Cruxhive

The hive mind for your team's AI context — approved, versioned, trusted.

Your coding AI forgets everything between sessions. Your teammates' AIs forget everything too. CruxHive fixes that — with a human approval gate, hybrid semantic search, and a knowledge base that compounds over time.

npx cruxhive init

One human-curated knowledge base. Every AI tool reads from it — Claude Code, OpenCode, Cursor, Windsurf, Gemini CLI. Local SQLite, MIT licensed, zero cloud dependency.

---

Install

# One command — installs the MCP server (via uv) and wires up your project
npx cruxhive init

Or install the engine directly:

uv tool install cruxhive-mcp      # Python 3.11+

Wire into any MCP client — add to .mcp.json:

{
  "mcpServers": {
    "cruxhive": { "command": "uvx", "args": ["cruxhive-mcp"], "type": "stdio" }
  }
}

How it works

AI tools never write to your knowledge base directly. They propose; a human approves.

1. Your AI calls context_propose → the entry lands in .llm/pending/.
2. You run cruxhive review → approve or reject, with full git history.
3. Approved entries get indexed and become searchable by every AI tool, every session.

Solo developer? cruxhive solo auto-approves your own proposals.

CLI

init        Bootstrap CruxHive in the current project
index       Index .llm/ markdown files into the local knowledge base
propose     Propose a new knowledge entry for human review
review      Interactively approve or reject pending proposals
sync        Sync org-layer context from the configured remote
health      Knowledge base health summary
stats       Usage observability — searches, hit rate, gaps, by AI tool
digest      Weekly markdown digest — gaps, decayed entries, queue health
status      One-line health summary (use --quiet for hooks)
doctor      Diagnose setup — symlinks, hooks, slash commands
workspace   Cross-project rollup — aggregate KPIs across all projects
direnv      Write a .envrc that auto-logs sessions for Cursor/Windsurf/Gemini
solo        Enable/disable solo mode — auto-approve your own proposals
ui          Open the approval-queue dashboard (localhost:3847)

MCP tools

Tool	What it does
context_search	Hybrid search (FTS5 BM25 + optional sqlite-vec, RRF fusion, entity boost)
context_workspace_search	Same, across every configured project
context_propose	Submit a knowledge entry for human review
context_review / context_approve / context_reject	The approval gate
context_index	(Re)index .llm/ markdown into the local store
context_check_faithfulness	NLI-based contradiction detection on captured context
context_radar	Scan recent commits, map to plan areas, surface uncovered work
context_next_slice · context_write_plan · context_sync_memory	Planning + org-layer sync

Status

Published. cruxhive-mcp on PyPI · @cruxhive/cli on npm. MIT, 0 cloud dependencies.

• Phase 1 — Internal dogfooding + MCP skill layer
• Phase 2 — Multi-tool support + bootstrap script
• Phase 3 — OSS extraction + public release
• Phase 4 — Cross-project org layer + hybrid semantic search

Why

Engineering work has 7 stages. Every AI tool covers stages 1–6. Stage 7 — Capture — has no tooling. That's where the value leaks out: hard-won context dies at the end of every session.

CruxHive is infrastructure for Stage 7.

What makes it different

	AGENTS.md	Grov	Kiro	CruxHive
Human approval workflow	✗	✗	✗	✓
Git versioning + audit trail	partial	✗	✗	✓
Semantic search (hybrid, local)	✗	✓ (API)	✗	✓
Org-layer (cross-project)	✗	✗	✗	✓
Faithfulness detection	✗	✗	✗	✓
Tool-agnostic via MCP	partial	partial	Kiro only	✓

Links

• Site: https://cruxhive.github.io/cruxhive/
• PyPI: https://pypi.org/project/cruxhive-mcp/
• npm: https://www.npmjs.com/package/@cruxhive/cli

License

MIT