SwarmVault CLI

Install

Works in

• Claude Code
• Codex
• Cursor
• GitHub Copilot
• Gemini CLI
• OpenCode
• Goose
• Amp
• Cline

npx skillfoxx add cli/swarmvault

What it does

SwarmVault installs from npm and turns documents, code, notes, transcripts and page addresses into a local three-layer store. `raw/` holds immutable copies of sources, `wiki/` holds Markdown pages, and `state/` holds the `graph.json` knowledge graph and a search index. The `swarmvault quickstart ./repo` command creates the store, ingests a file, a folder or a public GitHub repository, compiles the wiki and graph and opens a local graph viewer. Queries sit alongside: `swarmvault query` answers from the wiki, while `swarmvault graph query`, `graph path`, `graph explain` and `graph callers` walk the links, including function calls with file and line. The store reaches an agent through an explicit `swarmvault install --agent <name>` command, and neither `init` nor `quickstart` writes rules into a project by default. For Claude Code the `--hook` and `--mcp` flags add hooks and an entry in `.mcp.json`. The session-start hook advises asking the graph first and searching files only after, and after edits it starts a background graph refresh for the changed file. The `swarmvault mcp` command exposes the same store to any MCP client over stdio. Graph edges are tagged as extracted, inferred or ambiguous. New concepts land in `wiki/candidates/` first, and `compile --approve` stages changes as bundles for approval. The first run needs no keys: the built-in heuristic provider works offline. For sharper extraction you connect Ollama or cloud models, and a key is given as an environment variable name.

Who it is for. For people running a large project or research with dozens of documents in Claude Code, Codex or Cursor who want the agent to follow graph links instead of rereading files.

Example request

Build a SwarmVault store from the docs folder and sources of this repository and show who calls the parseConfig function

Limitations

• It needs Node 24 or newer. Install is global, through `npm install -g @swarmvaultai/cli`.
• Without a model the heuristic provider runs: it is fine for an offline start and light extraction, while sharper results need Ollama or a cloud model. A cloud provider receives the text of your sources.
• The `local-whisper` provider for audio transcription is marked experimental in the documentation and downloads a model of about 147 MB to work.
• Hooks only advise a search order by default. Blocking the first broad search is enabled with the `--graph-first` flag and fires once per session.
• The `swarmvault install` command appends the store folders to `.gitignore` and to `exclude` of strict `tsconfig.json` files, so the source copies in `raw/` do not break the project's type check.

How to disable. For one agent, run `swarmvault install status --agent claude` to see the written files and delete them: the hooks in settings, the `swarmvault` entry in `.mcp.json`, the `.claude/skills/swarmvault` folder. The store itself is plain `raw/`, `wiki/` and `state/` folders in the project, which you can delete along with `swarmvault.config.json`. The global package is removed with `npm uninstall -g @swarmvaultai/cli`.

README in short

The repository holds a TypeScript monorepo of the CLI and engine packages, documents on stability and scale, a Dockerfile with metadata for MCP server registries. The overview opens with a 30-second start, a comparison with Andrej Karpathy's original idea and a table of thirty-plus input formats. Further come model setup, connection to more than thirty agents and automation modes. Chinese and Japanese translations are included. MIT license.