SwarmVault
The local-first LLM Wiki toolchain: compile docs, code, transcripts, and URLs into a durable markdown vault, knowledge graph, and retrieval index — all on disk and auditable.
Evidence shows deliberate safety design: search enforcement is strictly opt-in (default advisory; SWARMVAULT_GRAPH_FIRST=off disables; repeat searches always allowed), install-time .gitignore/tsconfig edits are disclosed with opt-out via SWARMVAULT_OUT, raw sources stay immutable. Deductions: the skill delegates to a third-party npm CLI running as a resident binary/hooks whose supply chain is not auditable from the skill alone, and hooks modify agent settings (~/.claude); full marks not justified.
Good self-consistency: SKILL.md, references/commands.md, examples, and TROUBLESHOOTING agree; failure feedback is detailed (doctor, lint, graph validate, concrete versioned fixes like the 3.14.1 MCP patch). Static review cannot execute reproduction, so capped at 10; extra dependency on external CLI version alignment keeps it at the cap rather than higher.
Scenarios, triggers, and boundaries are clearly declared ('When To Use This Skill', graph-first heuristics, heuristic-provider limitations). Deductions: core function depends entirely on npm global install and GitHub fetches, a mainland-China reachability risk; no Chinese-language support; SKILL.md is truncated mid-sentence ('Defaults to pr'), leaving some boundary information missing.
Well-layered docs (SKILL.md → README → examples/references → TROUBLESHOOTING), explicit version 3.21.0, MIT license, clear update paths, candid limitation disclosure. Deductions: unverified publisher with maintenance responsibility shown only via npm/ClawHub channels; truncated SKILL.md; no changelog evidence in scope; the CLI itself is not bundled, implying external version coupling.
The claimed value (durable local wiki/graph/retrieval/task ledger) matches the toolchain, with zero-config demo/quickstart and explicit artifact checklists. Static review cannot verify representative outputs; comparative-benefit evidence is limited; capped at 7 statically, reduced to 6.
Cites the Karpathy gist, command-level artifact paths, and a validation/smoke-prompts.md file — auditable primary material. But committed CI workflows and test suites covering the skill's key paths do not appear in this evidence, so conclusions rest largely on author claims; statically capped at 5, scored 4.
- The skill does not bundle its CLI: real capability depends entirely on the external npm package @swarmvaultai/cli; verify swarmvault --version before use.
- install --agent --hook writes agent config (e.g. .claude/settings.) and edits .gitignore/tsconfig.; use install status for a dry check or set SWARMVAULT_OUT to skip hygiene edits.
- Search denial (--graph-first) happens only after explicit opt-in; if a search is denied, follow the redirect to graph query / graph callers — repeating the same search is always allowed.
- npm installs, GitHub clones, and YouTube fetching depend on overseas networks and may be unreachable or slow from mainland China.
- raw/ stores local copies of ingested sources — beware sensitive files entering the vault; this review is static, no commands were executed.
What does this skill do, and when should you use it?
SwarmVault is an open-source (MIT) CLI that turns Andrej Karpathy's LLM Wiki pattern into a working local toolchain: three layers of raw sources (raw/), a generated wiki (wiki/), and a schema file (swarmvault.schema.md). It ingests 30+ input formats (PDF, Word, Excel, transcripts, code, YouTube captions, and more) and compiles a provenance-tagged knowledge graph, local hybrid retrieval (SQLite FTS plus optional embeddings), dashboards, and reviewable change queues. The built-in heuristic provider runs fully offline with no API keys; optional model providers include Ollama, OpenAI, Anthropic, and others. It also ships an MCP server, an agent task ledger, token-budgeted context packs, and rule/hook installers for dozens of coding agents.
Initializes vaults and ingests files, directories, public GitHub repos, or docs hubs via quickstart/scan/clone/source add; runs AST analysis on code with tree-sitter/SQL parsers to emit module pages; compiles wiki pages, state/graph., and share cards; answers with save-first query/chat including persisted multi-turn sessions; builds cited, token-bounded context packs (context build) and a durable task ledger (task start/update/finish); exports static AI handoffs (llms.txt, JSON-LD) via export ai; exposes the vault over stdio MCP with graph, retrieval, and task tools; runs doctor --repair health checks; and supports git hooks, watch mode, approval queues (compile --approve, candidate), and automatic contradiction detection.
- A developer wants Claude Code or Codex to answer who-calls-this or blast-radius questions from a precompiled graph instead of repo-wide grep — exact file:line call-site evidence at a fraction of the tokens
- A researcher compiles papers, articles, meeting transcripts, and notes into an evolving personal research base with automatic contradiction detection across sources
- A reader builds a chapter-by-chapter fan wiki with character and theme pages that compound as they finish a book
- A consultant turns customer-call recordings, Slack exports, and email/calendar into a searchable business knowledge base (audio needs local Whisper or a hosted provider)
- An Obsidian user who wants graph views, provenance, and automation can export an Obsidian-native bundle with Dataview dashboards and typed links
- Teams need to expose the vault to any agent client over MCP, or feed crawlers via llms.txt and JSON-LD static exports
What are this skill's strengths and limitations?
- Runs fully local and offline by default with no API keys; code parsing never sends source to external APIs
- Every edge is tagged extracted/inferred/ambiguous, new concepts land in a candidates area first, and changes can go through approval queues — mitigating hallucination compounding
- Ingests 30+ formats including PDF, the full Office families, transcripts, 30+ programming languages, and YouTube captions
- Built-in MCP server plus rule/hook installers for 50+ coding agents; graph-first hooks materially reduce blind repo-wide searches
- Task ledgers, context packs, and chat transcripts land as git-friendly markdown/JSON on disk, reusable across sessions
- Requires Node >= 24, a fairly recent runtime
- The default heuristic provider has limited extraction quality; the docs themselves recommend adding Ollama or a cloud model for richer synthesis
- Audio/video ingest depends on external binaries (whisper.cpp, ffmpeg, yt-dlp) you must install separately
- The local-whisper provider is documented as experimental in 1.1.0; R language parsing lacks a safe packaged grammar and only emits a diagnostic
- The README does not publish a standalone test suite or full benchmarks (it points to SCALE.md for operating envelopes)
How do you install this skill?
CLI: npm install -g @swarmvaultai/cli (requires Node >= 24), verify with swarmvault --version. A Node-free desktop app for macOS/Windows/Linux is available from the website's download page. OpenClaw/ClawHub users can run clawhub install swarmvault. The skill file lives at skills/swarmvault/SKILL.md in the repo.
How do you use this skill?
Fastest path: run swarmvault quickstart ./your-repo in an empty folder — it does init + ingest + compile + opens the graph viewer. No repo handy? Use swarmvault demo. Common follow-ups: swarmvault next (read-only state guidance), swarmvault query "question" (saves answers to wiki/outputs/ by default), swarmvault graph serve (local workbench), swarmvault doctor (health check). Prefer swarmvault graph query/callers/path over reading source for code-understanding questions. Without API keys, the offline heuristic provider runs by default; for sharper synthesis configure Ollama + Gemma (ollama pull gemma4, then set providers/tasks in swarmvault.config.).
How does this skill compare with similar options?
Positioned by the project as the production-grade implementation of Karpathy's LLM Wiki gist (three layers go from described to runnable, with CLI commands, a typed graph, approval queues, and MCP). It also calls itself an Obsidian alternative, but in practice the relationship is complementary — graph export --obsidian produces Obsidian-native bundles with Dataview dashboards and Breadcrumbs/Juggl typed links.