ragmark
Local-first MCP server for retrieval over markdown wikilink vaults, offering hybrid vector+lexical search, note reading, neighbor expansion, and recent activity tracking with fully local embeddings and no network egress.
README
ragmark
Local-first retrieval over markdown / [[wikilink]] vaults — hybrid (vector + lexical)
chunk search, note-level similarity, wikilink-aware context expansion, and a first-party
MCP surface. Embeddings run locally (fastembed); the
package makes no network egress at query time. Sibling of
graphmark, which it consumes for deterministic
graph reads.
Status: seeded, pre-release. The public interfaces, security gate, correctness harness, and release machinery are in place; the engine behavior lands slice by slice against them.
Surfaces (one core, two faces)
| MCP tool | CLI | |
|---|---|---|
vault_search(query, k) |
ragmark search "query" -k 8 |
hybrid chunk hits with chunk_id citation handles |
vault_read(path) |
ragmark read path.md |
bare note text |
vault_neighbors(path, depth, token_budget) |
ragmark neighbors path |
wikilink BFS; structure always complete, budget limits content only |
recent_activity(days, limit) |
ragmark recent |
git-log-backed "what moved lately" |
| — | ragmark index [--force] |
index maintenance is CLI-only, deliberately |
Every surface routes through the same fail-closed machine-context gate; an unmarked vault reveals shared notes only.
MCP registration
claude mcp add --transport stdio vault -- ragmark mcp --vault ~/path/to/vault
The registration name vault keeps existing mcp__vault__* tool names stable. The vault
root always comes from --vault or $RAGMARK_VAULT — never inferred.
Correctness story
No parity oracle — the package improves on its predecessor in the same motion as absorbing
it. Instead: a hand-curated, executor-untouchable golden-query set asserted as recall@k
with the embedding model pinned, plus permanent safety invariants — no silent embedding
truncation (the predecessor silently dropped 27.8% of corpus characters), chunk↔vector
conservation, and context-gating tests that CI re-runs with the filter defanged and
requires to FAIL (scripts/teeth_check.py). Quality is measured, leaks are impossible to
test around, and a green gate means something.
Decision provenance: the-vault#136 (blueprint map) and the-vault#145 (build epic).
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。