nvk/llm-wiki: Parallel Multi-Agent Research for LLM Knowledge Bases

nvk/llm-wiki: Parallel Multi-Agent Research for LLM Knowledge Bases

Developers building with large language models face a persistent friction: research scattered across browser tabs, chat threads, and ephemeral context windows, with no durable structure for an AI agent to reason over. The result is repeated work, inconsistent answers, and knowledge that evaporates when the session ends. nvk/llm-wiki, an open-source Python^{↗ Bright Coding Blog} project with 858 GitHub stars and 89 forks, addresses this by compiling LLM knowledge bases through parallel multi-agent research, source ingestion, and structured wiki compilation that any AI agent can query and extend.

What is nvk/llm-wiki?

nvk/llm-wiki is a knowledge base compiler designed specifically for AI agent consumption. Created by nvk and released under the MIT License, it transforms raw research into structured, queryable wikis through automated multi-agent workflows. The project is built in Python and actively maintained, with its last commit dated July 15, 2026.

The tool occupies a distinct niche: it is not a vector database, not a simple note-taking app, and not a RAG pipeline in the conventional sense. Instead, it is a research orchestration layer that dispatches parallel agents to investigate topics, ingest sources, compile synthesized articles, and generate artifacts—all while maintaining strict immutability of raw inputs and confidence-scored outputs. It ships as native plugins for Claude Code and OpenAI Codex, as an OpenCode instruction file, and as a portable AGENTS.md for any other LLM agent.

The project traces conceptual lineage to Andrej Karpathy's LLM wiki concept and Elisha Long's "retardmaxxing" philosophy of rapid execution over prolonged deliberation. It also incorporates prior art from rvk7895/llm-knowledge-bases in Claude Code wiki plugins. The current implementation reflects a mature v0.16.0 release with established patterns for query optimization, session memory, and cross-runtime synchronization.

Key Features

Parallel Multi-Agent Research: The core differentiator is automated research dispatch. Standard mode runs 5 parallel agents (academic, technical, applied, news, contrarian); deep mode adds 3 more (historical, adjacent fields, data/stats); and "retardmax" mode deploys 10 agents with aggressive ingestion and minimal planning overhead. Each agent searches from a distinct angle, with results synthesized into cross-referenced articles.

Thesis-Driven Investigation: Unlike open-ended research, /wiki:thesis evaluates specific claims through structured evidence gathering. It decomposes claims into testable predictions, dispatches supporting and opposing agent cohorts, and delivers verdicts (supported/partially supported/contradicted/insufficient evidence/mixed). Anti-confirmation-bias logic automatically strengthens counter-evidence hunting in multi-round sessions.

Multi-Runtime Agent Support: The architecture maintains Claude Code as the principal user while generating thin packaging layers for OpenAI Codex, OpenCode, Pi/DS4 local models, and generic agents. A shared behavior layer ensures consistency; runtime-specific wrappers handle invocation metadata. Query-lite profiles (~2.8 KB) provide read-only access without full research context loading.

Session Memory & Feedback Curation: Automated capture stores redacted checkpoints under HUB/.sessions/, enabling context rehydration across chat compactions or runtime switches. High-signal corrections and approvals become reviewable feedback candidates; generic acknowledgements are filtered out. Promotion into topic knowledge requires explicit human action.

Obsidian-Compatible Structure: Each topic wiki uses dual-linking ([[wikilinks]] for Obsidian, standard markdown^{↗ Smart Converter} links for agents and GitHub) with isolated vault configurations. The hub remains lightweight—no content, just registry and session layers.

Deterministic Integrity Checks: Local helpers provide schema adoption, linting, archive management, and structural verification without requiring LLM invocation. Agentic workflows handle editorial depth.

Use Cases

Technical Research & Due Diligence: A developer investigating "hardware wallet threat models" can run /wiki:research "hardware wallet threat models" --new-topic --deep --min-time 2h to spawn 8 agents across academic, technical, contrarian, and adjacent angles, returning a compiled wiki with confidence-scored articles and cross-references rather than a pile of bookmarks.

Content & Media Cataloging: The collector workflow finds, deduplicates, and catalogs artifacts, media, memes, or tools. Running /wiki:collect "bitcoin memes" --wiki memes-bitcoin --media reference creates a provenance map without downloading binaries; --inventory corpus tracks large unstable collections as single linked records.

Evidence-Based Writing & Output Generation: Writers can query compiled wikis for structured outputs—/wiki:output report --topic gut-brain generates reports, while /wiki:output slides --retardmax ships rough slide decks for rapid iteration. The thesis mode produces actionable playbooks with testable claims from research questions.

Repository Gap Analysis: /wiki:assess /path/to/my-app --wiki nutrition compares a codebase against wiki research and market knowledge, identifying alignment gaps and competitive positioning. This supports architectural decision-making with structured evidence rather than intuition.

Cross-Runtime Knowledge Persistence: Teams using multiple AI coding tools (Claude Code for deep work, Codex for quick tasks, local Pi/DS4 for sensitive data) maintain a single shared knowledge base with runtime-appropriate access levels—full research capabilities or read-only query modes.

Installation & Setup

Claude Code (native plugin):

claude plugin install wiki@llm-wiki

This registers the plugin with Claude Code's native marketplace system. Restart Claude Code after installation.

OpenAI Codex (marketplace plugin):

codex plugin marketplace add nvk/llm-wiki
codex plugin add wiki@llm-wiki

The first command registers the catalog; the second installs and enables the cached plugin non-interactively. Start a new Codex thread, then use @wiki or type $ to select wiki-query.

For local checkout with managed bootstrap:

./scripts/bootstrap-codex-plugin.sh --scope user --verify

OpenCode (instruction file): Add to opencode.json (project-level or ~/.config/opencode/.opencode.json for global):

{
  "instructions": ["https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md"],
  "permission": {
    "external_directory": {
      "~/.config/llm-wiki/**": "allow",
      "~/Library/Mobile Documents/com~apple~CloudDocs/wiki/**": "allow"
    }
  }
}

OpenCode fetches the URL fresh on every session start—no manual updates needed. The external_directory permission is required because the wiki hub lives outside the project directory.

Pi / DS4 (local models):

pi --skill path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md

For read-only queries with DS4's isolated state:

./scripts/pi-ds4-wiki-query

Any LLM Agent (portable):

# Read-only queries: small default
cp profiles/query-lite/SKILL.md ~/your-project/AGENTS.md

# Research and maintenance: complete protocol
cp AGENTS.md ~/your-project/AGENTS.md

Real Code Examples

Basic research and query workflow:

/wiki:research "nutrition" --new-topic
/wiki:research "gut-brain axis" --wiki nutrition
/wiki:query "How does fiber affect mood?"

The first command creates a new topic wiki and initiates parallel research. The second adds focused research to the existing nutrition wiki. The third queries compiled articles for a specific question, returning a synthesized answer with citations rather than raw search results.

Thesis-driven investigation with time budget:

/wiki:thesis "fiber reduces neuroinflammation via SCFAs"
/wiki:thesis "cold exposure upregulates BDNF" --min-time 1h

The thesis command structures evidence gathering around a falsifiable claim. The --min-time flag enables multi-round investigation with automatic anti-confirmation-bias—Round 2 intensifies counter-evidence search if Round 1 skews supporting.

Collection and inventory workflow:

/wiki:collect "bitcoin memes" --wiki memes-bitcoin
/wiki:collect "bitcoin memes" --scale medium --media reference --inventory corpus
/wiki:inventory list --view actions --limit 10

Collection creates searchable catalogs with provenance tracking. The --media reference option avoids binary downloads, storing URLs and metadata instead. Inventory commands manage durable tracking records separately from ephemeral collection outputs.

Session management and feedback promotion:

/wiki:session status
/wiki:session rehydrate --cwd "$PWD"
/wiki:feedback list --unpromoted
/wiki:feedback promote fb-abc123 --topic meta-llm-wiki

Session commands recover operational context across chat interruptions. Feedback commands surface high-signal corrections captured from session hooks, with explicit promotion required before integration into topic knowledge.

Local deterministic checks (no agent required):

./scripts/llm-wiki lint /path/to/wiki
./scripts/llm-wiki lint --fix /path/to/wiki
./scripts/llm-wiki schema adopt /path/to/wiki

These local helpers perform structural verification and safe migrations without LLM invocation, useful in CI pipelines or pre-commit hooks.

Advanced Usage & Best Practices

Hub path configuration for iCloud or non-default locations: Rather than relying on ~/wiki fallback, explicitly configure the hub path to enable cross-device synchronization:

/wiki config hub-path "~/Library/Mobile Documents/com~apple~CloudDocs/wiki"

Use portable ~ expansion; avoid absolute paths in shared wikis.json since /Users/<name>/... paths break across machines.

Query mode selection: The query-lite protocol (~2.8 KB) reduces Claude /wiki:query instructions by 72.54% compared to full research context. Use explicit $wiki-query or /wiki:query --quick for simple lookups; reserve full profiles for research, ingestion, and compilation workflows that write to the wiki.

Sandbox permissions with nono: If running agents under sandbox wrappers, ensure both nono policies and OpenCode external_directory permissions allow the same paths. Codex specifically needs read-write access to $HOME/.codex for plugin state management. Use nono why --path <path> --op read to diagnose Seatbelt denials.

Version synchronization: The project uses generated packaging targets maintained through sync scripts (./scripts/sync-codex-plugin.sh, ./scripts/sync-opencode-plugin.sh). Drift is caught by deterministic tests. When upgrading, prefer GitHub CLI with HTTPS transport to avoid SSH known_hosts complications in sandboxed environments.

Comparison with Alternatives

The primary trade-off is complexity: llm-wiki's rich feature set requires understanding its hub/topic architecture and command taxonomy. For simple personal note-taking, lighter tools may suffice. For teams producing validated, multi-source research compendia with AI agent collaboration, the structure pays dividends.

FAQ

What Python version is required? v0.11.1 added compatibility for Python 3.9 and macOS system Python; newer runtimes are also supported.

Is the wiki format locked to any specific tool? No. Dual-linking ensures compatibility with Obsidian, Claude Code, Codex, GitHub markdown rendering, and plain text editors.

Can I use this without Claude Code? Yes. Native plugins exist for Codex; instruction files work with OpenCode and Pi/DS4; portable AGENTS.md supports any LLM agent with file access.

What license covers the project? MIT License. Copyright (c) 2026 nvk.

How does session capture affect privacy? Trusted hooks record harness metadata, directory context, and redacted events—not full transcripts by default. Opt out with /wiki:session disable.

What happens to archived topics? Archived wikis move to topics/.archive/ and are excluded from default queries, compilation, and research. Use --include-archived to search them explicitly.

How do I verify my installed version? For Claude Code, check /wiki status output or ~/.claude/plugins/installed_plugins.json. For Codex, run ./scripts/verify-codex-plugin.sh --scope user.

Conclusion

nvk/llm-wiki fills a specific gap in the AI tooling landscape: the translation from ephemeral research to structured, queryable, agent-compatible knowledge. Its 858-star GitHub presence reflects genuine utility for developers and researchers who need durable compendia rather than scattered browser tabs and chat fragments.

The tool rewards users who invest in understanding its architecture—hub configuration, topic isolation, query depths, and session management. For solo developers conducting deep technical research, teams building evidence-based documentation, or organizations standardizing knowledge across multiple AI runtimes, the parallel agent orchestration and structured compilation offer concrete productivity gains.

The project is actively maintained, MIT-licensed, and designed for extensibility. Start with the GitHub repository and the Claude Code or Codex plugin installation, or deploy the portable AGENTS.md if your workflow uses a different agent runtime.