services. **Built-in skills include:** - `design-eng` - UI polish and animation - `skeuomorphic-ui` - Dark skeuomorphic components (knobs, sliders, inset controls) with top-lighting, layered shadows, tactile depth (`/skeuo` · `/skeuomorphic-ui`) - `gateway-cache-awareness` - Protect prompt-cache hit rate when routing an agent through your own gateway/proxy - `akarso` - Post/schedule/reply across 14 social platforms; paired with the native **`akarso`** tool + `/akarso` (auto-installed) - `openseo` - SEO research/audits (keywords, backlinks, rank, site audit) via the OpenSEO MCP; `/openseo` - `dialkit` - Live-tune interface parameters (dials/sliders/timelines) across React/Svelte/Vue/Solid; `/dialkit` - `clone-website` - Website reverse-engineering - `cybersecurity` - Security investigations and DFIR - `sc-research` - Whitehat smart-contract / DeFi research router (repo: `skills/security/SCA/`) plus protocol/tool playbooks (vaults, AMMs, lending, bridges, gov, oracles, proxies, invariants, Halmos, BattleChain, X-linked bounty intel, historical vuln library, frontend/ops; lawful/in-scope only) - `cad` / `cad-viewer` / `dxf` / `gcode` / `urdf` / `srdf` / `sdf` - STEP-first CAD, fabrication, robot descriptions, and visual review from Text-to-CAD - `mobile-harness` - Android, iOS, and cloud-phone control through the capability-gated `mobilerun-core` facade - `apple-design` - Apple-style interface design - And 1,000+ more: the intent index knows 1,711 skills. About 120 are nur's own and live in this repo; the rest install from their upstreams at pinned commits... **Browse skills:** ```text /skills ``` --- ## AKM Agent Knowledge Management - a package manager for skills, commands, and tools across Claude, OpenCode, and Cursor. **Requires:** Node.js --- ## Manual management ```bash nur ecosystem ensure # install / repair all components nur ecosystem ensure --force # force re-install (useful after updates) nur ecosystem status # show readiness per component nur ecosystem install # print the pinned install plan for one component nur ecosystem install --yes # execute the printed plan ``` ## Pinned supply chain Everything `nur ecosystem ensure` (and the bootstrapper) installs is pinned in `src/ecosystem/lock.json`: npm packages by exact version, pypi packages by exact version, git-sourced components (skill packs, default plugins, dogwood, sgrep) by full commit SHA. Skill packs are cloned at their pinned commit with `git` (never via `npx skills add`, which resolves a mutable ref) and the checkout's `HEAD` is verified against the pin before any SKILL.md is copied. Nothing resolves `latest` at install time. ### Vendored skills in the repository Third-party skill trees are not committed. `skills/upstream-lock.json` pins each one (repository, commit, source path, sha256 per file, and the name and description the intent index reads) and `skills/.gitignore`, generated from the lock, keeps the files out of Git. A maintainer rebuilds the local mirror from the pins, and every file must match its recorded hash before it is written: ```bash python scripts/sync_upstream_skills.py --restore ``` Two kinds of locked file stay tracked: packs nur ships itself (`skills/bundled.json`) and files nur builds from reviewed sources (an entry's `built` list, such as gstack's offline diagram renderer). After a refresh, `--record-meta` re-records names and descriptions and regenerates the ignore file; `scripts/check_skill_snapshot.py` fails when a pinned file is committed, the ignore file is stale, or a present mirror file has drifted from its pin. `nur ecosystem install ` shows the plan (source, pin, target directory) and does nothing without `--yes`. Known names: `penecho`, `sgrep`, `ast-grep`, `dogwood`, `headroom`, `graphjin`, `egaki`, `akarso`, `ruflo`, `plur`, `executor`, `agent-browser`, `excalidraw`, `skills`, `akm`, `graphify`, `fractal`, the skill-pack labels (`web3-skills`, `infinite-headcount`, …) and catalog plugin ids (`superpowers`, `fable`, …). **Environment variables:** | Variable | Purpose | |----------|---------| | `CLAUDE_FLOW_DB_PATH` | Ruflo database path | | `CLAUDE_FLOW_MEMORY_PATH` | Ruflo home path | | `GRAPHJIN_CONFIG_PATH` | Override the GraphJin client config path | | `PENECHO_STATE_DIR` | Override the penecho state directory | | `HELIX_URL` / `NUR_HELIX_URL` | Select / activate the HelixDB memory resident (`[helix_memory] auto`) | | `HELIX_API_KEY` | Optional HelixDB bearer token | | `HEADROOM_PROTECT_READS` | Upstream Headroom knob: tools to keep out of inline compression; honored because the helper inherits the environment |