Turn any codebase into a living, LLM-maintained wiki — with architecture diagrams, cross-linked services, and automatic staleness detection.
You can't keep 30 projects in your head. Neither can your AI coding assistant.
Docs rot the moment you write them. llmwiki scans a project and generates a persistent, LLM-maintained markdown wiki — architecture diagrams, service maps, integration maps, cross-links — then flags entries when the code drifts away from the docs. Plain markdown, no database, no SaaS. Inspired by Karpathy's LLM Wiki pattern.
Your AI agents read it directly over MCP — llmwiki mcp serves every project's extracted domain, services, and flows to Claude Code, Cursor, and any MCP client, with no LLM call and no re-explaining the repo each session.
# 1. Install (macOS / Linux)
curl -fsSL https://raw.githubusercontent.com/emgiezet/llmwiki/main/install.sh | sh
# 2. Configure once (interactive)
llmwiki setup
# 3. Generate a wiki for any project
llmwiki ingest ~/workspace/my-apimy-api now has a structured wiki entry with diagrams and cross-links.
Prefer a package manager or container? npx llmwiki ingest . · brew install emgiezet/tap/llmwiki · docker run --rm ghcr.io/emgiezet/llmwiki:latest version · go install github.com/emgiezet/llmwiki@latest. All install options (pinned versions, pre-built binaries, Docker usage) → docs/installation.md.
One ingest turns a repo into a structured markdown file — domain & architecture, a service map, API docs, an integration map, auto-generated tags, and Mermaid diagrams that render right here on GitHub:
flowchart LR
GW[API Gateway] --> BILL[billing-api]
BILL --> DB[(Postgres)]
BILL --> Q[[Kafka]]
Multi-client setups also get executive summaries with C4 diagrams, and every file is cross-linked into a navigable knowledge graph. See how the pipeline works → docs/architecture.md. Want to see real output first? examples/llmwiki.md is what ingest produced on this very repo.
Under NDA? Keep your default backend on the Claude Code subscription and override just the secret project to a local Ollama model — that project's code never leaves your machine. No cloud calls, no NDA risk. → NDA / local-LLM recipe
- MCP server for agents — agents query the extracted wiki over stdio (
llmwiki mcp), filtered by client/project, with no LLM call. - AI-coding integration — inject Domain/Architecture/Services/Flows straight into
CLAUDE.md. - Automatic service detection — reads
docker-compose.yml+ code indicators, one wiki file per service. - Mermaid diagrams — architecture flowcharts, ERDs, and C4 landscapes; render in GitHub/GitLab/Obsidian.
- Cross-file linking — service mentions become clickable links across the knowledge graph.
- Knowledge layers — company, department, and client knowledge that belongs to no single repo, consulted in priority order and attributed by layer. Share a layer with your team as a git submodule. → Knowledge layers
- Incremental refinement — re-running
ingestrefines the previous entry instead of starting over. - Change tracking & freshness — knows which source files each entry describes and flags drift (
llmwiki check). - Docs alongside code — write wikis into the repo (
output_mode: local|both) so one PR shows code + doc. - Three LLM backends — Claude Code subscription, Claude API, or local Ollama.
- Sovereign / local-first — run fully offline on a local Ollama model; code never leaves the box.
- Not just code — build wikis from notes, research, and articles via document extraction (PDF/DOCX/ODT/EPUB).
- Client & project indexes — executive summaries across all of a client's projects.
Most "AI docs" tools either host your code on a server or generate a static site you still have to keep current by hand. llmwiki is local-first markdown that an LLM keeps in sync and that your agents can query directly.
| llmwiki | Hosted code-wiki (e.g. DeepWiki) | Static-site generator (mkdocs / Docusaurus) | Hand-written CLAUDE.md |
|
|---|---|---|---|---|
| Storage | Plain local markdown, no SaaS | Vendor-hosted | Local, but you write it | Local, you write it |
| Keeps itself current | LLM re-ingest + drift detection | Re-index on push | Manual | Manual |
| Drift / staleness flags | ✅ (check, git co-change) |
— | — | — |
| Multi-project / multi-client | ✅ index + C4 diagrams | per-repo | per-repo | per-repo |
| Agent access without an LLM call | ✅ MCP server | varies | — | reads the file |
| Runs fully offline (local LLM) | ✅ Ollama, NDA-safe | ❌ | n/a | n/a |
| Cost | Your own LLM / subscription | Subscription | Free | Free |
Different problem from context-compression layers (e.g. headroom): those shrink tokens in flight; llmwiki gives you and your agents persistent, self-maintaining knowledge of every repo. They compose well.
| Guide | What's inside |
|---|---|
| Installation | Download, one-liner installer, Go install, updating, releases |
| Configuration | Global / client / project config, presets, non-code projects & document extraction, NDA local-LLM recipe |
| Commands | Full command reference, wiki layout, freshness tracking, CLAUDE.md injection |
| Memory | graymatter persistent memory, modes, seeding, absorb queue |
| Integrations | Supported AI tools & session hooks, Obsidian, NanoClaw |
| Architecture | How the scan → generate → write pipeline works |
Consultants juggling many client codebases, tech leads who need docs that match the code, and developers tired of re-explaining project structure to their AI assistant every session.
Path-traversal rejection, scrubbed LLM prompt/response pipeline, loopback-only Ollama default, bounded subprocess/HTTP deadlines. See SECURITY.md and the threat model; the CI gate lives in .github/workflows/security.yml.
MIT