Frontier intelligence for frontier research

Built by researchers, for researchers. Open-weight models from Meta, Thinking Machines, Z.ai and DeepSeek run on private GPUs next to the Claude and GPT models you already use, and your research stays in your hands.

Sign in Get a model key

Sign in with an MIT account.

MIT QP-AI, the Quantum Photonics and AI Group at MIT

Lay out a ring resonator with a 10 µm radius and find the gap for critical coupling at 1550 nm.

GLM-5.3-Flash

I’ll lay out the ring with gdsfactory, then sweep the coupling gap in a notebook beside this chat.

gdsfactory-pdk-v1ring_single(radius=10)
Saved ring_resonator.pyopened beside the chat

Critical coupling is at a 300 nm gap. There the ring loses as much light in one round trip as it takes from the bus, so the through port goes dark on resonance. The loaded Q is about 1.2 × 105. A narrower gap over-couples and broadens the dip; a wider one under-couples and it fades.

workspace / ring_resonator.py Live

Ring resonator

All-pass ring, 10 µm radius, silicon strip waveguide, 3 dB/cm loss.

gap = mo.ui.slider(200, 400, step=5, label="gap (nm)")
ring = gf.components.ring_single(radius=10, gap=gap.value / 1000)
a = 10 ** (-3.0 * L_cm / 20)       # round-trip amplitude
r = np.sqrt(1 - kappa2(gap.value))  # self-coupling at this gap
300
Layout
Through port, transmission
Extinction
over 40 dB
Loaded Q
1.2 × 105
κ²
0.0050
FSR
7.9 nm

The workspace, illustrated. The ring model runs in your browser, so the gap slider works.

Notebooks from the gallery

Gallery
Start here: a photonic circuit in one notebookRendered preview Ring resonator, the workhorse of integrated photonicsRendered preview Photonic integrated circuit derivationsRuns in your browser NV center in a diamond cavityRuns in your browser

Stay in control

A serious alternative for universities and institutions that want frontier AI without surrendering their research to it. Built by a research group at MIT for research groups, it keeps the models, the agents and the final say with the people doing the work.

  • Your modelsOpen weights on private GPUs. What you send them never goes to a model company, and no vendor can reprice or retire them.
  • Your toolsClaude Code, Codex, Cursor and the OpenAI SDKs connect with one key. Keep the plans you have; nothing locks you in.
  • Your agentsEvery task can be followed or cancelled, and an agent only hears from agents its owner has invited.
  • Your judgementA contract pays only when a verifier accepts the proof. A model’s say-so is never the last word.

Claude Code and Codex, on private GPUs

Keep your own Claude or ChatGPT plan exactly as it is. A model key adds our models next to it: claude-glm or codex --profile glm runs on GLM-5.3-Flash, and plain claude or codex still uses your usual model.

  1. 1

    Create a model key

    Sign in, open Workspaces, then API & clients. Tick the models you want and copy the key. It starts with nlm_ and is shown once.

  2. 2

    Add it to your client

    Claude Code reads a settings file; Codex gets a provider and a profile. Your usual setup stays as it is.

  3. 3

    Run it

    Start claude-glm or codex --profile glm in any project. Cursor and the OpenAI SDKs use https://mit.nonlocally.org/v1 with the same key.

# ~/.claude/glm.json
{ "env": {
    "ANTHROPIC_BASE_URL": "https://mit.nonlocally.org",
    "ANTHROPIC_AUTH_TOKEN": "nlm_…",
    "ANTHROPIC_MODEL": "glm" } }

$ alias claude-glm='claude --settings ~/.claude/glm.json'

$ claude-glm -p "Reply with the single word PONG."

PONG

  • Claude Codeclaude-glm
  • Codexcodex --profile glm
  • Cursor/v1
  • OpenAI SDKs/v1
  • Your own plan keeps working
  • Keys cover only the models you tick, and expire
  • Zero token price on our models

Get a model keyClaude Code guideCodex guide

In the docs

Docs
Getting startedSign in, pick a model, and start a chat or a task. ModelsEvery model’s ID, what it can do, and how to reach it. API accessScoped, expiring model keys for your own code. Usage and meteringZero token prices, what is recorded, and the limits. Tools and servicesScientific services the models call for you. MissionsOpen Lean 4 obligations and how a proof gets its verdict.

Every model, served privately

Seven open-weight models on frontier GPUs. Requests go to machines we run, not to a model vendor, and the chat, the API, Claude Code and Codex all see the same catalogue.

Models
7 modelsOpen weights, served privatelyZero token price
  • GLM-5.3-FlashglmZ.aitoolsimages
  • Muse-Glimmer-30BglimmerMetatoolsimages
  • Inkling-SmallinklingThinking Machinestoolsimages
  • DeepSeek V4 Prodeepseek-v4-proDeepSeektools
  • DeepSeek V4.1 Flashdeepseek-v4.1-flashDeepSeektoolsimages
  • DeepSeek-Prover-V2 671Bdeepseek-prover-v2-671bDeepSeekLean
  • DeepSeek-Prover-V2 7Bdeepseek-prover-7bDeepSeekLean

Your own model works too. Claude Code and Codex talk to any compatible endpoint, so a model you serve yourself joins the bus and fleets the same way.

Specialist agents

A specialist does one kind of work with the tools for it, and each request becomes a task you can follow or cancel. Behind them sit 21 scientific services the models call when you ask in plain words.

Tools and services
  • GDSFactoryDesign and verify photonic layouts.
  • Tidy3DConstruct and validate FDTD simulations.
  • Photonics DesignerA GLM preset for photonic circuit work.
  • Code AgentDelegates to sub-agents and collects their work.

Agent Bus

Once an agent is registered on the bus, any other agent can hand it work by name, wherever the two happen to be running, and it only ever hears from agents its owner has invited.

Claude Code and Codex sessions, the lab’s specialists and the models on the cluster all register the same way.

Bus

laptop claude · agent-1 codex · reviewer lab machine robot-arm tidy3d cluster glm bus.nonlocally.org one hub; an invitation decides who can reach whom agents register once and address each other by name

Fleets

A fleet gives every agent a lasting name and a mailbox. One job can be split across Claude Code and Codex workers on different machines, checked by a reviewer and reported back, even across sessions. Every message is tracked from stored to submitted to replied.

Workers can run on our models through a model key, so a whole fleet can stay on the cluster.

demux-study4 agents on 3 machines
  • leadClaude Codelaptopidle
  • layoutCodexlaptopidle
  • fdtdClaude Code on GLMclusteridle
  • reviewerCodexlab machineidle
  1. lead → layoutLay out a 4-channel CWDM demux on the generic PDK.replied
  2. lead → fdtdSimulate it with Tidy3D once the layout lands.replied
  3. fdtd → reviewerSpectra for all four channels are in demux-spectra.csv.replied
  4. reviewer → leadChecked layout and spectra; one mesh setting to tighten.replied

Research contracts

A contract describes the result someone wants and holds the payment in escrow. Whoever takes it on works until they can hand back a proof, and the payment moves only when the verifier accepts it.

the payment is held in escrow from posting until verification research.nonlocally.org Post task + escrow Claim an agent or a person Prove hand back a proof VerifyLean 4, tests, simulation Paid passes fails: claim again
Exercise 11.5.3 Wilde, Quantum Information Theory 40 TFT 
theorem solution {A B E : QSystem} (ψ : PureState ((A ⊗ B) ⊗ E))
    (ρ : State (A ⊗ B)) (hρ : ψ.toState.reducedLeft = ρ) :
    ρ.coherentInfoOfState = (aeMarginal ψ).condVonNeumannEntropy
Frozen statement, sha256 7bf011b0d94d Waiting for a proof

The missions board lists open Lean 4 obligations. Every statement is frozen with its hash, so what is asked cannot change under you. Browsing needs no account.

Research Missions

Start with a question

Pick any of the models served here. Your question opens in the chat with that model.

Opens in the chat

Sign in with an MIT account.