Tin CanMCP

A prescriptive template for building large MCP servers (20+ tools) organized as a tiered catalog with a broker.

Every tin-canMCP server exposes exactly 7 tools to the AI client, regardless of catalog size. The catalog itself can contain hundreds of tools; the broker gates visibility so the AI only sees a bounded, task-relevant subset at any moment.

When to use this template

Use tin-canMCP when your MCP will wrap many tools (CLI binaries, API endpoints, AppleScript commands, SDK methods). The broker pattern keeps the AI's visible tool count under a cache-friendly ceiling.

Do NOT use tin-canMCP if you only need 3–10 tools with no grouping. A plain FastMCP server is simpler.

The 7 tools every server exposes

Tool	Purpose
catalog_status	Show current activation state and routing hints
catalog_list_groups	List every group in the catalog
catalog_search_tools	Search the full catalog by name/description (no activation needed)
catalog_activate	Turn on groups/tiers so their tools become runnable
catalog_escalate	Widen activation to the next tier (canonical → advanced → internal)
catalog_active_tools	Show which tools are currently runnable
catalog_run_tool	Execute an active tool by name with args

Domain tools are reached only through catalog_run_tool. See HOW-TO-USE.md for the full workflow.

Quick start

pip install -r requirements.txt
cp .env.example .env

# Scaffold a catalog for your domain
python scripts/scaffold_catalog.py \
  --name mydomain \
  --source cli \
  --tools list,get,create,delete \
  --out catalog-runtime \
  --force

# Validate the registry is structurally sound
python scripts/validate_registry.py catalog-runtime

# Run the server
python main.py --transport=stdio

At this point the server loads, exposes the 7 broker tools, and catalog_run_tool will refuse to run anything until you fill in the TODO markers in catalog-runtime/registry/tools.json and the per-tool manifests under catalog-runtime/tools/.

Next: read HOW-TO-USE.md — it walks through authoring tool manifests, grouping strategies, and tier assignment.

Repository layout

.
├── app/
│   ├── catalog_runtime/     # Broker, loader, executor (DO NOT modify)
│   ├── tools/               # Registers the 7 broker MCP tools (DO NOT modify)
│   ├── prompts/             # Optional MCP prompts
│   ├── config.py            # Env-backed settings
│   └── server.py            # FastMCP server
├── catalog-runtime/
│   ├── registry/            # YOUR 7 registry JSON files — edit these
│   ├── tools/               # YOUR per-tool manifests — edit these
│   └── schema/              # JSON Schemas for registry files (reference)
├── scripts/
│   ├── scaffold_catalog.py  # Generate a registry skeleton
│   └── validate_registry.py # Check invariants (run in CI)
├── ai-prompts/              # Master prompt + per-source hints for AI implementation
├── docs/
│   └── catalog-runtime-template.md  # Full schema contract reference
├── main.py
└── HOW-TO-USE.md            # Prescriptive authoring workflow

Configuration

Environment variables (via .env or shell):

• VAR_CATALOG_RUNTIME_DIR — path to your catalog runtime (default: repo-local catalog-runtime/)
• VAR_SERVER_NAME — MCP server name shown to clients (default: tin-canMCP)
• VAR_COMMAND_TIMEOUT_SECONDS — per-tool timeout (default: 20)
• VAR_MAX_OUTPUT_CHARS — truncation threshold for stdout/stderr (default: 12000)
• MCP_TRANSPORT — stdio, sse, or web (default: stdio)

What NOT to do

• Do not add @mcp.tool() decorators for your domain tools in app/tools/. That defeats the broker. Add catalog entries instead.
• Do not raise max_tools_per_activation above ~15. The whole point is a bounded visible surface.
• Do not put every tool in canonical. Use advanced and internal tiers to defer niche tools.
• Do not skip validate_registry.py. Cross-reference bugs are silent.

Acknowledgments

This template was inspired by josharsh's work on MCP broker patterns.

License

MIT. See LICENSE.