An autonomous knowledge base that grows as you work.
LLM Wiki is a Claude Code plugin that captures research, ideas, and decisions into an interlinked wiki with semantic search, automatic research, and a Wikipedia-style web UI. Knowledge compounds over time — the more you use it, the smarter it gets.
Autonomous knowledge base plugin for Claude Code - captures reserch, ideas, and decisions into an interlinked wiki with reserch-on-miss, semantic search, and a Wikipedia-style web UI. Knowledge compounds as you work.
星數
50
工具
—
可用率
—
Inspired by Andrej Karpathy's LLM Wiki pattern: raw sources are immutable, the LLM maintains the wiki layer, and a schema governs behavior.
Core Features
Knowledge Management
Automatic capture — saves research, ideas, decisions, and findings to the wiki as you work
Smart retrieval with research-on-miss — checks wiki first, automatically researches and ingests if not found
Full-text search — TF-IDF keyword search with content-aware scoring and snippet extraction
Block references & transclusion — [[page#heading]] links and ![[page#section]] embeds
Backlink panel with automatic unlinked mention detection
Frontmatter query language — Dataview-like queries: SELECT title, type FROM pages WHERE confidence = "high"
Intelligent freshness — 9-tier staleness system from live (15 min) to permanent (never expires)
Research-on-Miss
Automatic research — /wiki-read researches topics not in the wiki using available tools
Tool discovery — works with whatever tools the user has (WebSearch, WebFetch, Wikipedia API, MCP tools)
Auto-ingestion — saves findings to wiki with proper citations
Web UI Features
Wikipedia-style browsable website with 4 themes (light, dark, terminal, wikipedia)
Interactive knowledge graph (Cytoscape.js) with multiple layouts, clustering, and neighborhood highlighting
Canvas/whiteboard view for spatial page arrangement
Split-pane markdown editor with live preview and AI assist
Live research — click any red link to auto-research the topic
Smart caching with adaptive TTL and stale-while-revalidate
Circuit breakers for external API resilience
Git integration with auto-commit, attribution, and undo
How It Works
The wiki operates in a simple cycle: when you ask a question, it first checks its knowledge base. If found, it returns a cited answer. If not found, it automatically researches the topic, ingests the findings, and provides an answer—all without breaking your workflow.
sequenceDiagram
User->>wiki-reader: /wiki-read "What is X?"
wiki-reader->>wiki-index: Check knowledge base
alt Found in wiki
wiki-index-->>wiki-reader: Page exists
wiki-reader->>User: Cited answer from wiki
else Not found
wiki-index-->>wiki-reader: No results
wiki-reader->>search-orchestrator: Research needed
search-orchestrator->>search-channel: Fan out queries (web, academic, code, docs)
search-channel->>research-processor: Raw search results
research-processor->>wiki-writer: Processed findings
wiki-writer->>wiki-pages: Create/update page
wiki-writer->>User: Cited answer with new page
end
Quick Start
Installation
Install the plugin:
claude plugin install ./llm-wiki
Or copy manually:
cp -r llm-wiki .claude/plugins/
Restart Claude Code — dependencies install automatically on first session.
First Commands
Start using the wiki immediately with any of these:
Command
Purpose
/wiki-write https://example.com/article
Ingest a web page
/wiki-read "What is transformer attention?"
Ask — researches if not in wiki
/wiki-serve
Browse the wiki at localhost:8420
/wiki-maintain
Health check and optimization
Dependencies
Core dependencies (fastapi, uvicorn, mcp, etc.) are installed automatically via the plugin's SessionStart hook. For optional enhanced features:
Generate study guide, timeline, glossary, or comparison
How Research-on-Miss Works
When you ask a question that's not in the wiki, the entire research pipeline activates automatically. Here's the flow:
sequenceDiagram
actor User
participant WR as wiki-reader
participant Index as wiki index
participant SO as search-orchestrator
participant SC as search-channel
participant RP as research-processor
participant WW as wiki-writer
participant BM as backlink-manager
User->>WR: /wiki-read "What is X?"
WR->>Index: Check for matching pages
alt Page found
Index-->>WR: Return page
WR-->>User: Cited answer from wiki
else No match
Index-->>WR: No results
WR->>WR: Detect query intent & complexity
WR->>SO: Trigger research
SO->>SO: Route to search channels
SO->>SC: Dispatch to web, academic, code, docs channels
par Parallel Research
SC->>SC: Web search
SC->>SC: Academic search
SC->>SC: Code search
SC->>SC: Docs search
end
SC-->>RP: Raw results
RP->>RP: Deduplicate, condense, rank
RP-->>WW: Processed findings
WW->>WW: Synthesize findings into page
WW->>BM: Update backlinks
BM->>Index: Register page
WW-->>User: Cited answer with new wiki page
end
Architecture
System Overview
LLM Wiki consists of 5 entry points (skills), 10 autonomous agents, utilities in the bin/, and a persistent data layer in .wiki/.
flowchart TD
User([User]) -->|Invokes| Skills
subgraph Skills["5 Entry Points"]
W["/wiki-write<br/>Ingest & Update"]
R["/wiki-read<br/>Search & Ask"]
S["/wiki-serve<br/>Web UI"]
M["/wiki-maintain<br/>Health Check"]
V["/wiki-view<br/>Dashboard"]
end
Skills -->|Route to| Agents
subgraph Agents["10 Autonomous Agents"]
subgraph write["Write Pipeline"]
WW["wiki-writer<br/>(Sonnet)"]
BM["backlink-manager<br/>(Haiku)"]
end
subgraph read["Read Pipeline"]
WR["wiki-reader<br/>(Haiku)"]
SO["search-orchestrator<br/>(Sonnet)"]
SC["search-channel<br/>(Haiku)"]
end
subgraph research["Research Pipeline"]
RL["research-loop<br/>(Sonnet)"]
RP["research-processor<br/>(Haiku)"]
end
subgraph quality["Quality Pipeline"]
WA["wiki-auditor<br/>(Haiku)"]
FC["fact-checker<br/>(Sonnet)"]
CE["citation-explorer<br/>(Sonnet)"]
end
end
Agents -->|Read/Write| Data
Agents -->|Use| Bin
subgraph Bin["Utilities (bin/)"]
Search["search.py<br/>TF-IDF"]
Cache["cache.py<br/>Vectors"]
BL["backlinks.py<br/>Links"]
Gap["gaps.py<br/>Analysis"]
Git["git.py<br/>Tracking"]
end
subgraph Data[".wiki/ Data Layer"]
Pages["pages/<br/>Markdown"]
Index["index.md<br/>Catalog"]
Cache2["cache/<br/>SQLite"]
Raw["raw/<br/>Sources"]
Schema["SCHEMA.md<br/>Rules"]
end
S -->|Serves| UI["Web Server<br/>localhost:8420"]
UI -->|Renders| UIFeatures["4 Themes, Graph,<br/>Editor, Chat, Review"]
Directory Structure
llm-wiki/
.claude-plugin/ Plugin metadata (plugin.json, marketplace.json)
agents/ 10 autonomous agents
bin/ 23 CLI utilities (search, backlinks, gaps, cache, git, ...)
mcp/ MCP server for wiki operations
rules/ Workflow and integration rules
skills/ 5 user-facing skills
serve/ Web UI server and assets
scripts/ FastAPI server, WikiStore, RAG, chat, research workers
static/ JavaScript + CSS (4 themes)
templates/ 16 Jinja2 templates
Agent Collaboration During Research
When /wiki-read deep triggers a deep research operation, agents coordinate like this:
sequenceDiagram
participant OR as search-orchestrator
participant SC as search-channel
participant RP as research-processor
participant WW as wiki-writer
participant BM as backlink-manager
participant FC as fact-checker
OR->>OR: Classify query complexity
OR->>SC: Fan out to 4 channels (web, academic, code, docs)
par Parallel Search
SC->>SC: Execute web search
SC->>SC: Execute academic search
SC->>SC: Execute code search
SC->>SC: Execute docs search
end
SC-->>RP: Raw results stream
RP->>RP: Deduplicate & condense
RP->>WW: Processed findings
WW->>WW: Synthesize into wiki page
WW->>BM: Update backlinks
BM->>BM: Maintain reverse index
FC->>FC: Verify claims
FC->>WW: Flag uncertainties
WW-->>OR: Complete
Data Flow: From Source to Wiki
How content flows from raw sources into the wiki knowledge base:
External API calls are protected by circuit breakers that gracefully degrade when services fail:
stateDiagram-v2
[*] --> Closed
Closed --> Open: Threshold exceeded<br/>(5 failures in 60s)
Open --> HalfOpen: Timeout<br/>(30s backoff)
HalfOpen --> Closed: Trial succeeds
HalfOpen --> Open: Trial fails
Closed --> Closed: Success or<br/>slow failure
Open --> Open: Requests<br/>rejected
HalfOpen --> HalfOpen: Testing<br/>recovery
note right of Closed
Normal operation
Requests pass through
end note
note right of Open
Circuit tripped
Fast-fail all requests
Cache responses
end note
note right of HalfOpen
Recovery test mode
Allows one request through
Monitors outcome
end note
Data Model
Wiki data lives in .wiki/ — its location is derived from the plugin install scope:
Auto-init — create .wiki/ automatically if it doesn't exist
Freshness-aware — 9-tier staleness system governs when content needs refresh
Troubleshooting
Issue
Solution
No wiki found
Run any /wiki-* command — .wiki/ auto-creates
Web UI won't start
Check port 8420 availability; run /wiki-serve stop then retry
Search returns no results
Run /wiki-maintain to rebuild indexes
Broken links
Run /wiki-maintain lint to auto-fix
Stale content
Run /wiki-write --refresh-stale to update old pages
License
MIT
L
LLM Wiki
Oshayr·50 顆星·
Autonomous knowledge base plugin for Claude Code - captures reserch, ideas, and decisions into an interlinked wiki with reserch-on-miss, semantic search, and a Wikipedia-style web UI. Knowledge compou
Autonomous knowledge base plugin for Claude Code - captures reserch, ideas, and decisions into an interlinked wiki with reserch-on-miss, semantic search, and a Wikipedia-style web UI. Knowledge compounds as you work.
上次檢查時間:2026-10-27。 Autonomous knowledge base plugin for Claude Code - captures reserch, ideas, and decisions into an interlinked wiki with reserch-on-miss, semantic search, and a Wikipedia-style web UI. Knowledge compounds as you work.