LLMstudio is a monorepo that runs a proxy and a tracker as two services

Two flags, two Swagger ports, both on 0.0.0.0

The whole runtime is one command with two flags:

llmstudio server --proxy --tracker

Each flag turns on its own service with its own OpenAPI browser, and they do not share a port. With `--proxy` the Swagger page is at `http://0.0.0.0:50001/docs` on the default proxy port, and with `--tracker` it is at `http://0.0.0.0:50002/docs` on the default tracker port. Those are default ports, so a configuration file elsewhere has to be the thing that moves them.

Worth noticing before you run this on a shared machine: the addresses in the documentation are all-interface addresses, not `127.0.0.1`. A Swagger endpoint reachable on every interface is a local convenience that becomes an exposure the moment the port is forwarded, so treat binding as something to set deliberately rather than inherit.

The feature list behind those two services is what the project is selling: proxy access to models from OpenAI, Anthropic and Google, custom or local open source models through Ollama, a prompt playground, a Python SDK, monitoring and logging, LangChain integration, batch calls, and routing with fallback for availability. Type casting is on the same list and is marked as coming soon, so it is a promise rather than a feature today.

Extras decide whether you get the proxy and the tracker

Two pip commands, and the difference between them is which parts arrive:

pip install 'llmstudio[proxy,tracker]'

pip install llmstudio

The first is the full version with the proxy and tracker extras, the second is the lightweight core. Since those extras are the same components the `--proxy` and `--tracker` flags start, a core only install is the case where the flags will have nothing to run, so the choice has to be made before the first start rather than after a failed one.

The installation section also suggests creating and activating a new environment with conda first. That is advice rather than a requirement, and it matters because the supported interpreter range in the project metadata is Python 3.10 up to but not including 3.15.

Keys go in a .env in the directory you launch from

Configuration is a single file, and the rule is positional rather than a documented search path:

OPENAI_API_KEY="sk-api_key"
ANTHROPIC_API_KEY="sk-api_key"
VERTEXAI_KEY="sk-api-key"

The README is specific that the `.env` file is created at the same path you will run LLMstudio from, which means the working directory is part of the configuration. Move your shell somewhere else and the keys are not where the process last found them. The repository carries a `.env.template` at the top level, which is the shape to copy from, and `.gitignore` is present so the real file is not meant to be committed.

Three variables cover the hosted providers the feature list names, and nothing here covers a self hosted endpoint: for local or custom models the route described is Ollama rather than a base URL setting. So a deployment that mixes a local model with two hosted ones needs at least one of the three keys, and the README does not say what happens when only one is present.

package-mode is off, so the root project is not distributed

The project file is named `llmstudio-monorepo`, sits at version 1.0.0 and carries `package-mode = false`. That flag is the important line: the root is a workspace, not something that gets built and uploaded. The distributable pieces are the four libraries under `libs/`, and the development dependency group points at four local paths in develop mode:

llmstudio = { path = "libs/llmstudio/", develop = true }
llmstudio-core = { path = "libs/core/", develop = true }
llmstudio-tracker = { path = "libs/tracker/", develop = true }
llmstudio-proxy = { path = "libs/proxy/", develop = true }

The single console script the workspace declares is `llmstudio = "llmstudio.cli:main"`, which is the command the quickstart uses. The build backend is poetry-core, the only non development dependency constraint is the Python range, and the keywords list covers ml, ai, llm, llmops, openai, langchain and chatgpt, which tells you which integrations the author expected this to be used from.

The licence file says MPL-2.0, the project file says MIT

Two claims about licensing sit side by side. The repository licence is MPL-2.0, and the metadata in the project file declares `license = "MIT"`. For a package you install from a registry that gap is not academic: which terms apply depends on the file you actually receive and redistribute, and the two are not the same licence.

There is a second mismatch of the same kind. The repository homepage is tensorops.ai, while the project file points at llmstudio.ai and the documentation field at docs.llmstudio.ai, and the README quickstart tells you to check out the docs page. So the project's own metadata advertises a domain the repository does not list as its home, and the person reading the source has to guess which one is canonical.

The author of the project file is named as Cláudio Lemos with a tensorops.ai address, and the bug tracker points at the issues page of the repository.

Tests cover core and the main library, not the two services

Three Makefile targets is the entire developer entry point:

format:
	pre-commit run --all-files

unit-tests:
	pytest libs/core/tests/unit_tests

integration-tests:
	pytest libs/llmstudio/tests/integration_tests

The gap is where you would expect it. Unit tests live in `libs/core`, integration tests live in `libs/llmstudio`, and neither target mentions `libs/proxy` or `libs/tracker`, which are the two components you start as services on ports 50001 and 50002. Formatting is enforced through a pre-commit config rather than a Makefile-only step, so `make format` is the hook running all files.

Two other files at the root hint at what the services persist. `alembic.ini` is the migration tool configuration, which is how schema changes are managed for whatever the tracker stores, and `MANIFEST.in` controls what ends up in a source distribution. There is also a `deploy/` directory and a `docs/` directory, plus a `poetry.lock` for reproducible resolution.

The examples skip number five and the docs are marked coming soon

The `examples/` directory is the most complete documentation in the project. It holds notebooks for an introduction to the base, core, proxy and tracker versions of the library, an introduction with the proxy, one with tracking, an LLM comparison notebook, LangChain integration, a BigQuery integration, a Google Cloud guide and a LangGraph tutorial. Alongside them sit plain Python entry points: `core.py`, `sdk.py`, `llm_proxy.py`, `langchain_integration.py`, a `_config.yaml` and a README of its own.

Two things about that list are worth a second look. The numbering jumps from 04 to 06, so there is no fifth notebook in the tree, which suggests a file was removed or renumbered rather than never written. And the documentation bullet in the README is annotated as coming soon, while the quickstart points at docs.llmstudio.ai through a link whose target is the relative `docs` path, so following it lands in this repository's docs directory rather than on the hosted site.

Contributing goes through a CONTRIBUTING.md file and a Discord invite, and a training link points at an LLM Studio workshop page on the tensorops.ai site.

Editorial conclusion

Fit for a team that wants one interface in front of several model providers plus a record of what was sent, and for the LangChain or LangGraph side projects that the example notebooks already assume. A poor fit if you expected one process or one port, because the proxy and the tracker are separate services with separate Swagger endpoints on 0.0.0.0, and the prompt playground has no documented start command in the README. Before installing, decide whether you need the full extras or the core only, sort out the licence question by reading the LICENSE file rather than the pyproject, and bind those two ports to loopback if the machine is not yours alone.

Frequently asked questions

Official sources

1. Issues
2. License: MPL-2.0
3. Project website
4. README
5. TensorOpsAI/LLMstudio on GitHub