grimoire

Grimoire

Your knowledge, structured for machines and humans.

Grimoire turns any topic into a structured, compounding knowledge base. It's a Claude Code plugin that runs a five-stage pipeline — scout the web for sources, ingest them into wiki articles, compile a cross-referenced graph, present a study-oriented static frontend, and serve the whole thing to LLMs over MCP.

Every handoff between stages is a plain markdown file you can edit. Scout, ingest, and present all pause for your review before the next stage runs, so humans stay in control. The output is local-first, LLM-readable, and yours — a wiki you own on disk, a website you open from file://, and an MCP server any LLM client can query.

/plugin marketplace add vedantggwp/athanor
/plugin install grimoire@athanor
/grimoire:init

— or in natural language to Claude Code: "Create a new grimoire about reinforcement learning from human feedback."

Install

Grimoire is distributed through the Athanor marketplace. From Claude Code:

/plugin marketplace add vedantggwp/athanor
/plugin install grimoire@athanor

That's it. No npm install, no build step, no dependency setup. The plugin ships with pre-built bundles in dist/ that inline every runtime dependency (papyr-core, @modelcontextprotocol/sdk, zod), so compile, present, and serve run straight from node dist/*.js with nothing else on the machine.

Skills auto-discover from skills/*/SKILL.md once the plugin is loaded.

Developing Grimoire locally

Only needed if you're contributing to Grimoire itself:

git clone https://github.com/vedantggwp/grimoire.git
cd grimoire
npm install
npm run build

npm run build regenerates dist/{compile,present,serve}.js via esbuild. The bundles are committed so marketplace installs work without a build step — rerun npm run build whenever you change anything in lib/ and commit the updated bundles alongside the source.

Quick start

From any working directory with the plugin loaded:

> /grimoire "reinforcement learning from human feedback"

Or in natural language: "Build a grimoire about WebGPU for engineers."

That's it. Grimoire infers topic, scope, audience, and palette from your sentence, scouts sources, pauses once for you to curate them, ingests in batch, compiles the graph, generates the frontend, and pauses once more for your review.

Power user flow

For granular control, run each stage individually:

> /grimoire:init           # 7-question questionnaire + scaffold
> /grimoire:scout          # research sources, produce approved-sources.md
> /grimoire:ingest         # fetch, preserve raw, compile articles
> /grimoire:compile        # graph, backlinks, overview, gaps
> /grimoire:present        # build the static frontend
> /grimoire:serve          # start the MCP server

Flags for the one-command flow:

• --guided — run the full 7-question questionnaire
• --review-angles — approve scout search angles before searching
• --sequential — process sources one-at-a-time with per-source approval
• --from <path> — inherit design from an existing grimoire
• --palette <name> — override the default palette

The six skills

Skill	Status	What it does
run	Working	One-command pipeline: init → scout → ingest → compile → present, 2 taste checkpoints
init	Working	7-question interactive questionnaire and workspace scaffold
scout	Working	Web research with 6-signal confidence scoring, human review checkpoint before ingest
ingest	Working	Fetches approved sources, preserves raw text, compiles wiki articles with a human checkpoint
compile	Working	Papyr Core graph audit, backlink repair, overview evolution, gap analysis, emergent taxonomy
present	Working	Static frontend with 6 study modes (read, graph, search, feed, gaps, quiz)
serve	Working	MCP server exposing 7 tools for LLM knowledge access

init, scout, and ingest are Claude-driven workflows defined in SKILL.md. compile, present, and serve have matching TypeScript runtimes in lib/ that esbuild bundles into self-contained ESM files under dist/; each skill invokes its bundle directly via node ${CLAUDE_PLUGIN_ROOT}/dist/<skill>.js, so nothing needs to be installed on the user's machine.

How it works