One runtime instead of three stacks
Building serious AI tooling today means gluing three things: an agent SDK for the agents, an MCP framework for the tools, and then the part you end up writing yourself — access control, OAuth token handling, background execution, scheduling, storage, monitoring, approval steps. TAI42 ships all three as one runtime. You write the capability; the operational layer is already there. A tool is a plain decorated function; a one-file manifest declares what the server loads (nothing loads implicitly — that’s the point: the manifest IS the system definition):tai serve --manifest-path manifest.yml — and greet is an MCP tool
callable from Claude Desktop, Cursor, the bundled web console (the
Studio), the CLI, or an agent. The auth middleware, the OAuth token refresh, the cron wrapper,
the job queue, the audit trail you were about to write around it —
already running.
Already have an MCP server? TAI42 is built on FastMCP and speaks
standard MCP — any server you already run mounts as-is (over HTTP, a
unix socket, or as a stdio process the runtime launches, per the
mcp: tab above); its tools appear next to your own, behind the same
auth and monitoring. Nothing needs porting. And when a tool is
sensitive, it can pause for human sign-off — answered from the
Studio inbox or a connected chat channel (Slack, Telegram, WhatsApp).
The server is alive
With plain MCP frameworks, your server is a build artifact — change a tool, restart the process. A TAI42 server is operated live: reload a tool in place, replace the manifest across every worker process at once, re-probe and re-attach an external MCP server that was down, refresh config in place — no restart. Each of these operations is declared once and exposed three ways — HTTP API, CLI command, and optionally as an MCP tool your agents can call (an admin chooses which ones) — so the UI, your terminal, and your agents always drive the same thing.What’s in the box
Tools
A decorated function or a mounted server’s tool. Wrap one with an
extension (rate-limit it, transform it), save a configured
variant as a versioned preset.
Agents
An included agent framework — you author an
Agent class in
Python, the runtime runs it and streams every step (messages,
reasoning, tool calls) live. Models ride standard LangChain
providers (OpenAI, Anthropic, …) with your own keys, set in
config. Callable exactly like any tool.The operational layer
Key-based identities with scoped policies checked on every
request, OAuth connectors, scheduling, background execution,
storage, monitoring, human sign-off — supplied, not built by you.
uv or Docker.
Start here
Install
Install from PyPI and boot a working server in minutes.
Quickstart
Serve the hello example and call a tool over MCP.
Mental model
How the runtime fits together, in one page.
Marketplace
Browse plugins and install into a running server.

