MCP Plexus

Provides a Python framework for spinning up MCP servers that handle multiple independent tenants within a single process, with request-scoped context isolation to prevent cross-tenant data leakage. Each tenant request maintains isolated state through context managers and thread-local or async-context storage, enabling safe multi-tenant deployments without separate server instances.

Integrates OAuth 2.1 flows to authenticate users and exchange authorization codes for access tokens, with built-in token refresh, expiration tracking, and secure credential storage. The framework handles the full OAuth handshake (authorization request, callback handling, token exchange) and manages token lifecycle including refresh token rotation and expiration-based re-authentication.

Enables MCP tools to call external APIs (REST, GraphQL, RPC) with automatic credential injection from the OAuth token store, using a declarative binding pattern that maps tool definitions to external endpoints. Tools are defined with parameter schemas, and the framework automatically injects authenticated credentials (Bearer tokens, API keys, or custom headers) based on the current tenant and user context.

Routes incoming MCP requests to tenant-specific handlers and propagates tenant identity through the entire request lifecycle (tool invocation, credential lookup, logging). Tenant context is extracted from request headers, JWT claims, or URL paths and made available to all downstream components via context managers or async context variables, enabling tenant-aware logging, auditing, and resource isolation.

Provides a Python DSL or decorator-based system for defining MCP tool schemas (input parameters, output types, descriptions) with automatic JSON Schema generation and request/response validation. Tool definitions are declarative (not imperative), enabling the framework to generate OpenAPI/JSON Schema documentation and validate tool invocations against the schema before execution.

Implements async/await-based request handling using Python's asyncio, with connection pooling for external API calls to reduce latency and resource overhead. The framework manages a pool of HTTP connections (via aiohttp or httpx) and reuses them across multiple tool invocations, avoiding the overhead of creating new connections for each external API call.

Automatically logs all MCP operations (tool invocations, credential lookups, errors) with tenant context, timestamps, and execution metadata, enabling audit trails for compliance and debugging. Logs include tool name, parameters (with sensitive data masked), execution time, and tenant/user identifiers, and can be routed to multiple backends (files, cloud logging services, SIEM systems).

Implements structured error handling that distinguishes between credential-related failures (expired tokens, invalid API keys), transient API errors, and tool logic errors, with automatic recovery strategies. When a tool fails due to an expired token, the framework automatically attempts token refresh before retrying; for transient errors, it implements exponential backoff; for logic errors, it returns detailed diagnostics.

Allows different tenants to have different configurations (enabled tools, rate limits, feature flags) without code changes, using a configuration store (database, config file, or feature flag service). Configuration is loaded per-request based on tenant context and cached to avoid repeated lookups, enabling A/B testing and gradual feature rollouts.

Provides a plugin interface that allows developers to register custom tool implementations without modifying the core framework, using a discovery mechanism (module scanning, explicit registration, or entry points) to load plugins at startup. Plugins implement a standard tool interface and are automatically integrated into the MCP server with schema validation and error handling.