How it works

One agent.
No theatre.

LOLM is built for people who need code that runs and proof of what happened — not another chat that looks busy.

What you get

Multi-model raceOpening turns race GLM-class, gpt-oss-120b, and Llama 4 Maverick. Keep the candidate that compiles and runs.
Real sandboxWrite → run → fix in a network-isolated jail. Self-reported failures and broken checks block “shipped”.
Sealed receiptsEvery turn records files, commands, and control decisions in a hash-chained ledger you can audit.
NFET on codingAfter every run, LOLM measures latent dynamics of your code (or honest sandbox proxies) and can force verify, retrieve past receipts, branch to a new approach, or allow finalize — sealed into the receipt.

What NFET is

NFET is LOLM’s control layer. The five-stream graft (surface · latent SSM · regime · gate · memory) re-reads text and exposes entropy, hidden drift, gate balance, and regime entropy. A policy maps those to five actions: continue, retrieve, verify, branch, finalize. Chat has used this for a while. Coding now does too — so “shipped” is not model whim; it is measured dynamics plus a green sandbox.

What we deliberately hide

Lab dashboards live under More in the workspace. The product is the loop: describe a task, get working code, keep the receipt — with NFET decisions inside it.

Open models, no restraint on quality

Bring your own keys (Groq, Cerebras, OpenAI, Together, OpenRouter) and the same ensemble race runs on your account.

CLI

npx lolm-cli code "…" — same loop as the workspace, offline-friendly receipts.

Open workspace Public demo Pricing