MCP AI Suite is an open-source set of 10 standalone Python libraries for building autonomous AI agents on the Model Context Protocol (MCP). It covers memory, retrieval (RAG), sandboxed code execution, file management, scheduling, web search, planning, evaluation, and a hybrid orchestrator (kernelmcp). Each library installs independently from PyPI as mcpaisuite-<name>.
Yes. Every library is AGPL-3.0 and free for individuals and open-source projects. A commercial license is available for closed-source or commercial products — contact contact@mcpaisuite.com.
Install only what you need, e.g. pip install mcpaisuite-memorymcp, or the whole stack with pip install "mcpaisuite-kernelmcp[all]". The PyPI package names carry the mcpaisuite- prefix, but the import names stay bare — for example pip install mcpaisuite-ragmcp, then import ragmcp.
Memory, RAG, sandbox, scheduler, web search, planning and files are native and tested together with one audit trail and one config — not third-party plugins you wire and maintain yourself. The kernel also runs a hybrid ReAct + LTP engine: it can compile a goal into a deterministic plan in a single LLM call instead of calling the model on every step. In reproducible benchmarks on deep tool chains, kernelmcp Hybrid was the cheapest 100%-reliable system versus LangGraph and CrewAI on Sonnet.
kernelmcp is the orchestrator at the center of the suite. It drives a ReAct/LTP hybrid loop, enforces a constitution and a token/cost budget, auto-heals failed tool calls, routes between local, fast and cloud models, and exposes the other nine libraries as tools. It can be deployed as an autonomous agent, a tool router, or an MCP sampling server.
ReAct calls the LLM on every turn — flexible but token-heavy. LTP (Lean Task Protocol) compiles the whole plan in one LLM call and then executes it deterministically in Python — far cheaper and reproducible for structured tasks. Hybrid runs LTP first with a verified ReAct fallback, and was the most reliable mode in the benchmarks.
Be aware that this is a young project — version 1.0.x, about two months old and actively developed. The core paths are tested and measured with a reproducible benchmark harness, but it is not yet battle-hardened, so some edge cases may have rough edges. For stability, pin a version (for example mcpaisuite-kernelmcp==1.0.4). If something breaks, please report it on GitHub Issues — reports are triaged and fixed quickly.
Yes. Every library is fully standalone with its own Python API, CLI, and MCP server. You can adopt just one — for example ragmcp or memorymcp — without the rest of the suite.
Yes. Each library ships a standard MCP server, so it works with any MCP-compatible client including Claude Desktop, Cursor, and VS Code. The kernel can also use the host application’s model via MCP sampling, so it can run with no API key of its own.
It is a Python library suite (3.10/3.11+). It is model-agnostic via LiteLLM — Anthropic Claude, OpenAI, and local models through Ollama, among others — with optional smart routing between a cloud model, a fast model, and a local fallback.
Yes. At runtime the kernel can connect to any MCP server (2000+ in the community — GitHub, Slack, Notion, Postgres and more) and any LangChain tool (500+), with no wrapper code. These are reachable at runtime, not bundled with the install.