llm-wiki-mcp 0.1.1

MCP server + Claude Code skills for Karpathy-style LLM wikis: persistent markdown knowledge bases your agent grows over time.

llm-wiki-mcp

Persistent markdown wiki for your AI agent, built on Karpathy's LLM wiki gist. Four MCP tools (wiki_read, wiki_write_page, wiki_log_append, wiki_inventory) plus four Claude Code skills (wiki-init, wiki-ingest, wiki-query, wiki-lint). stdio transport, local filesystem.

The server handles the boring layer LLMs keep getting wrong: atomic writes, etag conflict checks, append-only log integrity, path containment. The skills give the agent a workflow to follow. The wiki schema lives in your own wiki/CLAUDE.md and grows with your domain. There is no Layer 3 schema validation in the server.

Status: alpha (v0.1.1). Local backend only. MIT licensed.

Quick start

Requires Python 3.11+ and uv.

Pick an absolute path for the wiki folder. The server creates pages/ and log.md under it on first run if they don't exist:

uvx llm-wiki-mcp --wiki-root /absolute/path/to/wiki

Wire it into your MCP client.

Claude Code:

claude mcp add llm-wiki -- uvx llm-wiki-mcp --wiki-root /absolute/path/to/wiki

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS) or Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "llm-wiki": {
      "command": "uvx",
      "args": ["llm-wiki-mcp", "--wiki-root", "/absolute/path/to/wiki"]
    }
  }
}

Restart the client. Four tools should appear: wiki_read, wiki_write_page, wiki_log_append, wiki_inventory.

Claude Code skills

Claude Code users can install the bundled workflow skills as a plugin:

claude plugin marketplace add https://github.com/flsteven87/llm-wiki-mcp
claude plugin install llm-wiki-skills@llm-wiki-mcp

Each skill reads wiki/CLAUDE.md for the active schema on every run, so you can evolve the schema without re-installing anything. Ask the agent things like:

Skill	What to ask	Needs MCP server?
wiki-init	"Create an LLM wiki for AI safety research at ~/wikis/ai-safety."	No
wiki-ingest	"Ingest https://arxiv.org/abs/2310.12345 into the wiki."	Yes
wiki-query	"What does the wiki say about steering vectors?"	Yes
wiki-lint	"Run a wiki health check."	Yes

wiki-init is a one-shot scaffolder; the other three are Karpathy's three operations.

Other MCP clients (Claude Desktop, Cursor) get the four tools but not the skills. The agent has to derive the workflow from tool descriptions alone, which works for one-off reads and writes but tends to skip the bookkeeping (log entries, backlink audits) the skills make explicit.

The four tools

Tool	Annotations	Purpose
wiki_read	read-only, idempotent	Read one page. Returns body, parsed frontmatter, outgoing links, etag.
wiki_write_page	destructive, idempotent	Atomic create or update with etag CAS. Pass etag=null to create, the read etag to update.
wiki_log_append	not idempotent	Append one entry to log.md in Karpathy's ## [YYYY-MM-DD] op | Title format.
wiki_inventory	read-only, idempotent	Snapshot the whole graph: pages, frontmatter, link edges, log entries, plus an optional plain-text mention scan for backlink audits.

index.md and raw/ are intentionally not exposed as tools. The index is LLM-curated content edited via the host's Read/Write. The raw layer is immutable from the server's perspective.

Wiki layout

wiki-init scaffolds a project that looks like this:

your-project/
├── raw/                    Immutable source files (papers, articles, transcripts)
│   └── ...
└── wiki/                   ← --wiki-root points here
    ├── pages/              Markdown pages, one per topic
    ├── log.md              Append-only session log
    ├── index.md            LLM-curated browse page
    └── CLAUDE.md           Schema doc the LLM reads on every operation

--wiki-root points at the curated wiki/ folder, not the parent project folder containing raw/. Easy to get wrong on first install; the troubleshooting section below covers the error you'll see.

Design boundary

The server enforces mechanics, not content shape:

• Atomic writes. tmp-file + fsync + rename for pages. O_APPEND single-write for log entries.
• Optimistic concurrency. Every page has an etag (sha256(body) || mtime_ns). Updates supply the etag they read; a mismatch raises WikiConflictError, and the agent re-reads, merges, and retries.
• Path containment. Slugs are regex-validated. Resolved paths are checked against the realpath of the root, blocking the CVE-2025-53109 symlink-escape class.
• Format-locked log line. ## [YYYY-MM-DD] operation | Title. Operation names are free strings; only characters that would break the line shape are rejected.

The server does not validate frontmatter shape, page categories, or link targets. That layer lives in your wiki/CLAUDE.md schema doc and grows with the LLM. Karpathy's gist is deliberately silent on content shape; baking a schema into the server would defeat the point.

Python API

If you want to wrap the MCP server with your own storage backend (SQLite, Notion, GDrive, a test fake), implement the WikiStorage Protocol and pass an instance to build_server:

from llm_wiki_mcp import WikiStorage, PageRead, LogEntry
from llm_wiki_mcp.server import build_server

class MyStorage:  # satisfies the WikiStorage Protocol
    async def read_page(self, slug: str) -> PageRead: ...
    async def write_page(self, slug, body, expected_etag=None) -> str: ...
    async def list_pages(self) -> list[str]: ...
    async def append_log(self, entry: LogEntry) -> None: ...
    async def read_log(self) -> str: ...
    async def write_raw_file(self, name, data) -> None: ...  # usually raises

server = build_server(storage=MyStorage())
server.run()

build_server is the composition root. The CLI main() is a thin caller that constructs LocalFilesystemStorage from --wiki-root and hands it in.

The bundled Claude Code skills ship as package data under llm_wiki_mcp/skills/ and load via importlib.resources if you want to wire them into a non-Claude-Code agent. Typed domain errors (WikiConflictError, WikiNotFoundError, WikiPermissionError, WikiPathError, WikiSchemaViolationError) are importable from the package root for catching at your own boundary.

Troubleshooting

llm-wiki-mcp: command not found after uv tool install. uv puts the binary in ~/.local/bin (or %USERPROFILE%\.local\bin on Windows). Add it to PATH, or use uvx llm-wiki-mcp ... to invoke without a persistent shim.

wiki_* tools don't appear after editing the client config. Restart the MCP client. Claude Desktop, Claude Code, and Cursor only re-read mcpServers at startup.

WikiPathError: path escapes wiki root. You pointed --wiki-root at the project folder containing raw/ instead of the curated wiki/ folder inside it. /Users/me/wikis/ai-safety/wiki is correct; /Users/me/wikis/ai-safety is not.

Skills not loading in Claude Code. Run claude plugin list. If llm-wiki-skills is missing, rerun the marketplace commands in the Claude Code skills section.

Development

git clone https://github.com/flsteven87/llm-wiki-mcp
cd llm-wiki-mcp
uv sync --extra dev
uv run pytest
uv run ruff check .
uv run pyright src/llm_wiki_mcp

License

MIT. See LICENSE.