We’re building Lor to do the work of an entire AI lab. Today, its agents write code, run experiments, and iterate on the results for days, on any machine, hundreds at a time. Now in private beta.
With agents doing the hands-on work, one person can take on what used to need a team, and a team what used to need an institution. People make the calls: what to try, and what to trust. We built Lor for that way of working, and Mirendil runs on Lor every day.
Agents in Lor operate for days on any machine, bring up GPUs when they need them, and split parallel work across dozens of agents, even hundreds.
Lor runs the frontier models you already use, like Claude Opus, Claude Fable, GPT-6 Sol, and GPT-6 Astra, as well as open-source models like Kimi, DeepSeek, and GLM. It’s also the home for Mirendil’s own models. Fern, the first, holds up to 10M tokens of context, enough for days of an agent’s work at once.
What you can hand off
- “Run these 12 ablations overnight and write a report on the results.”
- “Audit every file in this repo for security bugs and open a PR for each fix.”
- “Reproduce Table 2 of this paper on 8 GPUs and explain where our numbers differ.”
Massively multi-agent
Run hundreds of agents at once
A lot of the work of building AI models is parallel in nature: reading thousands of agent traces for failures, conducting hundreds of experiments, building training environments from real pull requests. In Lor, an agent can split one job across dozens of agents, and hundreds across your projects. You can follow every agent’s progress and stop any of them at any time.
HowEvery chat scales independently, and the whole stack is built for many streams at once.
Compute included
Agents provision their own compute
Lor comes with its own compute.
Agents request nodes, set them up, run the work, and release them when they’re done. No cloud account. No setup.
That’s what lets Lor build AI end to end. Start an experiment, a training run, or an eval sweep. The agent brings up the nodes, spreads work across them, and hands them back when the results are in.
HowAgents allocate compute on demand and get full control of every VM and node they bring up.
Long-running work
Agents keep working for days, even weeks
Lor is built for durable, asynchronous work. When something breaks, agents notice and unblock themselves instead of waiting for you.
Lor can turn raw user feedback into shipped fixes, babysit training runs overnight, and keep automations going for weeks.
HowThe agent runs in the cloud, and its commands run on the machine you pick. Pick a Lor-managed cloud computer or any always-on machine, and closing your laptop doesn’t stop anything.
Any machine
Agents work on any machine
Most agents can only work on the machine they run on. Agents in Lor run in the cloud and can work on any machine you want: Lor-managed cloud computers, your laptop, or any other machine you already have.
We built Lor so both you and your agents can use the right machine for every task. Set up a team of agents from your laptop, send training runs to different GPU nodes, and keep up with it all from your phone.
HowIn Lor, the agent and the machine it works on are separate. The agent decides what to run, and Lor sends each command to the machine you chose. Agents only operate on machines you’ve added, and you can remove one at any time.
A frontier AI lab for every scientist
AI is transforming almost every scientific field. The most ambitious projects behind it still take a frontier lab’s worth of effort. With Lor, we want to massively accelerate every scientist and engineer, so building AI is no longer the bottleneck.
We’ve seen how far this can go. Recently, a single Mirendil researcher pretrained Mirendil-Base, a competitive coding model, with agents handling almost all of the engineering. A project that once needed an entire team can now be driven by one person in three and a half months. Every scientist deserves that kind of leverage. We’re building Lor to be the best way to build AI.
Availability
- Access
- Currently in private beta, by invite only. Join the waitlist.
- Platform
- macOS, Linux, and web. Windows, iOS, and Android coming soon.
- Pricing
- Free during the private beta.
- Models
Mirendil Fern (10M-token context) Anthropic Opus 5.5, Fable 5.1, Sonnet 5.5, Haiku 5.5, Opus 5, Fable 5 OpenAI GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol, GPT-6 Luna Google Gemini 4 Argon xAI Grok 4.7, Grok 4.6 Meta Muse Spark 1.3, Muse Wonder Z.ai GLM-5.3, GLM-5.3-Flash, GLM-5.2, GLM-4.7-Flash Moonshot AI Kimi K3.1 Pro, Kimi K3 DeepSeek DeepSeek 4.1 Flash, DeepSeek 4 Pro
Lor is in private beta
We’re building Lor to carry the work of an entire lab, and we’d love to work with the people who’ll push it hardest.