trinitas-mcp

MCP server for context, memory, and multi-model orchestration. Designed as a reusable component — not tied to a single dashboard — and extensible for review workflows and dialog-state management.

Built with the official Model Context Protocol Python SDK (mcp[cli] / FastMCP).

Features

Area	Capability
Memory	Store and retrieve dialog turns per session (SQLite)
Resources	conversation://{session_id} — full session context as JSON
Routing	Provider recommendations via AI Provider Registry

Tools

• store_dialog_turn — persist a message turn with optional metadata
• retrieve_recent_context — fetch recent turns for LLM context
• search_memory — substring search within a session
• route_best_model — score providers for a task type
• list_available_providers — list registry providers and capabilities

Requirements

• Python >= 3.10
• Network access for registry tools (cached locally)

Quickstart

cd /opt/AI_trinitas-mcp

# Create virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install (editable)
pip install -e .

# Run server (stdio — for MCP clients)
trinitas-mcp

Environment variables

Variable	Default	Description
TRINITAS_DATA_DIR	data	Base directory for local data
TRINITAS_MEMORY_DB	{TRINITAS_DATA_DIR}/trinitas_memory.db	SQLite memory path
PROVIDER_REGISTRY_URL	GitHub Pages registry.json	Registry document URL
PROVIDER_REGISTRY_BASE	GitHub Pages base URL	Registry base for future detail fetches
REGISTRY_CACHE_TTL_SECONDS	60	In-memory registry cache TTL

No API keys are required for the MCP server itself. Provider credentials belong in the calling client or orchestrator.

Host orchestrator (review execution)

The MCP server prepares multi-provider reviews via execute_review but does not call external LLM APIs. Use the included host orchestrator example to run the prepared execution_request entries and persist findings.

Workflow:

propose_review_strategy → execute_review → execute_host_review → store_review_findings
       (MCP tool)           (MCP tool)        (host script)         (optional persist)

Mock run (no API key):

source .venv/bin/activate
python examples/host_orchestrator.py --session demo-review

Live run (OpenAI-compatible API):

export TRINITAS_ORCHESTRATOR_API_KEY=sk-...
export TRINITAS_ORCHESTRATOR_MODEL=gpt-4o-mini
# OpenRouter example:
# export TRINITAS_ORCHESTRATOR_BASE_URL=https://openrouter.ai/api/v1
python examples/host_orchestrator.py --session demo-review --live --persist

Variable	Description
TRINITAS_ORCHESTRATOR_API_KEY	API key (falls back to OPENAI_API_KEY)
TRINITAS_ORCHESTRATOR_BASE_URL	OpenAI-compatible base URL
TRINITAS_ORCHESTRATOR_MODEL	Default model for all reviewers
TRINITAS_MODEL_<PROVIDER>	Per-provider model override (uppercase id)

Core module: trinitas_mcp.orchestrator (execute_host_review, OpenAICompatibleClient, MockLLMClient).

Claude Desktop / Cursor

After pip install -e ., add to your MCP config:

{
  "mcpServers": {
    "trinitas-mcp": {
      "command": "trinitas-mcp",
      "env": {
        "TRINITAS_DATA_DIR": "/absolute/path/to/trinitas-mcp/data"
      }
    }
  }
}

See examples/claude_desktop_config.json for a full example.

Cursor: Settings → MCP → add the same server block, or edit .cursor/mcp.json.

Development (without install):

{
  "mcpServers": {
    "trinitas-mcp": {
      "command": "python",
      "args": ["-m", "trinitas_mcp.server"],
      "cwd": "/absolute/path/to/trinitas-mcp",
      "env": {
        "PYTHONPATH": "/absolute/path/to/trinitas-mcp/src"
      }
    }
  }
}

Project layout

src/trinitas_mcp/
├── server.py           # Entry point & tool/resource registration
├── config.py           # Environment-based settings
├── memory/             # SQLite dialog store
├── registry/           # AI Provider Registry client & routing
├── orchestrator/       # Host-side review execution (LLM calls)
├── tools/              # MCP tools (memory, registry, review)
├── resources/          # MCP resources (conversation, review)
└── prompts/            # Review workflow prompt templates

examples/
├── claude_desktop_config.json
└── host_orchestrator.py   # Runnable review orchestrator demo

Extending

Add new tools by creating a module under tools/ and calling register(mcp) from server.py. Review orchestration and dialog-state prompts can live under prompts/ using @mcp.prompt().

Links

• Model Context Protocol
• AI Provider Registry
• MCP Python SDK

License

MIT