The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the SeekLink listing page.
Status: paused. SeekLink is no longer actively maintained. For an actively-developed local Markdown search engine in the same space — local hybrid keyword + vector search, line-anchored results, CLI + MCP, multilingual — we recommend qmd.
SeekLink is a local semantic search CLI and optional read-only MCP stdio server
for Markdown vaults. It indexes a folder of .md files, searches with hybrid
keyword + vector retrieval, and returns line-anchored results that humans and
agents can read with simple shell commands.
It is built for personal knowledge bases, Obsidian-compatible vaults, bilingual
English/Chinese notes, and local agent workflows. MCP clients such as Claude
Code, Cursor, and VS Code can call the same read-only search/get/status/doctor
surface through seeklink[mcp]. It is also a useful search layer for Markdown
wiki patterns such as Andrej Karpathy's
llm-wiki:
an agent can search existing pages, read precise line windows, then update the
wiki without sending the vault to a hosted service.
Everything runs locally. No API key. No cloud search service. No Obsidian plugin required.
For Apple Silicon reranking support, install the optional MLX extra:
For Model Context Protocol (MCP) clients such as Claude Code, Cursor, or VS Code, install the optional MCP extra:
SeekLink requires Python's sqlite3 module to be linked against SQLite
3.45 or newer with FTS5 enabled. seeklink status --vault PATH checks this and
prints a clear error if the runtime SQLite is too old.
Daily use is simpler if you set a default vault:
seeklink search and single-file seeklink index path/to/file.md use a
resident daemon when --vault is not passed. The daemon keeps the embedder and
optional reranker warm in memory; on macOS this appears as a local Python
process. It is local-only, uses a Unix socket, and does not open a network port
or call a cloud service. By default it exits after 15 minutes of inactivity.
Full-vault seeklink index runs in-process so progress stays on stderr and the
final Done: summary stays on stdout. seeklink status and seeklink get
always stay cold-start: status only reads SQLite metadata, and get reads the
file directly from disk. Use --no-daemon, SEEKLINK_NO_DAEMON=1, or an
explicit --vault PATH when a script needs a one-shot cold-start path.
MCP users follow the same first step: build the index with
seeklink index --vault PATH before registering the MCP server.
Text search output is stable:
PATH is relative to the vault root.LINE is 1-indexed and points to the best matching chunk in the current file.0 for success, including no results; 1 for runtime
vault/config/file errors detected by SeekLink; and 2 for command-line usage
errors from argument parsing.Use JSON when an agent needs structured output:
Options:
Read a precise file window without using the database or daemon:
-l/--lines prints lines starting at LINE. -C/--context prints lines before
and after LINE, grep-style. Path escapes such as ../.. are rejected.
Status reports index counts, model names, index-configuration compatibility, SQLite WAL status, and freshness warnings. It does not load the embedding or reranking models.
Doctor checks Python, SQLite, the local database, index compatibility, daemon state, and optional MLX availability. It does not download or load models, but may initialize the local SeekLink database/schema if missing.
The optional Model Context Protocol (MCP) adapter lets agent clients discover and call SeekLink's read-only tools directly. The CLI keeps working independently; MCP is another surface for the same retrieval path, not a replacement.
Install it with seeklink[mcp]. Build the index with the CLI first:
seeklink index --vault PATH. The MCP adapter is read-only and exposes four
tools: search, get, status, and doctor. It does not expose index,
write notes, use HTTP/OAuth, or route through the Unix-socket daemon. Run one
MCP server per vault. search keeps its text summary compact with paths and
line anchors; result previews stay in structured content for agents that need
them. status and doctor may initialize or migrate the local SeekLink schema
when an existing .seeklink/seeklink.db needs it, but they do not index or
modify Markdown notes. If your MCP client does not inherit your shell PATH,
use the absolute path from which seeklink in the examples below.
Claude Code:
Cursor .cursor/mcp.json:
VS Code .vscode/mcp.json:
Full-vault indexing skips unchanged files by content hash unless the stored index was built with a different embedder, vector dimension, or chunker configuration, in which case SeekLink rebuilds the derived index contents. Single-file indexing updates one Markdown file only when the existing index configuration is compatible.
You normally do not need to start the daemon manually. search and single-file
index auto-spawn and auto-restart it when appropriate, then it exits after
SEEKLINK_DAEMON_IDLE_TIMEOUT seconds of inactivity. The default is 900 seconds
(15 minutes); set it to 0, off, false, or no to keep the daemon warm
until stopped.
Full-vault index still runs in-process for progress output. Passing --vault
to search or single-file index forces a one-shot cold-start path because the
daemon is bound to one vault at startup. --no-daemon and
SEEKLINK_NO_DAEMON=1 also force the same cold-start path. Use
seeklink daemon status to inspect the warm process and seeklink daemon stop
to release its memory immediately.
SeekLink fuses four channels with Reciprocal Rank Fusion:
| Channel | Purpose |
|---|---|
| BM25 / FTS5 | Exact words, code terms, acronyms, CJK lexical matches |
| Vector search | Semantic matches across different wording |
| Title / aliases / headings | Exact note and section lookup |
| Wikilink indegree | Small graph-quality prior from existing [[links]] |
The default embedder is jinaai/jina-embeddings-v2-base-zh through
fastembed. CJK full-text search uses a jieba FTS5 tokenizer when the local
Python/SQLite build can safely register it; otherwise SeekLink falls back to
SQLite's built-in trigram tokenizer instead of crashing.
The default vector dimension is 768. Advanced custom-embedder experiments can
set SEEKLINK_EMBEDDING_DIM, but it must match the embedder output and requires
a full seeklink index rebuild.
On Apple Silicon, SeekLink can rerank candidates with
mlx-community/Qwen3-Reranker-0.6B-mxfp8 when installed with seeklink[mlx].
Reranking is local and optional; if MLX is unavailable, SeekLink falls back to
first-stage hybrid RRF ranking. Use --no-rerank for one query or set
SEEKLINK_RERANKER_MODEL="" to disable it globally.
Markdown frontmatter is optional. When present, SeekLink uses it for tags and aliases:
tags support filtered search: seeklink search "memory" --tags aialiases are indexed for search and used when resolving wikilinksSeekLink writes one SQLite database inside the vault:
The database contains source metadata, chunks, FTS5 tables, sqlite-vec vectors,
and a wikilink graph. Delete .seeklink/ and run seeklink index to rebuild.
| Area | Status |
|---|---|
| Python | 3.11, 3.12, 3.13, 3.14 |
| SQLite | Python sqlite3 linked against SQLite 3.45+ with FTS5 |
| OS | macOS and Linux |
| Windows | Not supported as a first-class path |
| File format | Markdown .md |
| Vault style | Plain folder or Obsidian-compatible vault |
| CJK | Native path via jieba, with trigram fallback on static SQLite builds |
| Reranker | Optional seeklink[mlx] extra on Apple Silicon; disabled elsewhere |
| Daemon | Single vault per machine |
| MCP | Optional seeklink[mcp] stdio adapter, one server per vault |
Agents can use SeekLink through ordinary subprocess calls:
MCP clients can use the optional read-only adapter:
To make an agent choose SeekLink for a Markdown vault, add this to the
project's AGENTS.md, CLAUDE.md, or editor rules:
For hot loops, the daemon exposes a length-prefixed JSON protocol over the Unix
socket at ~/.rhizome/seeklink.sock. Most agents should prefer the CLI JSON
surface unless they specifically need socket-level latency.
See llms.txt for the compact agent contract.
Search-quality tests live in tests/blind/; the method is documented in
docs/blind-test.md. Release claims should be backed by
the bundled fixture queries or by clearly labeled private-vault measurements.
Keep runtime dependencies small, keep public docs user-facing, and add a
CHANGELOG.md entry for user-visible changes.
MIT