by Innominds
RELEASE 2026.11.20.2 · 20 NOVEMBER 2026

Practise without Docker, and without an LLM key

The first PyPI release with the practice project: tv practice-project (built in 2026.11.20.1, which was not published) gives you an LLM agent to onboard, and it now runs from PyPI on any machine with Python: tv practice-agent starts it with TruVelocity's recording proxy in front of its tools. And tv practice-project --scripted gives it an offline rule-based model, so the whole exercise needs no LLM key and costs nothing.

0Docker or LLM keys needed for the full exercise
15 / 15cases pass on the clean build offline: gate passes
6 of 15cases fail on the buggy build offline: gate fails
1064automated tests pass
What's new

Two new ways to practise

The same project, catalog, steps and planted defects as with Docker and a real model. Only how the agent runs changes.

The agent on PyPI

pip install harbor-expense-agent installs the expense desk (fictitious) with its tools, seed data and LLM support. Its command harbor-expense-agent starts the tools and the agent; Ctrl+C stops both.

tv practice-agent

The no-Docker twin of the project's Compose file: from the project's .env it starts tv proxy on 9710 and the agent on 8700 behind it, with the records and the proxy's log in the project's data folder, where TruVelocity already reads them. It runs the agent as a black box, over HTTP.

No LLM key: SCRIPTED

tv practice-project --scripted writes DEFAULT_LLM_PROVIDER=SCRIPTED and LLM_SCRIPT=expense_agent.policy:factory: a deterministic stand-in for the model that follows the agent's written procedures. It makes no LLM call and nothing is billed, and the planted defects still apply.

What the offline model doesn't show

It shows no model mistakes of its own and no variation between tries (no flaky results). Use a real model to see those: the same .env takes any provider the agent supports.

The guide

TruVelocity-Onboard-a-New-Agent.html has a Without Docker section and a No LLM key option. Where it says docker compose, it now also says what to do with tv practice-agent, and Troubleshooting covers this route.

Both routes offered

tv practice-project prints the Docker and the no-Docker commands. Without Docker on the machine it prints only the pip route.

Get started

No Docker, no key

Install or upgrade TruVelocity, with its UI and proxy

pip install -U "truvelocity-tevv[ui,proxy]"

Install the practice agent

pip install harbor-expense-agent

Make the practice project with the offline model

tv practice-project --scripted

Go into it

cd expense-desk-practice

Start the agent and the proxy (leave this window open)

tv practice-agent

In a second window in the same folder: open the UI and follow TruVelocity-Onboard-a-New-Agent.html

tv ui --allow-runs --open

With a real model, use pip install "harbor-expense-agent[all]", leave out --scripted and put your provider and key in .env. With Docker Desktop, docker compose up -d --wait still works as before.

Not yet

Coming in later releases

Adding an agent from the UI · sign-off by several approvers · editing settings and calibrating evaluators in the UI · per-metric targets · comparing runs and shadow runs from Conduct · alerts by email or chat
Quality

How this release was checked

Lint (ruff, format, mypy strict, the core import rule) and the full suite: 1064 passed, none failed (24 pilot tests that need Docker were skipped). The practice route was run with no Docker and no LLM key, from a clean environment: the agent's wheel and TruVelocity installed, then tv practice-project --scripted and tv practice-agent. The clean build passed 15 of 15 cases and the release bar. The buggy build (loose_meal_limit, phantom_notice, silent_write) failed 6 cases with 4 critical defects, and the release bar failed. harbor-expense-agent 0.1.0 was installed from PyPI into a fresh environment and handled an event end to end.