Limbo

The LLM-controlled data layer.

An MCP server where AI agents define schemas, generate their own CRUD tools, and manage structured application data — no database required.

---

What is Limbo?

Most AI assistants hit a wall when a user wants to track something new. The schema doesn't exist. The API wasn't built for it. The assistant can only apologize.

Limbo removes that wall. When a user says "start tracking my calories", the LLM creates the data domain, generates its own tools to interact with it, and starts managing the data — all on the fly.

Limbo is not a memory layer. Memory systems (Mem0, LangMem) store facts about users. Limbo stores user data the LLM manages — calorie logs, budgets, habit trackers, inventories.

Limbo is not text-to-SQL. There's no pre-defined schema. The LLM doesn't query a human-designed database — it builds and operates its own.

---

How it works

┌─────────────────────────────────────────────────┐
│                  LLM Agent                      │
│          (converses with the user)              │
└──────────────────┬──────────────────────────────┘
                   │ MCP Protocol
                   ▼
┌─────────────────────────────────────────────────┐
│              Limbo MCP Server                   │
│                                                 │
│  ┌─────────────┐  ┌──────────────────────────┐  │
│  │  Registry   │  │  Domain Manager          │  │
│  │  (manifest) │  │  (create/evolve domains) │  │
│  └─────────────┘  └──────────────────────────┘  │
│                                                 │
│  ┌─────────────────────────────────────────────┐│
│  │  Tool Runtime                               ││
│  │  (loads & executes LLM-generated JS tools)  ││
│  └─────────────────────────────────────────────┘│
│                                                 │
│  ┌─────────────────────────────────────────────┐│
│  │  Store (pluggable)                          ││
│  │  v1: JSON files on disk                     ││
│  │  v2: Redis / SQLite / Postgres (roadmap)    ││
│  └─────────────────────────────────────────────┘│
└─────────────────────────────────────────────────┘

Three components:

1. Registry — a registry.json manifest tracking every domain the LLM has created: name, purpose, data pattern, and tool list.
2. Store — flat JSON files on disk. No database engine needed. Swappable to Redis/SQLite/Postgres later without changing the tool interface.
3. Tool Runtime — the LLM generates JavaScript tool handlers, Limbo registers and executes them. The tools and data co-evolve as the LLM's needs change.

---

Quickstart

Prerequisites: Bun installed.

git clone https://github.com/Codinghaze-AI/llm-limbo.git
cd llm-limbo
bun install
bun start

The server runs over stdio (standard MCP transport). Connect it to any MCP-compatible agent.

---

Connect to Claude

Claude Desktop

Add to your claude_desktop_config.json:

Mac: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "limbo": {
      "command": "bun",
      "args": ["run", "/absolute/path/to/llm-limbo/index.ts"],
      "env": {
        "LIMBO_DATA_DIR": "/absolute/path/to/llm-limbo/data"
      }
    }
  }
}

Restart Claude Desktop. You'll see Limbo's tools available in the tool picker.

Claude Code (CLI)

Add to your project's .mcp.json or run:

claude mcp add limbo -- bun run /absolute/path/to/llm-limbo/index.ts

---

Available tools

Tool	Description
limbo_create_domain	Register a new data domain with a name, purpose, and file pattern
limbo_list_domains	List all domains and their metadata
limbo_describe_domain	Get full details on a domain including its tools
limbo_delete_domain	Delete a domain and all its data (requires confirm: true)
limbo_generate_tool	Generate and register a JS tool handler for a domain
limbo_execute_tool	Execute a previously generated tool by name
limbo_list_tools	List all generated tools, optionally filtered by domain

---

Example: Calorie tracker

Tell Claude: "I want to start tracking my calories."

Claude will:

1. Call limbo_create_domain to register a calories domain
2. Call limbo_generate_tool to generate calories_log_meal, calories_daily_total, and calories_weekly_summary
3. Use those tools to log meals and query totals — all in the same conversation

The data lives in data/domains/calories/data/YYYY-MM-DD.json. The tools live in data/domains/calories/tools.json. Everything is readable, editable, and portable.

---

Configuration

Env var	Default	Description
LIMBO_DATA_DIR	./data	Where Limbo stores all domain data and the registry

---

Roadmap

• Phase 0 (now) — Core MCP server, registry, flat JSON store, tool runtime
• Phase 1 — Tool versioning, schema migration, input validation, CI hardening
• Phase 2 — JS sandbox (isolated-vm), multi-tenancy, auth, audit logging
• Phase 3 — Pluggable storage backends (Redis, SQLite, Postgres)
• Phase 4 — Cross-domain queries, domain analytics, schema inference
• Phase 5 — npm package, Docker image, docs site, agent integration guides

---

Contributing

See CONTRIBUTING.md.

License

MIT