hive-vault 1.1.0

Unified MCP server for AI-assisted development — on-demand Obsidian vault access + worker delegation

hive-vault

Unified MCP server for AI-assisted development — on-demand Obsidian vault access + worker delegation to local/cloud models.

The Problem

AI coding assistants load context statically. A typical CLAUDE.md grows to 800+ lines of standards, patterns, and project knowledge. Most of it is irrelevant to the current task. Every session pays the full token cost.

The Solution

Hive replaces static context with on-demand vault queries via the Model Context Protocol. Your knowledge stays in an Obsidian vault. Your AI assistant loads only what it needs, when it needs it.

For tasks that don't need your primary model's full reasoning, Hive delegates to cheaper models (local Ollama or cloud OpenRouter) with automatic routing and budget controls.

Works with any MCP client: Claude Code, Codex CLI, Gemini CLI, Cursor, Windsurf, and others.

Measured result: 67–82% token reduction on targeted queries vs static context loading.

Quick Start

Claude Code:

claude mcp add hive -- uvx hive-vault

Other MCP clients: point your client's MCP config at uvx hive-vault via stdio transport. See your client's docs for the exact syntax.

To configure the vault path (defaults to ~/Projects/knowledge):

# Claude Code example
claude mcp add hive -e VAULT_PATH=/path/to/your/vault -- uvx hive-vault

Tools

Vault Tools (11)

Tool	Description
vault_list_projects	List all projects in the Obsidian vault
vault_query	Read sections or files on demand (supports shortcuts: context, tasks, roadmap, lessons)
vault_search	Full-text search across the vault with metadata filters (type, status, tag)
vault_health	Health metrics for all vault projects (file counts, staleness, coverage)
vault_update	Write to vault with YAML frontmatter validation + auto git commit
vault_create	Create new files with auto-generated frontmatter + auto git commit
vault_summarize	Smart summarization — returns small files directly, delegates large ones
vault_smart_search	Ranked search with relevance scoring (status weight + recency + match density)
session_briefing	One-call context briefing: tasks + lessons + git log + health
vault_recent	Files changed in the vault in the last N days (git + frontmatter)
vault_usage	Tool usage analytics — call counts, token estimates, breakdowns

Worker Tools (3)

Tool	Description
delegate_task	Route tasks to cheaper models with automatic tier selection
list_models	List available models across all providers
worker_status	Worker health: budget remaining, connectivity, usage stats

Resources

URI	Description
hive://projects	All vault projects with file counts and available shortcuts
hive://health	Vault health metrics for all projects
hive://projects/{project}/context	Project context document (00-context.md)
hive://projects/{project}/tasks	Project task backlog (11-tasks.md)
hive://projects/{project}/lessons	Project lessons learned (90-lessons.md)

Prompts

Prompt	Description
retrospective	End-of-session review — extracts lessons and appends to vault
delegate	Structured protocol for delegating tasks to cheaper models
vault_sync	Post-sprint vault synchronization — reconcile docs with shipped code
benchmark	Estimate token savings from hive MCP tools in the current session

Configuration

Variable	Default	Description
VAULT_PATH	~/Projects/knowledge	Path to your Obsidian vault
HIVE_OLLAMA_ENDPOINT	http://localhost:11434	Ollama API endpoint
HIVE_OLLAMA_MODEL	qwen2.5-coder:7b	Default Ollama model
HIVE_OPENROUTER_API_KEY	—	OpenRouter API key (also reads OPENROUTER_API_KEY)
HIVE_OPENROUTER_MODEL	qwen/qwen3-coder:free	Default OpenRouter model
HIVE_OPENROUTER_BUDGET	5.0	Monthly budget cap in USD
HIVE_DB_PATH	~/.local/share/hive/worker.db	SQLite database for budget tracking
HIVE_RELEVANCE_DB_PATH	~/.local/share/hive/relevance.db	SQLite database for adaptive context scoring

Architecture

MCP Host (Claude Code, Codex CLI, Cursor, ...)
    └── hive (MCP server, stdio)
            ├── Vault Tools ──── Obsidian vault (~/Projects/knowledge/)
            │     query, search, update, create, summarize,
            │     smart_search, briefing, recent, usage, health
            │
            └── Worker Tools ─── delegate_task → routing:
                  list_models        1. Ollama (local, free)
                  worker_status      2. OpenRouter free tier
                                     3. OpenRouter paid ($5/mo cap)
                                     4. Reject → host handles it

Maximizing Hive with CLAUDE.md

MCP servers don't activate themselves — your AI assistant needs guidance on when and how to use each tool. The CLAUDE.md file (or equivalent in your MCP client) is the key lever.

Add instructions like these to your project's CLAUDE.md:

## Vault & Knowledge (Hive MCP)

When hive-vault MCP is available, use it for on-demand context:
- `vault_query(project="myproject", section="context")` — project overview
- `vault_query(project="myproject", section="tasks")` — active backlog
- `vault_search(query="...")` — cross-vault search
- `session_briefing(project="myproject")` — full context in one call

When writing to the vault: lessons → `90-lessons.md`, decisions → `30-architecture/`.

Without these instructions, your assistant might use Hive, but inconsistently. With them, it uses Hive predictably for every relevant query.

How it works: Your MCP client loads all available tools at session start. The assistant sees tool names and descriptions, but CLAUDE.md instructions tell it which tools to prefer for which situations. Multiple MCP servers coexist — they don't compete. Each serves its domain, and your instructions guide the routing.

Worker Routing

Tasks are routed through a tiered system that minimizes cost:

1. Ollama (local) — Free. Runs on homelab hardware. Best for trivial tasks.
2. OpenRouter free — Free tier models (Qwen3 Coder 480B). Real code work.
3. OpenRouter paid — DeepSeek V3.2 at $0.28/1M tokens. Only when max_cost_per_request > 0 and monthly budget allows.
4. Reject — Returns error so the host handles the task directly.

Vault Structure

Hive expects an Obsidian vault with this layout:

~/Projects/knowledge/          # vault root (configurable via VAULT_PATH)
├── 00_meta/
│   └── patterns/              # cross-project patterns
├── 10_projects/
│   ├── my-project/
│   │   ├── 00-context.md      # section shortcut: "context"
│   │   ├── 10-roadmap.md      # section shortcut: "roadmap"
│   │   ├── 11-tasks.md        # section shortcut: "tasks"
│   │   ├── 90-lessons.md      # section shortcut: "lessons"
│   │   └── 30-architecture/   # arbitrary paths
│   └── another-project/
└── ...

Development

See CONTRIBUTING.md for setup, code standards, and PR workflow.

git clone https://github.com/mlorentedev/hive.git
cd hive
make install   # create venv + install deps
make check     # lint + typecheck + test

License

MIT