hive-vault

Your AI coding assistant forgets everything between sessions. Hive fixes that.

Every session, your assistant loads 800+ lines of static context. Most of it is irrelevant. You pay the full token cost every time. And next session? It starts from zero again.

Hive is an MCP server that connects your AI assistant to an Obsidian vault. Instead of loading everything upfront, it queries only what's needed — architecture decisions, lessons learned, project context — all on demand via MCP.

The numbers:

Measured on a real vault with 19 projects, 200+ files. See benchmarks.

Install (30 seconds)

One command. No cloning, no venv, no config files. Use user scope (-s user) so Hive works across all your projects — that's where cross-project knowledge shines.

Claude Code:

claude mcp add -s user hive -- uvx --upgrade hive-vault

Gemini CLI:

gemini mcp add -s user hive-vault uvx -- --upgrade hive-vault

OpenAI Codex CLI — add to ~/.codex/config.toml:

[mcp_servers.hive-vault]
command = "uvx"
args = ["--upgrade", "hive-vault"]

GitHub Copilot (VS Code) — add to .vscode/mcp.json:

{
  "servers": {
    "hive-vault": {
      "command": "uvx",
      "args": ["--upgrade", "hive-vault"]
    }
  }
}

Other MCP clients (Cursor, Windsurf, etc.): point your client at uvx --upgrade hive-vault via stdio transport.

Then ask your assistant:

"Use vault_list_projects to see my vault"

That's it. You're running.

What You Get

16 Vault Tools — your knowledge, on demand

3 Worker Tools — delegate to cheaper models

Routing: Ollama first (free) → OpenRouter free tier → OpenRouter paid ($1/mo cap) → reject.

Your primary model handles architecture. Cheaper models handle boilerplate. extract_lessons uses workers to batch-extract lessons from session notes — saving your primary model's tokens.

Before / After

Before Hive — static CLAUDE.md:

# My Project
## Architecture
[200 lines of decisions you made 3 months ago]
## Standards
[150 lines of coding patterns]
## Lessons
[100 lines of past bugs]
## Tasks
[50 lines of backlog]
# ...loaded every single session, whether relevant or not

With Hive — dynamic, on demand:

# Only when the assistant needs architecture context:
vault_query(project="my-project", section="context")

# Only when searching for a past decision:
vault_search(query="database migration strategy")

# Start of session — just the essentials:
session_briefing(project="my-project")

Configure Your Vault

Default vault path: ~/Projects/knowledge. To change it:

# Claude Code
claude mcp add -s user hive -e VAULT_PATH=/path/to/vault -- uvx --upgrade hive-vault

# Gemini CLI
gemini mcp add -s user -e VAULT_PATH=/path/to/vault hive-vault uvx -- --upgrade hive-vault

Enable Worker Delegation (optional)

claude mcp add -s user hive \
  -e VAULT_PATH=/path/to/vault \
  -e HIVE_OLLAMA_ENDPOINT=http://your-ollama:11434 \
  -e OPENROUTER_API_KEY=sk-or-... \
  -- uvx --upgrade hive-vault

All Configuration

See full configuration reference for all 15 environment variables.

Recommended Workflow

The highest-value setup combines three tools:

1. Obsidian — local-first knowledge base with 1M+ community, Markdown native, no lock-in
2. Obsidian Git — auto-commits your vault changes on a schedule (version history for free)
3. Hive — bridges your vault to any AI coding assistant via MCP

Your assistant writes lessons and decisions to the vault → Obsidian Git auto-commits → next session, everything is there. No manual sync. No context lost.

Hive works with any directory of Markdown files — Obsidian is recommended, not required.

Vault Structure

For best results, follow this layout:

~/Projects/knowledge/
├── 00_meta/patterns/          # cross-project patterns
├── 10_projects/
│   ├── my-project/
│   │   ├── 00-context.md      # vault_query section="context"
│   │   ├── 10-roadmap.md      # vault_query section="roadmap"
│   │   ├── 11-tasks.md        # vault_query section="tasks"
│   │   ├── 90-lessons.md      # vault_query section="lessons"
│   │   └── 30-architecture/   # any path works with vault_query path="..."
│   └── another-project/
└── ...

Make Your Assistant Use Hive Consistently

MCP tools don't activate on their own. Add this to your project's CLAUDE.md (or equivalent):

## Vault & Knowledge (Hive MCP)

When hive-vault MCP is available:
- `session_briefing(project="myproject")` — start every session here
- `vault_query(project="myproject", section="context")` — project overview
- `vault_search(query="...")` — find past decisions
- `capture_lesson(...)` — capture insights inline, don't wait until session end

Without these instructions, your assistant uses Hive inconsistently. With them, it uses Hive every session, predictably.

Resources & Prompts

5 MCP Resources for auto-discoverable data:

4 MCP Prompts for guided workflows:

Architecture

MCP Host (Claude Code, Gemini CLI, Codex CLI, Cursor, ...)
    └── hive-vault (MCP server, stdio)
            ├── Vault Tools (16) ── Obsidian vault (Markdown + YAML frontmatter)
            │     query, search, smart_search, list_files, patch,
            │     update, create, capture_lesson, extract_lessons,
            │     summarize, validate, session_briefing, recent,
            │     usage, health, list_projects
            │
            └── Worker Tools (3) ── Task delegation + routing:
                  delegate_task        1. Ollama (local, free)
                  list_models          2. OpenRouter free tier
                  worker_status        3. OpenRouter paid ($1/mo cap)
                                       4. Reject → host handles it

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 (265 tests, 92% coverage)

License

MIT