Omnispindle

A todo system that went wonderfully, intentionally wrong.

Omnispindle is the coordination spine of the Madness Interactive ecosystem — a Python FastMCP server with 33 tools that lets AI agents manage tasks, capture knowledge, coordinate sessions, and navigate the whole workshop from a single, standardized interface. PyPI packaged. Auth0 integrated. Runs anywhere a MCP config lives.

It started as "let's do todos properly." It became the central nervous system for a multi-project AI-assisted development lab. Both of these things are fine.

---

What it actually does

Todo management — the boring part that enables everything else. Agents can create, query, update, complete, and audit tasks across any project in the ecosystem with full metadata, priority, target agent tracking, and change detection.

Knowledge capture — lessons learned get stored with language, topic, and tag metadata. Searchable by regex, text, or vector embedding. The institutional memory doesn't evaporate when the conversation ends.

Session tracking — AI work sessions in Inventorium can be forked, spawned, linked to todos, and traced through a full genealogy tree. Every thread of work has a parent and a lineage.

Semantic search — find_relevant uses vector embeddings to surface todos and lessons by meaning, not just keywords. get_context_bundle gives an agent the full project picture in one call.

---

Installation

pip install omnispindle

CLI commands available after install:

• omnispindle-stdio — MCP stdio server for Claude Desktop
• omnispindle / omnispindle-server — HTTP web server for authenticated endpoints

Claude Desktop (zero config)

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "omnispindle": {
      "command": "omnispindle-stdio",
      "env": {
        "OMNISPINDLE_MODE": "api",
        "OMNISPINDLE_TOOL_LOADOUT": "basic",
        "MCP_USER_EMAIL": "[email protected]"
      }
    }
  }
}

First tool call opens your browser for Auth0 login. Token saves locally. That's it.

Development

git clone https://github.com/DanEdens/Omnispindle.git
cd Omnispindle
pip install -r requirements.txt
python -m src.Omnispindle.stdio_server

---

Tools

Full loadout is 33 tools across 6 categories. Control what's available — and your agent's token budget — with OMNISPINDLE_TOOL_LOADOUT.

Todo Management (9 tools)

Tool	What it does
add_todo	Create a task with project, priority, target agent, notes, and metadata
query_todos	MongoDB-style filter queries with projection, limit, offset, since change detection, and graph_root for dependency subgraph traversal
update_todo	Patch any fields; metadata is deep-merged. Supports $push/$pull on array fields (e.g. metadata.blockers) for dependency linking
delete_todo	Remove a task
get_todo	Fetch a single task by ID
complete_todo	Stage for review with optional comment; writes to audit log. Sets status to review, not completed
list_todos_by_status	Filter by status: pending initial in_progress blocked review completed
search_todos	Tokenized multi-word fuzzy text search
list_project_todos	Recent tasks for a specific project

Knowledge / Lessons (7 tools)

Tool	What it does
add_lesson	Capture a lesson with language, topic, and tags
get_lesson	Fetch by ID
update_lesson	Patch lesson content or metadata
delete_lesson	Remove
search_lessons	Text search across lesson fields
grep_lessons	Regex pattern search
list_lessons	Browse all lessons, with brief mode for summaries

Inventorium Sessions (8 tools)

Tool	What it does
inventorium_sessions_list	List sessions, optionally filtered by project
inventorium_sessions_get	Fetch a session by ID
inventorium_sessions_create	Start a new AI work session
inventorium_sessions_spawn	Spawn a child session from a parent, linked to a todo
inventorium_sessions_fork	Fork a session with optional message inheritance
inventorium_sessions_genealogy	Full ancestor/descendant trace for a session
inventorium_sessions_tree	Visual session tree for a project
inventorium_todos_link_session	Link a todo to a session

Context & Search (2 tools)

Tool	What it does
get_context_bundle	One call: recent todos, lessons, session state, project stats for an agent's working context
find_relevant	Semantic RAG search via vector embeddings — finds related todos and lessons by meaning

System / Admin (5 tools)

Tool	What it does
query_todo_logs	Audit log queries: filter by type, project, date range
list_projects	Enumerate known projects from filesystem
explain / add_explanation	Topic explanation system: persistent knowledge cards
point_out_obvious	Logs an observation with configurable sarcasm. Useful for marking known issues during automated runs.

Custom Code (1 tool)

Tool	What it does
bring_your_own	Inject a Python, JavaScript, or shell function as a live MCP tool at runtime

---

Tool Loadouts

Set OMNISPINDLE_TOOL_LOADOUT to control what's registered:

Loadout	Tools	Use case
full	33	Everything
basic	10	Core todo CRUD + get_context_bundle
minimal	4	Add, query, get, mark complete
lessons	7	Knowledge management only
admin	14	Admin tasks + session management
write_only	6	Create/update/delete only
read_only	15	Query/get only
lightweight	13	Token-optimized core
agent_preflight	6	Session startup: context bundle, RAG check, todo ops

---

Operation Modes

Set via OMNISPINDLE_MODE:

• api — HTTP calls to madnessinteractive.cc/api. No local database needed. Best for cloud-native or multi-user setups.
• hybrid (default) — API-first with MongoDB fallback. Reliable when the network isn't.
• local — Direct MongoDB connections only. Good for offline development.
• auto — Benchmarks both and picks the faster one.

---

Authentication

Zero-config device flow: On first tool call, a browser window opens for Auth0 login. Token is saved locally. All subsequent calls are authenticated without any configuration.

Manual token setup (optional):

python -m src.Omnispindle.token_exchange

Environment variables:

• MADNESS_AUTH_TOKEN or AUTH0_TOKEN — JWT from Auth0 device flow
• MCP_USER_EMAIL — required for per-user data isolation
• MADNESS_API_URL — override API base (default: https://madnessinteractive.cc/api)

All data is scoped per user at the database level. Your todos stay yours.

---

Configuration

# Operation mode
OMNISPINDLE_MODE=hybrid          # api | hybrid | local | auto
OMNISPINDLE_TOOL_LOADOUT=basic   # see loadouts table above
OMNISPINDLE_FALLBACK_ENABLED=true
OMNISPINDLE_VALIDATE_PROJECT_NAMES=false  # optional: enable slower DB-backed project validation

# Authentication
MADNESS_AUTH_TOKEN=<jwt>
[email protected]

# Local/hybrid database
MONGODB_URI=mongodb://localhost:27017
MONGODB_DB=swarmonomicon

# Real-time events
MQTT_HOST=localhost
MQTT_PORT=1883

---

Ecosystem

Omnispindle is the AI interface layer. The rest of the workshop:

Inventorium — React web dashboard and SwarmDesk 3D spatial workspace. Humans use this. It reads todos and sessions via the REST API directly — not MCP, which is strictly for AI agents. If you're clicking buttons, you're in Inventorium. If you're an agent, you're calling MCP tools.

cartogomancy — JS/TS codebase analysis tool (npm install -g @madnessengineering/cartogomancy). Point it at any JavaScript or TypeScript project; get back a rich JSON map of structure, complexity, git blame, and cross-references. Feed that map to SwarmDesk and your codebase becomes a 3D city you can walk through.

SwarmDesk — The 3D spatial visualization layer inside Inventorium. Todos, sessions, and cartogomancy code maps rendered as explorable architecture. Your project as a place, not a list.

MadnessVR — Quest 2 VR version of SwarmDesk. Put on the headset, walk through your codebase as actual geometry. This one is on the roadmap, not the release notes. But the plan exists and it's detailed.

Cogwyrm — AI chat companion integrated into the ecosystem. Uses Omnispindle MCP tools during conversations to read and write shared context.

Swarmonomicon — The core database layer. MongoDB, swarmonomicon database. Everything persistent lives here.

The architecture rule

UI (Inventorium, forms, dashboards) → REST API → Database
AI agents (Claude, Cogwyrm, any MCP client) → MCP tools → Backend → Database

Never route UI through MCP. Never call the REST API directly from an AI agent. The separation is intentional and load-bearing.

---

Development

# Tests
pytest tests/

# Stdio server (Claude Desktop)
python -m src.Omnispindle.stdio_server

# Web server
python -m src.Omnispindle

# Check tool count
python -c "from src.Omnispindle.tool_loadouts import _BASE_LOADOUTS; print(len(_BASE_LOADOUTS['full']), 'tools in full loadout')"

PyPI Publishing

# Build and check
python -m build
python -m twine check dist/*

# Publish
python -m twine upload dist/*

---

Privacy & Security

This repo contains Auth0 configs and infrastructure-as-code for our ecosystem. It's open source for learning and forking — not for deploying as-is.

If you're actually running this for your own setup:

1. Fork it
2. Replace all auth providers and credentials
3. Point it at your own domain and database
4. Review tool permissions (we're permissive by default)

This is a working system for our lab. For yours, make it yours.

---

Philosophy

Most people build a todo app with 5 features. We built one with 33 MCP tools, three operation modes, session genealogy trees, vector embeddings, zero-config OAuth, and a roadmap that includes walking through your code in VR.

This is the right amount of complexity. Every piece is load-bearing.

Working now:

• Todo management for AI agents — robust, audited, per-user
• Persistent knowledge capture across projects
• Session tracking with full lineage
• Semantic search via vector embeddings
• Zero-config Auth0 that doesn't require a setup guide

Coming:

• MadnessVR: SwarmDesk on Quest 2
• Teaching prompt engineering through Terraria-style inventory mechanics
• cartogomancy → SwarmDesk → VR as one continuous pipeline

We write careful code. We're not afraid to push the boundaries when it's practical.

---

"Over-engineered? Maybe. Under-ambitious? Never."