Hive

What mlorentedev/hive MCP server does

The mlorentedev/hive MCP server gives an MCP client access to an Obsidian vault stored as a directory of Markdown files. It supports targeted reads rather than loading the entire vault into every session. Agents can retrieve project context, tasks, roadmaps, lessons, or a specific file with vault_query; search content with metadata filters, regular expressions, ranked results, recent-change ordering, or lesson-reinforcement ranking; and browse files with glob filters.

The server also provides write operations. vault_write creates, appends to, or replaces files, while vault_patch performs a focused find-and-replace. Both can defer the git commit, allowing several changes to be grouped before vault_commit creates one commit. capture_lesson records lessons, extracts multiple lessons from text, or finds existing lessons by keyword. session_briefing combines tasks, lessons, git history, and health information in one request.

How it works

The server runs over MCP using stdio and reads the configured vault directly from disk. The default vault location is ~/Projects/knowledge; VAULT_PATH or HIVE_VAULT_PATH can select another directory. A structured layout using directories such as 00_meta, 10_projects, 50_work, and 80_agents is supported but not required. Without that structure, vault operations continue with flat scope routing.

Lesson reads through vault_query, vault_search, or capture_lesson(find=...) update reinforcement data. This data is stored in a SQLite side-table, with the default path at ~/.local/share/hive/lesson_reinforcement.db. Search can use reinforcement, confidence, or a hybrid ranking mode.

Optional worker features route tasks or vault-file summarization through Ollama first, then OpenRouter according to available connectivity and budget. worker_status reports budget, connectivity, and available models. Ollama can provide a local worker, while OpenRouter requires an API key.

Setup and configuration

Install the mlorentedev/hive MCP server with Python 3.12 or newer and the hive-vault package through uvx. The server can start before a vault is configured; vault tools return an explanatory error until a path is available. A git repository inside the vault is recommended for commit support, but writing still works without git and skips per-write commits.

Obsidian is optional for authoring the files manually. The obsidian-git plugin can perform periodic commits, while calls using commit=False keep file persistence separate from commit timing. A backup git remote can protect vault history from disk loss. Daemon mode is also available through the hive command for long-lived ownership of the vault and concurrent sessions, but the standard uvx --upgrade hive-vault process is the basic setup.

Tools and capabilities

• Query project context or individual files.
• Search Markdown content with filters, regex, and ranking options.
• List projects and files with glob matching.
• Inspect identity, health metrics, drift, usage, and worker information.
• Create, append, replace, patch, and commit vault content.
• Capture, extract, and reinforce reusable lessons.
• Delegate tasks or summarize vault files with optional workers.

Limitations and notes

The mlorentedev/hive MCP server requires Python 3.12 or later and a Markdown-file directory for useful vault operations. Git commits depend on git being initialized in the vault. Deferred commits persist file changes immediately, but a crash before the next commit can lose the commit itself. A cancelled tool call may have changed the disk before the cancellation acknowledgement arrives, so state should be checked with vault_query before retrying.

Ollama and OpenRouter workers are optional rather than prerequisites. OpenRouter use requires the user's OPENROUTER_API_KEY; the baseline vault tools do not require an external service. The server reports health and worker details through its MCP tools, and it includes five resources and four prompts for guided workflows according to the project documentation.