Infer MCP Server

This is an MCP (Model Context Protocol) server providing resource access via SSH and RDBMS connections. It is designed to integrate with AI Copilot tools like GitHub Copilot in VS Code and Cursor.

Setup

1. Install dependencies: npm install
2. Build: npm run build
3. Run: npm start

Features

• SSH command execution enforced through configured profiles, command allowlists, and output quotas
• PostgreSQL database queries limited to configured connections and statement patterns
• Per-profile concurrency limits with cancellation-aware execution and progress notifications
• Classifier training orchestration via SSH profiles

Configuration

The server loads configuration from either INFER_MCP_CONFIG_PATH (JSON file) or INFER_MCP_CONFIG (inline JSON string). A starter config is available at config/sample-config.json; copy .env.example to .env and update paths/secrets as needed.

Secrets can be provided inline, via environment variables, or read from disk. Example:

{
	"sshProfiles": {
		"training-cluster": {
			"host": "cluster.example.com",
			"username": "trainer",
			"privateKey": {
				"path": "./secrets/training-cluster.key"
			},
			"policy": {
				"allowedCommands": ["^python\\s+train.py\\b"],
				"maxExecutionMs": 900000,
				"maxOutputBytes": 1048576
			}
		}
	},
	"databaseProfiles": {
		"training-metadata": {
			"connectionString": { "env": "TRAINING_METADATA_DB_URL" },
			"allowedStatements": ["^\\s*SELECT\\b", "^\\s*WITH\\b"],
			"maxRows": 1000,
			"maxExecutionMs": 20000
		}
	},
	"training": {
		"defaultCommandTemplate": "python train.py --dataset={{datasetPath}} --class={{subclass}}",
		"defaultTimeoutMs": 600000
	}
}

• sshProfiles define reusable credentials for tools such as sshExecute and trainClassifier. For password, privateKey, or passphrase, supply either a raw string, { "env": "VAR_NAME" }, or { "path": "relative/or/absolute" }. Base64-encoded files are supported with { "path": "...", "encoding": "base64" }. Policies control command allowlists, maximum runtime, captured output size, and per-profile maxConcurrent slots.
• databaseProfiles centralise PostgreSQL access. Statements must match the configured regex allowlists and respect row/time limits, with maxConcurrent restricting simultaneous queries per profile.
• training controls defaults for classifier jobs.

Integration

Configure in your AI tool's MCP settings to connect to this server.

For VS Code GitHub Copilot: Add to mcp.json in .vscode folder.

Debugging

You can debug this MCP server using VS Code's debugger.

Simulator

Build the project (npm run build) and use the simulator to exercise tools locally without an agent client:

npm run simulate -- list
npm run simulate -- call sshExecute '{"profile":"training-cluster","command":"python train.py --help"}'
npm run simulate -- call dbQuery '{"profile":"training-metadata","query":"SELECT * FROM jobs LIMIT 5"}'

Override defaults with environment variables:

• MCP_SERVER_COMMAND – binary to launch (default node)
• MCP_SERVER_ARGS – comma-separated arguments (default build/index.js)
• MCP_SERVER_CWD – working directory for the spawned server

For staging validation, copy config/staging-config.json, populate the referenced secrets, and set INFER_MCP_CONFIG_PATH=$(pwd)/config/staging-config.json before running npm run simulate -- … commands.

Integration Testing

• Build the project (npm run build) so the simulator artifact exists.
• Configure staging credentials via INFER_MCP_CONFIG_PATH or INFER_MCP_CONFIG (the staging sample references environment variables and secret files under config/secrets/).
• Run npm run test:integration to execute Vitest suites that shell out to the simulator (guarded by INTEGRATION=1).
• The integration suite expects the simulator to list sshExecute, dbQuery, and trainClassifier; extend tests/integration/ with additional cases as you add tools.