2026.10.10
pi-llm-wiki: Turning Raw Sources into a Maintainable Knowledge Base
pi-llm-wiki connects source capture, ingestion, search, linting, and Obsidian-compatible storage for Pi and MCP workflows. It suits developers and researchers who want durable context, but deployment still requires decisions about hosts, models, data governance, and automation.
What it does
pi-llm-wiki is a self-maintaining knowledge-base extension for pi. Its goal is to turn one-off source research into a wiki that can grow over time. The project describes its approach as an implementation of Andrej Karpathy’s LLM Wiki pattern and says the resulting vault is Obsidian-compatible. Raw inputs are retained, while source pages, canonical knowledge pages, and generated metadata support later retrieval and synthesis.
Image source: pi-llm-wiki. Published by the project; the interface may vary by version.
Who it suits
The project is aimed at developers, researchers, and technical writers who use Pi or MCP-capable tools and want a durable personal knowledge base. Existing Obsidian users can work with a compatible vault. Users of Claude Code, Cursor, Windsurf, Zed, Cline, and other MCP clients can access the same knowledge model through the packaged stdio MCP server. The README also describes layered recall across personal and project vaults, which may suit people who want selected knowledge to follow them between projects.
Documented capabilities
Source capture supports URLs, local files, PDFs, Markdown, text, HTML, XML, JSON, and pasted text. Batch ingestion processes uncaptured source packets into knowledge pages. Search operates through the generated registry, while linting checks such as orphan pages, broken links, duplicate aliases, and coverage gaps help maintain the vault. Status views, dashboards, activity logs, metadata rebuilds, and scheduled-update helpers are also documented.
Open Knowledge Format (OKF) v0.2 is a central part of the project. It can create documents with canonical frontmatter, standard Markdown links, and stable source citations. It can also read legacy and OKF pages without forcing an automatic migration. The official site highlights Obsidian integration as well. This portable, text-based structure may appeal to users wary of application lock-in, but the practical interoperability of the wider OKF ecosystem is not established by these sources alone.
An optional agent-trajectory feature records how a task was solved and can distill that record into reusable skills or cases. The README states that these tools are disabled by default. The v0.12.5 release lists Pi tools, MCP tools, slash-command templates, and vault templates, and adds an opt-out for automatic recall.
Limits and deployment considerations
The first consideration is the host environment. The README documents separate integration paths for Pi, oh-my-pi, and MCP clients, so users should verify that their chosen client can load the relevant extension or stdio server. The supplied sources do not establish operating-system support, hardware requirements, model-provider compatibility, or performance figures, so those should not be assumed.
Knowledge quality still depends on the sources, model processing, and human review. Automated ingestion, summarization, linking, and linting can reduce maintenance work, but they are not a substitute for fact-checking. For sensitive material, operators should define access controls, source retention, model-routing, and backup policies. The project documents guardrails against direct edits to raw sources and generated metadata, but that should not be treated as a complete security guarantee.
Before deployment, decide how vaults are scoped, how personal and project knowledge are separated, whether trajectory capture is appropriate, and how scheduled updates will be managed. GitHub metadata identifies the project as MIT-licensed. Select a release only after validating it in the intended environment and workflow.
Sources
Sources checked: 2026-10-10