One knowledge base.
Every AI tool.

Stop re-explaining your project to Claude Code, Cursor, OpenCode every session. CruxHive gives all your AI tools one shared, human-curated, searchable knowledge base — local SQLite, MIT, zero cloud.

cruxhive-mcp 0.12 @cruxhive/cli 0.12 32 tests MIT 0 cloud
$ npm install -g @cruxhive/cli Read the docs →
~/projects/backend · cruxhive search --workspace
$ cruxhive search --workspace "auth"

Searching 8 projects · BM25 + entity + recency …

1. backend .llm/decisions/auth.md match
2. mobile .llm/plans/oidc.md match
3. analytics .llm/memory/sessions.md match
4. backend .llm/context/identity.md entity
5. web-app .llm/plans/oauth-flow.md match

✓ 5 results across 4 projects · 0.06 s · 124 entries indexed

$
01 / How it works

One server. Many tools. Three tiers.

AI TOOLS Claude Code OpenCode Cursor Windsurf Gemini CLI + any MCP client CruxHive MCP server · SQLite BM25 + entity + recency KNOWLEDGE LAYERS Personal ~/.cruxhive/personal/ Project .llm/ (in repo) Org shared remote · synced read CONTEXT.md + call MCP tools search · propose · review · approve · index human approval gate indexed locally git-versioned

Every AI tool reads from the same SQLite-backed index. You approve what goes in. Nothing leaves your machine.

02 / Tool-agnostic

Every AI tool reads from the same place.

One canonical .llm/CONTEXT.md per project. CruxHive symlinks CLAUDE.md, AGENT.md, .cursor/rules/, .windsurfRules, and GEMINI.md at it. Switch tools mid-day — same context.

For tools without their own session hook (Cursor, Windsurf, Gemini), cruxhive direnv wires a per-project .envrc so every shell entry is logged correctly.

5 AI tools 1 source of truth 0 drift
your-project/
CLAUDE.md → .llm/CONTEXT.md
AGENT.md → .llm/CONTEXT.md
.cursor/rules/cruxhive.mdc→ .llm/CONTEXT.md
.windsurfRules → .llm/CONTEXT.md
GEMINI.md → .llm/CONTEXT.md
.mcp.json cruxhive-mcp
.claude/commands/ /radar /next-slice /extract
.opencode/commands/ /radar /next-slice /extract
03 / Trust gate

AI proposes. Humans approve. Knowledge compounds.

The single rule: no AI tool writes to the knowledge base without you saying yes. /extract distills a session into candidates, /summarize into a structured research entry — all queued for one-keystroke review.

Solo? Run cruxhive solo once — your own proposals auto-approve. Stay in control without the bureaucracy of a team approval queue.

NLI conflict detection at review time · confidence decay over age
.llm/pending/ · awaiting review
constraint_auth.md [constraint]
topic: oauth · age: 1h
source: ai-proposed (/extract)
decision_db.md [decision]
topic: database · age: 3h
⚠ conflicts with .llm/decisions/orm.md
[a]pprove [r]eject [s]kip [e]dit
04 / Cross-project

One workspace. Every project searchable from every other.

Working in backend, wondering "how did I solve auth in mobile?" — cruxhive search --workspace "auth" hits every project's index and merges results.

Or open cruxhive ui --workspace for a live dashboard with the 5 KPIs that matter: hit rate, gaps, queue age, decay, sessions. Trend deltas with cruxhive digest --compare.

localhost:3847 · MCP context_workspace_search · JSON-exportable
Hit rate
74%
↑ +6 vs 7d ago
Gaps (30d)
3
↓ -2 vs 7d ago
Pending
2
oldest 4d
Decayed
8
6% of 124
05 / Why CruxHive

Two real questions answered.

vs. just using CLAUDE.md
A file is fine.
A system compounds.

One CLAUDE.md works great while you use one AI tool. The moment you also open Cursor or OpenCode, your context fragments — three files, three subtly different versions, immediate drift.

CruxHive symlinks five AI tools at one canonical .llm/CONTEXT.md and adds search, decay, cross-project retrieval, and a per-AI-tool observability log. Plain markdown files stay your source of truth; CruxHive is the layer around them.

Stay on CLAUDE.md if: you only use one AI tool, one project, and your context fits in 200 lines.
vs. Mem0 / Zep / autonomous memory
Mem0 auto-writes.
CruxHive auto-suggests.

Autonomous-memory tools extract and commit memory without asking. That is the right model for chatbots and personal AI agents.

CruxHive is for code projects, where you want every entry that becomes team knowledge to be reviewed first. The AI proposes via /extract; you approve via cruxhive review (or cruxhive solo auto-approves your own). The approval gate IS the product — hallucinations and drift never reach the shared base.

Stay on Mem0 / Zep if: you're building a chatbot, not curating a code project's memory.

One sentence: CruxHive is what you reach for when AI memory needs to be governed, searchable, multi-tool, and local.

Install in 30 seconds.

Two packages. Both public. MIT. Choose either.

RECOMMENDED · npm
@cruxhive/cli

Bootstraps everything. Installs the Python MCP server, wires 5 AI tools, sets up hooks & gitignore.

$ npm install -g @cruxhive/cli
DIRECT · PyPI
cruxhive-mcp

Just the MCP server. Point your AI tool's .mcp.json at cruxhive-mcp.

$ uv tool install cruxhive-mcp
After install ~5 min · then 2 min/day

Your first day with CruxHive

1
Bootstrap a project: cd your-project && cruxhive init
Wires 5 AI tools (Claude Code / Cursor / OpenCode / Windsurf / Gemini), installs hooks, creates .llm/CONTEXT.md and .gitignore entries. Idempotent — safe to re-run.
2
Edit .llm/CONTEXT.md — describe your project
Stack, key paths, conventions, the rules you'd tell a new hire. Every AI tool reads this at session start. This is where 80% of the value comes from. ~5 minutes.
3
Solo dev? Run cruxhive solo --name "you"
Skips the human-approval queue. Your own proposals (from /extract, /summarize, cruxhive propose) land as already-approved. Disable later with cruxhive solo --disable when you add a teammate.
4
End-of-session habit: type /extract in Claude Code or OpenCode
Distills the conversation into knowledge entries. The next session — whatever AI tool — starts knowing what this one learned. Run /summarize for a single structured research entry instead of N candidates.
5
Find anything later: cruxhive search "auth"
Searches the current project. Add --workspace to search across every CruxHive-initialized project on your machine. Use cruxhive ui for the visual dashboard, cruxhive digest weekly to see what's growing.
Stuck or want detail? Daily workflow → Full CLI reference → All docs → Verify setup: cruxhive doctor
✓Phase 1–4 · context layer · multi-tool bootstrap · MCP + SQLite · approval workflow + NLI faithfulness + web UI
✓Memory intelligence · entity-aware search · recency boost · confidence decay · ephemeral entries
✓Observability · session log · workspace dashboard · digest with --compare · 14 CLI commands
✓Solo mode + cross-project search · direnv coverage for Cursor / Windsurf / Gemini