Docs › MTS Lab
The agent
Your own Claude Code, working in the project with you through the Lab's tools — a research partner that discusses, edits code you review, runs only when asked and commits only when you say so.
The chat
Each project (and library) has an Agent tab: one continuous conversation, kept in your browser, with the input at the bottom. New chat starts over. The agent is Claude Code on the computer the workspace is on, signed in with your own account; MTS Lab adds nothing in between.
What it can do
- Read: the project's files, outputs and history, uncommitted changes, experiments and their comparisons, per-day cfgs and logs, notebooks, the engine's graph of a cfg, the engine's classes and the SDK, the data format.
- Change: write and delete project files (to the working copy — they show as uncommitted changes), move a plugin into a library.
- Run: experiments and reruns, on this computer or a server — when you ask.
- Remember: when PriorStates Memory is installed, findings go to the project's research journal.
It cannot reach your shell, files outside the workspace or the web: only the Lab's tools are offered.
How it works with you
- Incrementally. It discusses first — what the data could refute, look-ahead, overfitting, r² that will not become P&L — then edits a small step at a time and says what changed where.
- Runs only when asked ("run fit", "rerun the last experiment from sim"), reports the numbers and compares with earlier experiments rather than from memory.
- Commits only when you say the change is settled — or use Review & commit yourself.
Models
The menu by the input picks the model for the chat: Fable (most capable: hard design and research questions), Opus (deep work: conversions, multi-file changes), Sonnet (fast and balanced: most edits), Haiku (fastest, fewest tokens: small questions and lookups), Claude Code's default, or any model name. Switching keeps the conversation. ↻ refresh checks which models your Claude Code can run now, with one tiny turn each, and says why the others cannot (for example, a newer Claude Code needed). After each reply a small line shows the model, tokens and turns.
Things to ask
Why did the held-out r² drop in the last experiment compared with the one before? Add a 60 s horizon to the trends and basis, then run vargen and fit only. The maker loses to fees on day 7: look at its fills and say what you would change. Convert notebooks/explore.ipynb into the engine and make parity pass. Get Binance USD-M BTCUSDT trades and book tickers for 2024-01-15 into data/. Add a variable for the spread in bps to qlib, with a test.