Skip to main content
Install the runtime from PyPI. tai42-skeleton pulls in the tai42-contract, tai42-kit, and tai42-cli packages it builds on — tai42-cli provides the tai command — so one install puts the whole runtime and the tai command on your path.

​
Requirements

  • Python 3.13 or newer.
  • uv for the uv add, uv venv, and uv pip install commands below. The pip install needs no uv; the from-source flow does.
  • Redis 7.0+ and Postgres only for the features that need them — access control, connectors, and interactions. Local tools and mounted MCP servers need no services at all.

​
Install

Run that inside an activated virtual environment — a stock system Python refuses a bare pip install with error: externally-managed-environment. Or, in a uv project:

​
Add the standard toolbox

The toolbox extra pulls in the tai42-toolbox contrib package — its generic tools and its composition extensions (chain, batch) load straight from a manifest:
See the standard toolbox guide for what ships in it.

​
Add an identity provider

Access control is on by default, and the runtime refuses to boot when no identity provider is registered — the skeleton ships none of its own. Install the Redis-backed API-key provider and name it in the manifest — that is the pairing every install starts from:
manifest.yml
The provider stores key hashes in Redis, so an access-controlled server needs a reachable Redis instance. Access control explains the policy model, and the access-control guide mints the first key. For local work against tools that need no auth, turn the gate off with ACCESS_CONTROL_ENABLE=false instead.

​
Install from source

To run the skeleton itself — to read the code, run the shipped examples, or build on an unreleased change — clone the tai42 monorepo and install the skeleton (at core/skeleton) into a fresh environment with --no-sources:
--no-sources is what makes this single-package install stand on its own. tai42-skeleton’s [tool.uv.sources] points its sibling tai42-* dependencies at their monorepo workspace members for development; the flag ignores those overrides and resolves them from PyPI instead. Install ".[toolbox]" instead of . for the standard toolbox, or ".[dev]" for the test and lint tooling. Run the checkout’s commands with uv run --no-sync, which uses the environment exactly as installed. A bare uv run re-syncs the project first, which puts the [tool.uv.sources] overrides back and fails in a clone that has no sibling checkouts beside it.

​
CLI-only install

To operate a running server from another machine, install just the command-line client — it pulls in none of the server runtime:
Or, in a uv project:
This puts the tai command on your path with the remote command groups — thin clients over a server’s /api/* routes. The full local tree (tai serve, tai db, …) appears only when tai42-skeleton is installed alongside.

​
The tai command

Installing the package puts the entrypoint on your path under two names, tai and tai42. Both run the same command tree — tai is the documented one, and tai42 is an alias for environments where an unrelated tai already sits on your PATH. Three subcommands run the long-lived processes:
  • tai serve — run the MCP server.
  • tai backend — run the agent / worker backend process.
  • tai metrics — serve the Prometheus metrics endpoint.
Run tai --help for the full command tree, or browse the CLI reference. Next, boot a server and call a tool in the quickstart.