Show HN: WUPHF — Karpathy-Style LLM Wiki with Markdown + Git as Source of Truth

WUPHF is an open-source collaborative office for AI agents (Claude Code, Codex, OpenClaw, local LLMs via OpenCode) that includes a Karpathy-style wiki layer. The wiki uses Markdown and Git as the source of truth, stored at ~/.wuphf/wiki/, with a bleve (BM25) + SQLite index on top. No vector or graph DB is used yet — the goal is to see how far Markdown + Git can go before adding heavier infrastructure.
Key Features
- Each agent gets a private notebook at
agents/{slug}/notebook/plus shared team wiki atteam/. - Draft-to-wiki promotion flow: notebook entries are reviewed (by agent or human) and promoted to canonical wiki with back-links. A state machine handles expiry and auto-archive.
- Per-entity fact log: append-only JSONL at
team/entities/{kind}-{slug}.facts.jsonl. A synthesis worker rebuilds entity briefs every N facts. - Commits are attributed to a distinct Git identity ("Pam the Archivist") for provenance via
git log. - [[Wikilinks]] with broken-link detection (rendered in red).
- Daily lint cron for contradictions, stale entries, and broken wikilinks.
/lookupslash command + MCP tool for cited retrieval. Heuristic classifier routes short queries to BM25 and narrative queries to a cited-answer loop.
Retrieval Tuning
Current benchmark with 500 artifacts and 50 queries achieves 85% recall@20 on BM25 alone, which is the internal ship gate. If a query class drops below that, sqlite-vec is the pre-committed fallback.
Substrate Choices
- Markdown for durability — the wiki outlives the runtime; users can
git cloneand walk away with every byte. - Bleve for BM25.
- SQLite for structured metadata (facts, entities, edges, redirects, supersedes).
- Canonical IDs are first-class: fact IDs are deterministic (include sentence offset), slugs are assigned once and never renamed (redirect stubs used). Rebuild is logically identical, not byte-identical.
Known Limits
- 85% recall is not a universal guarantee — tuning ongoing.
- Synthesis quality depends on agent observation quality. The lint pass helps but is not a judgment engine.
- Single-office scope; no cross-office federation yet.
Demo & Install
A 5-minute terminal walkthrough is available at asciinema (script at ./scripts/demo-entity-synthesis.sh).
Install with: npx wuphf@latest
Build from source: git clone https://github.com/nex-crm/wuphf.git; go build -o wuphf ./cmd/wuphf
The wiki ships as part of WUPHF but can be used standalone. MIT license, self-hosted, bring-your-own keys.
📖 Read the full source: HN LLM Tools
👀 See Also

Claude Sessions: Lightweight Desktop App for Browsing Claude Code History
Claude Sessions is a new desktop application that lets developers browse their Claude Code session history locally. It reads from ~/.claude/projects, organizes sessions by project, handles large sessions up to 500k+ tokens without lag, and includes search functionality and keyboard navigation.

Distilled Qwen 3.5 27B Model Shows Strong Performance with Cursor AI Coding Agent
A user reports that the opus 4.6 distilled version of Qwen 27B works effectively as the model driving Cursor, with performance comparable to Gemini 3 Flash. Setup took about 10 minutes using Cursor to configure ngrok tunnel and localllama.

Open-source methodology for agentic AI partnership with Claude
A developer has published a 25,000-word paper and open-sourced templates for building a persistent partnership system with Claude that uses shared memory across sessions, cognitive monitoring, and multi-AI consultation.

MCP-India-Stack: Offline-first server for Indian financial data in AI agents
MCP-India-Stack is an offline-first MCP server that provides Indian financial and government API functionality without authentication or external API calls. It bundles datasets locally for tax calculations, validation tools, and lookups.