Languages: English | 简体中文 | 日本語

The local-first LLM Wiki, knowledge graph builder, and RAG knowledge base for AI agents. SwarmVault turns docs, code, transcripts, notes, and URLs into a durable markdown wiki plus a local graph you can inspect, query, and hand to agents. Start with one command, then learn the deeper graph, review, context-pack, and automation workflows when you need them.

Documentation on the website is currently English-first. If wording drifts between translations, README.md is the canonical source.

npm install -g @swarmvaultai/cli
swarmvault quickstart ./your-repo

quickstart initializes a vault in the current directory, ingests the input, compiles the wiki and graph, writes share artifacts, and opens the local graph viewer. It is the beginner-friendly alias for swarmvault scan.

No repo handy?

swarmvault demo

After your first compile, the most useful next commands are:

swarmvault next
swarmvault query "What are the key concepts?"
swarmvault graph serve
swarmvault doctor
swarmvault candidate list

Not sure what state the vault is in? swarmvault next is read-only and tells you whether to initialize, ingest, compile, query, review, or refresh.

No API keys are required for the first run. The built-in heuristic provider runs locally and offline.

What you get on disk:

• raw/ - immutable copies of ingested material
• wiki/ - generated markdown pages, saved outputs, graph reports, context packs, and task notes
• state/graph.json - the machine-readable knowledge graph
• state/retrieval/ - local search index
• wiki/graph/share-card.md, wiki/graph/share-card.svg, and wiki/graph/share-kit/ - copyable and visual first-run summaries

SwarmVault uses three layers, following the pattern described by Andrej Karpathy:

1. Raw sources (raw/) — your curated collection of source documents. Books, articles, papers, transcripts, code, images, datasets. These are immutable: SwarmVault reads from them but never modifies them.
2. The wiki (wiki/) — LLM-generated and human-authored markdown. Source summaries, entity pages, concept pages, cross-references, dashboards, and outputs. The wiki is the persistent, compounding artifact.
3. The schema (swarmvault.schema.md) — defines how the wiki is structured, what conventions to follow, and what matters in your domain. You and the LLM co-evolve this over time.

In the tradition of Vannevar Bush's Memex (1945) — a personal, curated knowledge store with associative trails between documents — SwarmVault treats the connections between sources as valuable as the sources themselves. The part Bush couldn't solve was who does the maintenance. The LLM handles that.

Turn books, articles, notes, transcripts, mail exports, calendars, datasets, slide decks, screenshots, URLs, and code into a persistent knowledge vault with a knowledge graph, local search, dashboards, and reviewable artifacts that stay on disk. Use it for personal knowledge management, research deep-dives, book companions, code documentation, business intelligence, or any domain where you accumulate knowledge over time and want it organized rather than scattered.

SwarmVault turns the LLM Wiki pattern into a local toolchain with graph navigation, search, review, automation, and optional model-backed synthesis. You can also start with just the standalone schema template — zero install, any LLM agent — and graduate to the full CLI when you outgrow it.

If you liked Karpathy's LLM Wiki gist, SwarmVault is the production-grade version. Here's how it addresses the most common concerns from the community:

"Won't hallucinations compound?" — Every edge is tagged extracted, inferred, or ambiguous. Contradiction detection flags conflicting claims. compile --approve stages all changes into reviewable approval bundles. New concepts land in wiki/candidates/ first. lint --conflicts audits for contradictions on demand.

"Does it scale past 100 pages?" — Yes. Hybrid search merges SQLite full-text with semantic embeddings, so queries work without fitting every page into context. compile --max-tokens trims output to fit bounded windows. Graph navigation (graph query, graph path, graph explain) lets you traverse rather than search.

"Is it just for personal use?" — Git-backed workflows (--commit), watch mode with git hooks, scheduled automation, and an MCP server make it usable for teams. Agent integrations cover direct-rule targets plus the extended skill-bundle roster.

"Do I need API keys?" — No. The built-in heuristic provider is fully offline. For sharper extraction, pair with a free local LLM via Ollama. Cloud providers are optional.

Download the desktop app for macOS, Windows, or Linux — bundles its own runtime:

Download Desktop App | GitHub Releases

SwarmVault requires Node >=24.

npm install -g @swarmvaultai/cli

Verify the install:

swarmvault --version

Update to the latest published release:

npm install -g @swarmvaultai/cli@latest

The global CLI already includes the graph viewer workflow and MCP server flow. End users do not need to install @swarmvaultai/viewer separately.

Run this from an empty folder or a scratch folder where you want the vault artifacts to live:

mkdir my-vault
cd my-vault
swarmvault quickstart ../your-repo
swarmvault next

That is the easiest path for a new user. It does the same work as swarmvault scan: initialize the vault, ingest the input, compile the wiki and graph, write share artifacts, and open the graph viewer unless you pass --no-serve or --no-viz.

my-vault/
├── swarmvault.schema.md       user-editable vault instructions
├── raw/                       immutable source files and localized assets
├── wiki/                      compiled wiki: sources, concepts, entities, code, outputs, graph
├── state/                     graph.json, retrieval/, embeddings, sessions, approvals
├── .obsidian/                 optional Obsidian workspace config
└── agent/                     generated agent-facing helpers

If you want to keep generated artifacts outside the source tree, run with SWARMVAULT_OUT=.swarmvault-out. swarmvault.config.json and swarmvault.schema.md stay in the project root; raw/, wiki/, state/, agent/, and inbox/ resolve under the output directory.

Once the fast path makes sense, the same workflow can be run step by step:

swarmvault init --obsidian --profile personal-research
swarmvault ingest ./src --repo-root .
swarmvault ingest ./meeting.srt --guide
swarmvault add https://arxiv.org/abs/2401.12345
swarmvault compile
swarmvault next
swarmvault query "What is the auth flow?"
swarmvault graph serve

Use swarmvault source add https://github.com/karpathy/micrograd, swarmvault source add https://example.com/docs/getting-started, swarmvault source list, swarmvault source reload --all, and swarmvault source session transcript-or-session-id when the same repo, folder, or docs hub should stay registered and refreshable. For public GitHub repos, swarmvault clone https://github.com/owner/repo --no-viz and swarmvault source add https://github.com/owner/repo --branch main --checkout-dir .swarmvault-checkouts/repo are the reusable checkout paths.

Want the minimal LLM-Wiki starter instead? swarmvault init --lite creates just raw/, wiki/, wiki/index.md, wiki/log.md, and swarmvault.schema.md - no config, no state, no agent installs.

When the vault lives inside a git repo, ingest, compile, and query support --commit. compile --max-tokens <n> trims lower-priority pages for bounded context windows. swarmvault ingest ./customer-call.mp3, swarmvault ingest https://www.youtube.com/watch?v=dQw4w9WgXcQ, and swarmvault ingest --video https://example.com/product-demo.mp4 cover audio, YouTube transcript, and video workflows when the required providers or helper binaries are available.

You do not need API keys or an external model provider to start using SwarmVault. The built-in heuristic provider supports local/offline vault setup, ingest, compile, graph/report/search workflows, and lightweight query or lint defaults.

If you want a fully local setup with sharp concept, entity, and claim extraction, pair the free Ollama runtime with Google's Gemma model. No API keys required.

ollama pull gemma4

{
  "providers": {
    "llm": {
      "type": "ollama",
      "model": "gemma4",
      "baseUrl": "http://localhost:11434/v1"
    }
  },
  "tasks": {
    "compileProvider": "llm",
    "queryProvider": "llm",
    "lintProvider": "llm"
  }
}

When you run compile/query with only the heuristic provider, SwarmVault surfaces a one-time notice pointing you here. Set SWARMVAULT_NO_NOTICES=1 to silence it. Any other supported provider (OpenAI, Anthropic, Gemini, OpenRouter, Groq, Together, xAI, Cerebras, openai-compatible, custom) works too.

For local semantic graph query without API keys, use an embedding-capable local backend such as Ollama instead of heuristic:

{
  "providers": {
    "local": {
      "type": "heuristic",
      "model": "heuristic-v1"
    },
    "ollama-embeddings": {
      "type": "ollama",
      "model": "nomic-embed-text",
      "baseUrl": "http://localhost:11434/v1"
    }
  },
  "tasks": {
    "compileProvider": "local",
    "queryProvider": "local",
    "embeddingProvider": "ollama-embeddings"
  }
}

With an embedding-capable provider available, SwarmVault can also merge semantic page matches into local search by default. tasks.embeddingProvider is the explicit way to choose that backend, but SwarmVault can also fall back to a queryProvider with embeddings support. Set retrieval.rerank: true when you want the configured queryProvider to rerank the merged top hits before answering.

For cloud-hosted models, add a provider block with your API key:

{
  "providers": {
    "primary": {
      "type": "openai",
      "model": "gpt-4o",
      "apiKeyEnv": "OPENAI_API_KEY"
    }
  },
  "tasks": {
    "compileProvider": "primary",
    "queryProvider": "primary",
    "embeddingProvider": "primary"
  }
}

See the provider docs for optional backends, task routing, and capability-specific configuration examples.

For audio files (voice memos, meeting recordings, interviews), install whisper.cpp and let SwarmVault drive it locally — no API keys, no network traffic:

# macOS
brew install whisper-cpp
# Debian / Ubuntu
sudo apt install whisper.cpp

swarmvault provider setup --local-whisper --apply

That command verifies the binary, downloads the base.en ggml model (~147 MB) into ~/.swarmvault/models/, and registers the provider in swarmvault.config.json under providers.local-whisper with tasks.audioProvider pointed at it. From then on, swarmvault add voice-memo.m4a (or dropping audio into raw/inbox/) transcribes end-to-end offline; the existing ingest-time redactor scrubs secrets spoken aloud before they reach raw/ or wiki/. Tune accuracy with --model {tiny.en,small.en,medium.en,large-v3}, threads with localWhisper.threads, and override binary/model discovery via localWhisper.binaryPath / localWhisper.modelPath / SWARMVAULT_WHISPER_BINARY. The local-whisper provider type is documented as experimental in STABILITY.md for 1.1.0.

Prefer a hosted transcription provider instead? Point tasks.audioProvider at any provider with audio capability (OpenAI, Groq, etc.). YouTube transcript ingest does not require a model provider.

The fastest way to make SwarmVault useful is the managed-source flow:

swarmvault source add ./exports/customer-call.srt --guide
swarmvault source add https://github.com/karpathy/micrograd
swarmvault source add https://example.com/docs/getting-started
swarmvault source list
swarmvault source session file-customer-call-srt-12345678
swarmvault source reload --all

source add registers the source, syncs it into the vault, compiles once, and writes a source-scoped brief under wiki/outputs/source-briefs/. Add --guide when you want a resumable guided session under wiki/outputs/source-sessions/, a staged source review and source guide, plus approval-bundled canonical page edits when profile.guidedSessionMode is canonical_review. Profiles using insights_only keep the guided synthesis in wiki/insights/ instead. Set profile.guidedIngestDefault: true in swarmvault.config.json to make guided mode the default for ingest, source add, and source reload; use --no-guide when you need the lighter path for a specific run. It now works for recurring local files as well as directories, public repos, and docs hubs. Use ingest for deliberate one-off files or URLs, and use add for research/article normalization.

Set up your coding agent so it knows about the vault:

swarmvault install --agent claude --hook    # Claude Code + graph-first hook
swarmvault install --agent codex --hook     # Codex + graph-first hook
swarmvault install --agent cursor           # Cursor
swarmvault install --agent copilot --hook   # GitHub Copilot CLI + hook
swarmvault install --agent gemini --hook    # Gemini CLI + hook
swarmvault install --agent trae             # Trae
swarmvault install --agent claw             # Claw / OpenClaw skill target
swarmvault install --agent droid            # Droid / Factory rules target
swarmvault install --agent kiro             # Kiro IDE + always-on steering
swarmvault install --agent hermes           # Hermes user-scope skill
swarmvault install --agent antigravity      # Google Antigravity rules + /swarmvault workflow
swarmvault install --agent vscode           # VS Code Copilot Chat chatmode

Or expose the vault directly over MCP:

swarmvault mcp

Using OpenClaw or ClawHub? Install the packaged skill with:

clawhub install swarmvault

That installs the published SKILL.md plus a ClawHub README, examples, references, troubleshooting notes, and validation prompts. Keep the CLI itself updated with npm install -g @swarmvaultai/cli@latest.

Knowledge graph with provenance - every edge traces back to a specific source and claim. Nodes carry freshness, confidence, and community membership.

God nodes and communities - highest-connectivity bridge nodes identified automatically. Graph report pages surface surprising connections with plain-English explanations.

Contradiction detection - conflicting claims across sources are detected automatically and surfaced in the graph report. Use lint --conflicts for a focused contradiction audit.

Semantic auto-tagging - broad domain tags are extracted alongside concepts during analysis and appear in page frontmatter, graph nodes, and search.

Schema-guided compilation - each vault carries swarmvault.schema.md so the compiler follows domain-specific naming rules, categories, and grounding requirements.

Save-first queries - answers write to wiki/outputs/ by default, so useful work compounds instead of disappearing. Supports markdown, report, slides, chart, and image output formats.

Agent context packs - swarmvault context build "<goal>" --target <path|node|page> writes a cited, token-bounded handoff for coding or research agents. Packs include graph orientation, included evidence, explicit omissions when the budget is too small, and durable artifacts under wiki/context/ plus state/context-packs/.

Agent task ledger - swarmvault task start|update|finish|resume records task goals, linked context packs, decisions, graph evidence, touched paths, outcomes, and follow-ups as git-friendly JSON plus markdown under state/memory/tasks/ and wiki/memory/tasks/. Compile includes task nodes and decisions in the graph, and the viewer exposes the task history. Existing memory commands remain compatibility aliases.

Vault doctor and workbench - swarmvault doctor [--repair] checks graph artifacts, retrieval, review queues, watch state, migrations, managed sources, source/page counts, and task state. The graph viewer workbench surfaces prioritized next actions, every check with details, copyable suggested commands, safe repair, explicit capture modes, title/tag capture fields, budgeted context-pack creation, and task-start actions.

Reviewable changes - compile --approve stages changes into approval bundles. New concepts and entities land in wiki/candidates/ first. Nothing mutates silently.

Configurable profiles - compose vault behavior with profile.presets, profile.dashboardPack, profile.guidedSessionMode, profile.guidedIngestDefault, profile.deepLintDefault, and profile.dataviewBlocks in swarmvault.config.json instead of waiting for hardcoded product modes. personal-research is a built-in preset alias.

Guided sessions - ingest --guide, source add --guide, source reload --guide, source guide <id>, and source session <id> create resumable source sessions under wiki/outputs/source-sessions/, stage source reviews and source guides, and route approval-bundled updates either to canonical source/concept/entity pages or to wiki/insights/, depending on the configured guided-session mode. Set profile.guidedIngestDefault: true in swarmvault.config.json to make guided mode the default for ingest and source commands; use --no-guide to override.

Deep lint defaults - set profile.deepLintDefault: true in swarmvault.config.json to make swarmvault lint include the LLM-powered advisory pass by default. Use --no-deep when you want one structural-only lint run without changing the profile.

Web-search augmented lint — lint --deep --web enriches deep-lint findings with external evidence via a configured web-search provider (http-json or custom). Web search is currently scoped to deep lint; other commands query only local vault state.

Knowledge dashboards - wiki/dashboards/ gives you recent sources, a reading log, a timeline, source sessions, source guides, a research map, contradictions, and open questions. The pages work as plain markdown first, and profile.dataviewBlocks can append Dataview blocks when you want a more Obsidian-native view.

Retrieval, hybrid search, and rerank - local retrieval stores its SQLite FTS shard and manifest under state/retrieval/. When an embedding-capable provider is available, it can merge full-text hits with semantic page matches. tasks.embeddingProvider is the explicit way to choose that backend, but SwarmVault can also fall back to a queryProvider with embeddings support. Set retrieval.rerank: true when you want the configured queryProvider to rerank the merged candidate set before query answers. Use swarmvault retrieval status|rebuild|doctor to inspect or repair the index.

Token-budgeted compile and auto-commit - compile --max-tokens <n> trims lower-priority pages to keep generated wiki output inside a bounded token budget, and ingest|compile|query --commit can immediately commit wiki/ and state/ changes when the vault lives in a git repo.

Graph report health signals - graph report artifacts now include community-cohesion summaries, isolated-node and ambiguity warnings, and sharper follow-up questions when the graph has weakly connected or ambiguous regions.

Visual + post-ready share kit - every compile writes wiki/graph/share-card.md, wiki/graph/share-card.svg, and wiki/graph/share-kit/; swarmvault graph share --post prints concise text, swarmvault graph share --svg [path] writes a 1200x630 visual card, and swarmvault graph share --bundle [dir] writes markdown, post text, SVG, HTML preview, and JSON metadata for easy posting, linking, or screenshotting.

Graph blast radius, status, stats, validation, refresh, query filters, tree, merge, clustering, and report export - graph blast <target> traces reverse import impact through module dependencies, graph status [path] performs a read-only stale check over graph/report artifacts and tracked repo changes, and check-update [path] is the top-level compatibility alias for the same cron-safe status check. graph stats prints lightweight counts and relation mix, graph validate [graph] --strict checks duplicate ids, dangling references, confidence bounds, and conflicted-edge evidence before export/merge/push workflows, and graph update [path] / graph refresh [path] runs the code-only repo refresh cycle for graph artifacts with a 25% shrink guard unless --force is explicit; update [path] is the top-level compatibility alias. watch [path] --once can target one repo root without persisting watch config. graph query can filter traversal by relation, context group, evidence class, node type, or language, graph tree writes an interactive source/module/symbol HTML tree with expand/collapse controls and a node inspector, and tree is the top-level alias. graph merge combines SwarmVault or node-link JSON graphs into one namespaced artifact, and merge-graphs is the top-level alias. graph cluster [--resolution <n>] recomputes communities, degrees, god-node flags, and graph report pages from an existing graph without re-ingesting sources, and cluster-only [vault] is the top-level compatibility alias. graph export --report writes a self-contained HTML report with graph stats, key nodes, communities, and warnings; graph export --neo4j <path> is an alias for the Cypher export used before Neo4j imports.

Graph diff - swarmvault diff compares the current knowledge graph against the last committed version, showing added/removed nodes, edges, and pages so you can see exactly what a compile changed.

Worktree artifact roots - SWARMVAULT_OUT=<dir> relocates generated raw/, wiki/, state/, agent/, and inbox/ artifacts while keeping swarmvault.config.json and swarmvault.schema.md in the project root. Use it for isolated smoke tests, shared source trees, and repo worktrees where generated vault state should not sit beside source files.

Obsidian graph export - graph export --obsidian writes an Obsidian-friendly bundle that preserves wiki folders, appends graph connections with typed link frontmatter for Breadcrumbs/Juggl, emits community notes and orphan-node stubs, copies assets, generates Dataview dashboard pages, and includes a full .obsidian config with types.json, node-type color groups, and cssclasses on every page.

Adaptive graph communities - SwarmVault auto-tunes Louvain community resolution for very small or sparse graphs, then splits oversized or low-cohesion communities so graph reports stay scannable on larger repos. You can pin a specific value with graph.communityResolution in swarmvault.config.json or override one recompute with swarmvault graph cluster --resolution <n>.

Optional model providers - OpenAI, Anthropic, Gemini, Ollama, OpenRouter, Groq, Together, xAI, Cerebras, generic OpenAI-compatible, custom adapters, or the built-in heuristic for offline/local use.

Agent integrations - install rules for Codex, Claude Code, Cursor, Goose, Pi, Gemini CLI, OpenCode, Aider, GitHub Copilot CLI, Trae, Claw/OpenClaw, Droid, Kiro, Hermes, Google Antigravity, VS Code Copilot Chat, and the extended skill-bundle roster. Optional graph-first hooks bias supported agents, including Codex, toward the wiki before broad search. Antigravity installs under .agents/rules/ and .agents/workflows/; older fully managed .agent/ files are cleaned up during reinstall.

MCP server - swarmvault mcp exposes the vault to any compatible agent client over stdio, including graph stats, graph clustering refresh, community lookup, hyperedges, context-pack, task-ledger, compatibility memory-task, vault doctor, and retrieval health tools.

Built-in browser clipper - graph serve exposes a local /api/bookmarklet page and /api/clip endpoint so a running vault can capture the current browser URL, page title, selected text, markdown, HTML excerpts, and tags from the workbench or bookmarklet. URL-only bookmarklet clips use normalized add; selected text is imported through the inbox path.

Automation - watch mode, git hooks, recurring schedules, and inbox import keep the vault current without manual intervention.

Managed sources - swarmvault source add|list|reload|review|guide|session|delete turns recurring files, directories, public GitHub repos, and docs hubs into named synced sources with registry state under state/sources.json, source briefs under wiki/outputs/source-briefs/, resumable session anchors under wiki/outputs/source-sessions/, and guided integration artifacts under wiki/outputs/source-guides/. Public GitHub repo sources support --branch, --ref, and --checkout-dir for pinned branch/tag/commit scans and reusable checkouts.

Source artifact types:

External graph sinks - export to full HTML, lightweight standalone HTML, self-contained report HTML, SVG, GraphML, Cypher, JSON, Obsidian note bundles, or Obsidian canvas, or push the live graph directly into Neo4j over Bolt/Aura with shared-database-safe vaultId namespacing.

Large-repo hardening - long repo ingests and compile passes emit bounded progress on big batches, provider-backed non-code analysis chunks long extracted text before model calls, nested .gitignore and .swarmvaultignore files are respected with .swarmvaultinclude allowlists for intentional exceptions, parser compatibility failures stay local to the affected sources with explicit diagnostics, code-only repo watch cycles skip non-code re-analysis, and graph reports roll up tiny fragmented communities for readability.

Every edge is tagged extracted, inferred, or ambiguous - you always know what was found vs guessed.

Codex, Claude Code, OpenCode, Gemini CLI, and Copilot also support --hook for graph-first context injection. Agents in the extended roster install a project-level skill bundle at the tool's conventional skills directory (e.g. .cline/skills/swarmvault/SKILL.md, .codeium/windsurf/skills/swarmvault/SKILL.md).

Each folder has real input files and actual output so you can run it yourself and verify.

See the examples guide for step-by-step walkthroughs.

Providers are optional. SwarmVault routes by capability, not brand. Built-in provider types:

heuristic openai anthropic gemini ollama openrouter groq together xai cerebras openai-compatible custom

See the provider docs for configuration examples.

SwarmVault processes your data locally by default:

• Code files are parsed on your machine via the TypeScript compiler API, tree-sitter, or the SQL parser. Source code contents are never sent to external APIs.
• Documents and text are sent to your configured provider for semantic extraction. With the built-in heuristic provider, everything stays local.
• Images are sent to a vision-capable provider only when one is configured.
• Heuristic mode (the default) is fully offline — no API keys, no network calls.

When you add a model provider (OpenAI, Anthropic, Ollama, etc.), only non-code content is sent for LLM analysis. All graph building, community detection, and report generation happen locally.

• Docs: https://www.swarmvault.ai/docs
• Providers: https://www.swarmvault.ai/docs/providers
• Troubleshooting: https://www.swarmvault.ai/docs/getting-started/troubleshooting
• npm package: https://www.npmjs.com/package/@swarmvaultai/cli
• GitHub issues: https://github.com/swarmclawai/swarmvault/issues

pnpm install
pnpm lint
pnpm test
pnpm build

See CONTRIBUTING.md for PR guidelines, and docs/live-testing.md for the published-package validation workflow.

See STABILITY.md for the public-API contract, semver promise, and deprecation policy. Stable surfaces follow semantic versioning once 1.0.0 is cut; experimental surfaces may change in any minor release.

See SCALE.md for tested operating envelopes across small, medium, and large vaults plus knobs to tune when the defaults aren't enough. See docs/pdf-extraction.md for the 1.0 PDF extractor choice and known limitations.

• Website: https://www.swarmvault.ai
• Docs: https://www.swarmvault.ai/docs
• npm: https://www.npmjs.com/package/@swarmvaultai/cli
• GitHub: https://github.com/swarmclawai/swarmvault

MIT