LLM Wiki
Turns documents, web clips, highlights, and notes into a searchable, cited wiki that AI can keep current.
The local API is explicitly loopback-only, the MCP tool surface is enumerated, and workflow permissions are scoped by purpose, supporting a good but incomplete least-privilege showing; full hosted, extension, and cloud Routine permission boundaries are absent. The README explains flows among files, indexes, Postgres, S3, Mistral OCR, and Claude scheduling, and identifies create/edit/delete effects, but it lacks a complete privacy, retention, encryption, or secret-handling policy. Test names indicate attention to delete approval, authentication limits, isolation, SSRF, and upload hardening, yet no implementation is supplied, so confirmation and sensitive-data handling receive only partial credit. Dependency security is thorough: daily OSV scans, PR dependency review, npm and pip audits, lockfile integrity enforcement, and moderate-severity failure thresholds. Editable/Git-friendly Markdown and rebuildable derived indexes offer rollback options, but recovery after deletion is undocumented. Karpathy inspiration, source-linked citations, and Apache-2.0 licensing provide strong attribution.
README descriptions of local/hosted storage, MCP tools, and client architecture broadly agree with CI coverage for SQLite, Postgres, web, extension, isolation, and lifecycle behavior, supporting self-consistency. This remains a static review: no test results were supplied or independently run. Python, Node, LibreOffice, Mistral, Postgres, and S3 requirements or optional conditions are identified, while CI pins runtimes and uses lockfiles; degradation when external services are unavailable is not described. Apart from limited troubleshooting such as the PowerShell execution-policy workaround, the evidence does not show concrete user-facing errors for MCP, conversion, indexing, or hosted-service failures.
The documentation thoroughly identifies individual, AI, and organizational audiences and covers local, hosted, upload, browser-clipping, and scheduled-maintenance scenarios. Supported formats, tool behavior, protected files, loopback-only local access, and optional OCR/LibreOffice requirements establish useful capability boundaries, but hallucination risk, erroneous merges, scale limits, and hosted quotas are not adequately documented. The Routine prompt, guide-first convention, and distinct MCP tool responsibilities support precise triggering, although autonomous-write selection and conflict rules are under-specified. Concrete macOS, Linux, Windows, runtime-version, local/remote backend, and preview instructions provide strong environment fit.
The README is well structured around features, setup, ingestion, formats, disk behavior, tools, architecture, and roadmap, with actionable cross-platform installation commands. Names are internally consistent, but no API stability or compatibility promise is provided. Several examples and one Windows troubleshooting note are present, though there is no systematic FAQ. Loopback restrictions, optional OCR, LibreOffice requirements, and current ingestion-channel limits are disclosed, while AI-accuracy, capacity, and production-deployment limitations are missing. Apache-2.0 metadata matches a complete LICENSE, justifying full license credit. No version scheme, release history, or changelog is supplied, so versioning scores zero. The repository owner offers limited attribution, but the publisher is unverified and there is no contribution, support, vulnerability-reporting, or maintenance policy, leaving responsibility thin.
Outputs are ordinary Markdown with source footnotes, cross-links, visualizations, and deterministic linting, with the filesystem acting as the local source of truth; this strongly supports browsability, manual editing, Git use, and portability. Combining uploads, clipping, search, citations, and scheduled maintenance provides plausible marginal value, but claims about long-term compounding and improved institutional memory are largely promotional and not demonstrated by the supplied static evidence. The product can be tried free or self-hosted and offers a local mode, but it requires Python and Node, some formats need LibreOffice, and enhanced OCR, cloud Claude, databases, and object storage may add costs that are not quantified.
Many claims trace to concrete architecture, disk-layout, tool-table, and CI-test targets, but core implementation files and test outcomes are absent, so not every README claim is established. Security and isolation statements receive some corroboration from named tests and the security workflow, while LICENSE corroborates the license claim; nevertheless, documentation and configuration dominate the evidence and do not provide independent implementation-level confirmation. The roadmap is clearly separated from current features, and local versus hosted behavior is distinguished, but some marketing claims about outcomes are not explicitly labeled as hypotheses or unverified expectations, preventing full fact/inference separation.
- A nightly Claude Routine can autonomously create, edit, and delete wiki content; verify delete approval, conflict handling, backups, and recovery before using it for important material.
- Hosted operation may expose content to Postgres, S3, Claude cloud execution, and optional Mistral OCR; the supplied material does not fully document encryption, retention, deletion, or third-party processing policies.
- Security and reliability evidence largely consists of workflow configuration and test names, without implementation or test results; this assessment does not establish that the controls actually pass.
- AI-generated synthesis may introduce incorrect summaries, citations, or overwrites of manual edits; the README does not document model-accuracy limits or a human-review process.
What does this agent do, and when should you use it?
LLM Wiki is an AI-maintained knowledge-base system that organizes local files, uploads, and Chrome extension captures into a persistent wiki. Its components include a Next.js web app, HTTP API, MCP server, Chrome extension, and a Converter for PDF and Office extraction, unified behind the VaultFS storage abstraction. Local mode uses SQLite and the user's filesystem, while hosted mode uses Postgres and S3; the local service binds only to 127.0.0.1 and cannot serve a LAN or remote client directly. Claude, Codex, or another MCP-compatible application can search and read sources, create or revise wiki pages, and check citations and links. The principal outputs are ordinary Markdown pages plus assets such as SVG and CSV, while the durable store remains authoritative and the search index is derived and rebuildable. Unattended upkeep is documented through a scheduled Claude Routine that processes new material and refreshes affected pages, citations, and cross-references.
The workflow begins with files in a workspace, web uploads, or pages and PDFs captured by the Chrome extension. LLM Wiki accepts Markdown, PDF, Word, PowerPoint, spreadsheet, HTML, text-data, and image formats; the Converter handles heavier PDF and Office extraction, while a background watcher re-indexes changes made outside the app. An MCP client starts with guide and list_knowledge_bases, then uses search to browse paths, run full-text searches, or inspect the citation graph, and read to retrieve individual files, glob batches, or PDF and Office page ranges. It uses create to produce cited wiki pages, notes, SVG diagrams, or CSV assets; edit, append, and delete change stored content; and lint checks citation resolution, dangling links, orphaned or stale pages, and frontmatter consistency. In local mode, writes immediately become files under wiki/, with .llmwiki/index.db and cached artifacts accelerating retrieval; hosted mode persists them in Postgres and S3. The web app exposes compiled pages, underlying sources, cross-links, a concept graph, and visualizations including SVG and Mermaid diagrams.
- A researcher can collect papers, notes, PDF highlights, and margin comments in one workspace and have Claude build a topic-oriented wiki with source citations.
- A frequent reader can save web pages or PDF passages through the Chrome extension, then use a nightly Routine to merge new material into existing pages.
- A privacy-conscious individual can run local mode, preserve the original directory structure, and keep generated knowledge as editable Markdown files.
- An organization can consolidate internal documents into a hosted knowledge layer with full-text search, citation relationships, and cross-linked pages.
- A Claude Code, Claude.ai, or Codex user can connect through MCP so the model can consult personal source material and apply the user's recorded concepts and annotations.
What are this agent's strengths and limitations?
- Local mode leaves source files in place and stores generated pages as ordinary Markdown; the
.llmwiki/index can be deleted and rebuilt with./llmwiki reindex. - VaultFS and the same MCP tool set span both SQLite-plus-filesystem local storage and Postgres-plus-S3 hosted storage.
- The Chrome extension preserves highlights and margin notes alongside captured material, allowing the wiki to reflect both the source and the reader's reactions.
- Citation backlinks, source inspection, cross-linking, graph queries, and deterministic
lintchecks provide concrete mechanisms for auditing and maintaining generated knowledge. - The documented MCP interface works with Claude.ai, Claude Code, Codex, and other MCP-compatible applications rather than requiring a single bundled chat interface.
- Local installation requires both Python 3.11+ and Node.js 20+, plus separate Python and npm dependency stacks.
- The local API is deliberately loopback-only and does not support LAN or remote binding, limiting direct multi-device access.
- Word and PowerPoint extraction requires a separate LibreOffice installation, and better OCR for difficult PDFs may require a Mistral API key.
- The documented autonomous nightly workflow depends on Claude Routines or a desktop scheduled task; equivalent native scheduling for other model platforms is not established.
- Output quality depends on the captured source material and the model's synthesis, so generated pages still benefit from citation review and
lintchecks.
How do you install or deploy this agent?
Python 3.11+ and Node.js 20+ are required. On macOS/Linux, run git clone https://github.com/lucasastorian/llmwiki.git && cd llmwiki && python -m venv .venv && source .venv/bin/activate && pip install -r api/requirements.txt -r mcp/requirements.txt && cd web && npm install && cd ... On Windows PowerShell, run git clone https://github.com/lucasastorian/llmwiki.git; cd llmwiki; python -m venv .venv; .venv\Scripts\Activate.ps1; pip install -r api/requirements.txt -r mcp/requirements.txt; cd web; npm install; cd ... LibreOffice is optional but required to extract Word and PowerPoint files. An optional MISTRAL_API_KEY enables higher-quality OCR for PDFs with complex layouts or tables.
How do you use this agent?
Start a local workspace with ./llmwiki open ~/research on macOS/Linux or python llmwiki open C:\Users\you\research on Windows. This initializes and indexes the folder, launches the API and web app, and opens http://localhost:3000. Next, run ./llmwiki mcp-config ~/research or python llmwiki mcp-config C:\Users\you\research, then paste the resulting JSON into Claude Desktop's claude_desktop_config.json or Claude Code's .claude/settings.json; each workspace requires its own MCP server entry. After connecting, instruct Claude to read the guide, ingest the sources, and begin building the wiki. Add content by placing files in the workspace, uploading through the web app, or configuring the Chrome extension's Local mode to send clips to http://localhost:8000. For ongoing maintenance, schedule a Claude Routine to find sources, clips, and highlights added since its previous run and update the relevant pages, citations, and cross-references.
How does this agent compare with similar options?
Compared with Andrej Karpathy's LLM Wiki concept, which the project explicitly cites as its inspiration, this implementation places greater emphasis on autonomous maintenance and supplies a runnable system with document ingestion, Chrome clipping, MCP operations, scheduled Claude Routines, local and hosted storage backends, and citation hygiene checks.
FAQ
Does local mode upload or rewrite my original documents?
wiki/ directory for generated pages and a rebuildable .llmwiki/ directory for the index and cache.Is an API key mandatory?
MISTRAL_API_KEY is optional and improves OCR for complex PDFs, although AI writing and scheduled maintenance still require access to Claude or another connected MCP-capable model application.Can I reach the local server from another device?
127.0.0.1 and explicitly does not support LAN or remote binding; cross-device use calls for the remote self-hosted path or hosted service.Will deleting .llmwiki/ destroy the wiki?
wiki/. The directory contains derived index and cache data and can be reconstructed with ./llmwiki reindex ~/research.