A personal knowledge base that writes itself.
Drop a source. Claude does the bookkeeping.
Your knowledge compounds — in plain markdown you own.
"Obsidian is the IDE. Claude is the programmer. The wiki is the codebase."
The Graph view. Every note is a star sized by its links, each community its own hue, [[wikilinks]] are the connective tissue.
Most LLM-plus-documents setups re-derive knowledge on every query. RAG finds chunks, the model stitches an answer, nothing is kept. Ten queries against the same docs → ten rediscoveries.
myco inverts this. You add a source once. Claude reads it, integrates it into a persistent wiki, flags contradictions against older pages, wires up citations, and commits the result. By query #10 the wiki itself answers — the bookkeeping already happened.
Based on Andrej Karpathy's LLM Wiki pattern. The idea goes back to Vannevar Bush's 1945 Memex, which the project was originally named after.
raw/ Original sources. Immutable.
│ Ingest
▼
wiki/ Claude-maintained pages. Cited [^src-*], cross-linked.
│
▼
myco desktop + Obsidian (optional) + your shell / git client
All see the same files. myco never locks the vault.
Grab the bundle for your platform from the latest release:
- macOS (Apple Silicon):
myco_x.y.z_aarch64.dmg— unsigned; Gatekeeper warns on first open (right-click the app → Open → Open, or System Settings → Privacy & Security → "Open Anyway"). - Windows x64:
myco_x.y.z_x64-setup.exe— unsigned; SmartScreen warns on first launch (More info → Run anyway).
On first launch myco creates ~/Documents/myco/ seeded with maintenance rules
(CLAUDE.md), raw/, wiki/ (interconnected starter notes so the Graph is
populated on day one — delete them anytime), daily/, and ingest-reports/.
To use a different folder (e.g. an existing Obsidian vault):
Settings → Account → Change…
 |
 |
| Overview — stats, jump-back, recent activity | Provenance — citation coverage per page |
 |
 |
| Reader — source / split / preview + backlinks | Settings — separate Query / Ingest models |
Ingest — drop a file or paste text; it lands in raw/ and the active model
integrates it into wiki/ with citations, a log entry, and a WHY report.
Inputs are multimodal: PDF, Office docs, spreadsheets, images (vision
provider), audio/video (installed whisper CLI), YouTube URLs. A bundled
offline embedding index (bge-m3, in-process llama.cpp) powers semantic search
and per-page Related notes.
Ask & Agent mode — chat over your wiki with any connected model. Agent
mode turns a tool-capable provider into an autonomous researcher: it searches,
reads pages, traverses links, and answers with citations; optional write tools
are confirmed per call and never touch raw/. Audio overview renders an
answer's cited pages as a two-host spoken deep-dive, offline.
Graph — the whole vault as a 3D cosmic web (three.js + d3-force-3d). Louvain communities get hues and auto-named cluster labels, hubs bloom, orphans and ghost links show like Obsidian. Drag stars (the sim re-heats), isolate neighbourhoods, light shortest paths, play a timelapse of the vault building itself. Alternative layout engines include a static 2D atlas and a grown mycelium mat. 60 fps to ~10k nodes.
Study — generate flashcards from any page; plain-markdown decks with FSRS scheduling that round-trip with Obsidian's spaced-repetition plugin. Review due cards and generated quizzes in-app.
Schedules — recurring digests: a standing query, "what changed", staleness
sweeps, topic trackers. Each run writes a cited markdown note into digests/.
Distillation — an idle-time housekeeping pass scores new raw/session inflow against the wiki's topic clusters: junk is filtered before it costs an embedding, near-duplicate or off-topic items are quarantined then trashed, and raw sources already folded into the wiki move to a dated archive. Review its proposals on the Feedback page; tune intensity, gate strictness, and schedule from Settings. It also writes daily session digests, drafts topic maps for its clusters, and — with an optional profile — personalises Ask and ingest toward your stated role and interests.
Reader — CodeMirror source / live preview / split, [[wikilink]]
autocomplete, backlinks and Related notes panels. raw/ PDFs open in an
in-app pdf.js viewer: select text → highlight & cite mints a pinpoint link;
highlights live in a sidecar so raw/ stays immutable.
Providers — bundled offline embeddings (no key, no install), Claude Code CLI (your Pro/Max subscription), Anthropic / OpenAI / Google AI / OpenRouter APIs, and local Ollama. Separate model pick for Query vs Ingest, monthly cost budget. API keys go to the OS keychain — never plaintext on disk.
Web clipper — browser extension + bookmarklet (clipper/) that sends any
page or selection into the vault inbox via a deep link.
⌘K command palette (files, routes, full-text + semantic), EN / 한국어 / 日本語
UI, light/dark, responsive to 320px.
Use the same vault from Claude Desktop, Claude Code, or any MCP client.
Easiest path — the app hosts it. The desktop app runs the MCP server in-process (no Python needed). Open Settings → MCP → Connect to Claude Code. The server follows whichever vault the app has open; the token persists, so you connect once.
Standalone Python server (from-source / non-app clients)
Requires Python 3.10+ (stdlib-only runtime).
bash mcp-server/install.sh # creates mcp-server/.venv
bash mcp-server/serve.sh # serves http://127.0.0.1:22360/sse
claude mcp add --transport sse myco http://localhost:22360/sseOr stdio for Claude Desktop — add to claude_desktop_config.json:
{
"mcpServers": {
"myco": {
"command": "<repo>/mcp-server/.venv/bin/python",
"args": ["<repo>/mcp-server/myco_mcp.py"]
}
}
}28 tools: read (list_pages read_page search folder_tree …), write
(add_raw_source create_page update_page git_commit …), inbox, no-LLM
quality checks (lint_citations trust_report contradictions …), and
multi-project governance (resolve_cross_links export_project
register_vault …). The standalone server manages multiple independent wikis
under projects/<slug>/, each with its own wiki/ raw/ CLAUDE.md.
Prerequisites: Node 20+, Rust 1.77+, and the Tauri prerequisites for your OS. Clone with Git LFS (the bundled embedding model is stored in LFS).
cd app
npm install
npm run tauri dev # hot-reload dev window
npm run tauri build # release bundle in src-tauri/target/release/bundle/Tests and lint:
cd app && npm run lint && npx tsc -b && npx vitest run
cd app/src-tauri && cargo fmt --check && cargo clippy --all-targets -- -D warnings && cargo testSee app/README.md for the development guide and
docs/SIGNING.md for the release/signing process.
Issues and PRs welcome.
- Pattern: Andrej Karpathy — LLM Wiki.
- Ancestor: Vannevar Bush, "As We May Think", 1945.
- Built with: Claude Code.