writestead

LLM Wiki

Inspired by Karpathy's LLM OS wiki concept — persistent knowledge that compounds over time. Humans curate sources, agents maintain the structure.

Writestead gives you:

• Structured wiki: frontmatter, wikilinks, page types (source, entity, concept, analysis)
• Raw ingest: add local files or URLs, extract text from PDF/DOCX/PPTX/images via liteparse
• MCP server: expose wiki tools to any MCP client
• Obsidian sync: headless Obsidian Sync via ob CLI
• Lint: detect orphans, broken links, stale logs, duplicate titles, missing frontmatter

Install

# npm (macOS, Linux, WSL)
npm i -g @ahkohd/writestead

# homebrew (macOS, Linux)
brew install ahkohd/writestead/writestead

# cargo
cargo install writestead --locked --force

# verify
writestead --version

Optional tools (install any you need, writestead doctor checks availability):

• lit — PDF/DOCX/PPTX/image text extraction
• poppler-utils — PDF utilities
• rg, fd — faster search/listing
• ob — headless Obsidian Sync

Quick start

New vault (no existing Obsidian Sync)

writestead init --vault-path ~/Documents/writestead --sync-backend obsidian
writestead doctor
writestead start

Existing Obsidian Sync vault

If you already have a vault syncing via Obsidian Sync, set up sync before init so existing files are preserved:

# 1. login and link to remote vault
ob login
ob sync-list-remote
ob sync-setup --path ~/Documents/writestead --vault <vault-id>
ob sync --path ~/Documents/writestead

# 2. init without --force (skips files that already exist)
writestead init --vault-path ~/Documents/writestead --sync-backend obsidian

# 3. start
writestead doctor
writestead start

Docker

docker run -d \
  -v writestead-vault:/vault \
  --name writestead \
  ghcr.io/ahkohd/writestead:latest

# setup sync inside container
docker exec -it writestead bash
ob login
ob sync-list-remote
ob sync-setup --path /vault --vault <vault-id>
ob sync --path /vault
writestead init --vault-path /vault --sync-backend obsidian
exit

# restart to pick up synced vault
docker restart writestead

Commands

• writestead init — create vault structure and config
• writestead start / stop / status — daemon lifecycle (start --foreground for attached mode, status --json for machine output)
• writestead doctor — health checks for vault, sync, extractors, accelerators (--json for structured output)
• writestead sync — run sync backend
• writestead help-wiki — print workflow guide and conventions

Wiki

• writestead read <path> — read wiki page with line pagination
• writestead search <query> — case-insensitive content search
• writestead edit <path> --old-text ... --new-text ... --log-action ... --log-description ... — exact-match replacement
• writestead write <path> --content-file ... --log-action ... --log-description ... — write full page
• writestead list — list wiki pages with pagination
• writestead lint — run structural lint checks
• writestead lint --fix — apply safe lint fixes
• writestead lint --fix --dry-run — preview safe lint fixes
• writestead index — read wiki/index.md

Raw

• writestead raw add <source> — add local file or URL to raw/ (--name, --force)
• writestead raw list — list raw source files with pagination
• writestead raw read <path> — extract text from raw source with pagination

Config

• writestead config path / show / get <key> / set <key> <value> / unset <key>

HTTP API

The CLI talks to a local HTTP daemon (default: http://127.0.0.1:8765).

Endpoints:

• GET /health
• GET /metrics (Prometheus format)
• POST /mcp (MCP over HTTP JSON-RPC)
• GET /mcp (returns 405)
• DELETE /mcp (terminate MCP session)

Configure bind address with config keys (host, port) or env (WRITESTEAD_HOST, WRITESTEAD_PORT).

MCP over HTTP

POST /mcp exposes the writestead MCP server. Tools are discoverable via tools/list:

Tool	Description
wiki_read	Read wiki page (1-indexed line pagination)
wiki_search	Case-insensitive content search
wiki_edit	Exact oldText/newText replacement with log
wiki_write	Write full page with log
wiki_list	List pages (0-indexed item pagination)
wiki_lint	Validate vault structure, frontmatter, links, orphans, and stale logs
wiki_index	Read wiki/index.md
wiki_sync	Run sync backend
wiki_help	Print workflow guide
raw_list	List raw source files (0-indexed pagination)
raw_read	Extract text from raw source (1-indexed line pagination)
raw_upload	Add source via url, path, or base64 content

MCP clients receive workflow instructions automatically on initialize.

MCP client setup

Local no-auth:

writestead:
  url: http://127.0.0.1:8765/mcp
  tools:
    resources: false
    prompts: false

Bearer auth:

writestead:
  url: http://127.0.0.1:8765/mcp
  headers:
    Authorization: Bearer ${WRITESTEAD_BEARER_TOKEN}
  tools:
    resources: false
    prompts: false

Configuration

Config file

Default path: ~/.config/writestead/config.json (or $XDG_CONFIG_HOME/writestead/config.json).

{
  "name": "writestead",
  "vault_path": "~/Documents/writestead",
  "host": "127.0.0.1",
  "port": 8765,
  "sync": { "backend": "obsidian" },
  "mcp": { "auth": { "mode": "none" }, "session_ttl_seconds": 3600 },
  "search": { "backend": "auto" },
  "raw": { "upload_max_bytes": 52428800, "url_timeout_seconds": 30 }
}

Config keys

• name — vault display name
• vault_path — path to vault root
• host — daemon bind address (default: 127.0.0.1)
• port — daemon port (default: 8765)
• sync.backend — obsidian | none (default: obsidian)
• mcp.auth.mode — none | bearer (default: none)
• mcp.session_ttl_seconds — session expiry (default: 3600)
• search.backend — auto | builtin | rg-fd (default: auto)
• raw.upload_max_bytes — upload size cap (default: 52428800)
• raw.url_timeout_seconds — URL download timeout (default: 30)
• raw.pdf_liteparse_max_pages — max PDF pages routed to liteparse (default: 30)
• raw.pdf_liteparse_timeout_ms — liteparse timeout (default: 60000)
• raw.pdf_liteparse_mem_limit_mb — liteparse memory cap (default: 4096)

Environment variables

• WRITESTEAD_CONFIG_FILE — config file path override
• WRITESTEAD_RUNTIME_DIR — runtime directory override
• WRITESTEAD_PID_FILE — PID file path override
• WRITESTEAD_LOG_FILE — log file path override
• WRITESTEAD_BEARER_TOKEN — bearer token (required when mcp.auth.mode=bearer)
• WRITESTEAD_MCP_AUTH_MODE — auth mode override

Bearer auth

Token is env-only. Setting mcp.auth.bearer_token in config is blocked by design.

writestead config set mcp.auth.mode bearer
export WRITESTEAD_BEARER_TOKEN='your-token'
writestead start

Raw source conventions

• raw add detects mode by prefix: http:// / https:// downloads, otherwise copies local file
• raw read supports:
  • .md / .txt / .json / .yaml / .csv / .html / .xml / .rst / .tex / .log — direct text read
  • .pdf — lit parse or pdftotext by size
  • .docx / .pptx / .xlsx — lit parse
  • images (.png / .jpg / .tiff / .webp) — lit parse with OCR
  • unknown types rejected
• raw upload (MCP) accepts exactly one of: url, path (vault-relative), or content (base64)
• raw/assets/ is excluded from listing and reading (deferred)
• PDF page windows: writestead raw read manual.pdf --page-start 1 --page-end 20

Search acceleration

When search.backend=auto (default), writestead uses rg and fd if found in PATH, falling back to built-in search. Set search.backend=rg-fd to require them.

# install (arch)
pacman -S ripgrep fd

# install (macOS)
brew install ripgrep fd

# verify
writestead doctor

Pagination

• wiki_read / raw_read / writestead read: offset is 1-indexed (line number)
• wiki_list / raw_list / writestead list: offset is 0-indexed (item index)
• All paginated responses include: offset, limit, total (or total_lines), has_more

Sync backend

• obsidian (default): runs ob sync --path <vault_path> — headless Obsidian Sync
• none: explicit no-op

Observability

GET /metrics exports Prometheus counters and gauges:

writestead_uptime_seconds
writestead_mcp_sessions_active
writestead_mcp_requests_total
writestead_mcp_tool_calls_total
writestead_mcp_tool_calls_by_tool_total{tool="..."}
writestead_mcp_tool_errors_total
writestead_mcp_tool_errors_by_tool_total{tool="..."}
writestead_raw_uploads_total
writestead_raw_upload_bytes_total
writestead_raw_reads_total
writestead_raw_reads_by_format_total{format="..."}
writestead_sync_runs_total{trigger="..."}
writestead_sync_errors_total{trigger="..."}
writestead_sync_duration_seconds_sum
writestead_sync_duration_seconds_count

Alert suggestions:

• Tool error spike: increase(writestead_mcp_tool_errors_total[5m]) > 10
• Per-tool regressions: watch writestead_mcp_tool_errors_by_tool_total{tool=...}
• Upload pressure: sustained growth in writestead_raw_upload_bytes_total

Troubleshooting

• Run writestead doctor first
• If daemon won't start, check writestead status and ~/.config/writestead/writestead.log
• If MCP auth fails, verify WRITESTEAD_BEARER_TOKEN is set and mcp.auth.mode=bearer
• If raw reads fail for PDF/DOCX, install lit (npm i -g @llamaindex/liteparse)
• For large PDFs, install poppler-utils
• If search is slow on large vaults, install rg and fd