@leumas/lmx

Sixteen little languages with one runtime. A script's extension picks the interpreter — nothing to configure and no project file. hello.lmx is the full language; rename it hello.lmm and the same runtime reads it as a matrix machine.

Public face: lmx.leumas.tech — a browser sandbox that runs the local modes with no account, and the lmx CLI as a single downloadable file.

The sixteen

That table is generated from src/host/interpreters.catalog.js, which is also what the API's /api/lmx/interpreters route, the Studio picker and INTERPRETERS.md read. There is one list.

Three tiers, and the difference is not politeness

Local (lmx lmy lmz lmm lmc lmn lmg) — no host access at all. These run anywhere: the browser sandbox, the CLI, the API.

Install-backed (lmq lmai lma lmi) — their vocabulary is a running Leumas. lma has no verb table: its verbs are every adapter tool, every stored functioncall, every automation action, every MCP tool and every capability the install currently holds, so registering an adapter makes existing scripts able to call it with nothing rebuilt. These need an identity, because they read somebody's data, spend somebody's model credits and call somebody's tools. They are refused by name in the public sandbox and by the CLI unless it is pointed at an install with --studio.

Privileged (lms lmk lmd lmnet lmt) — these touch the host. Admin role and an explicit env flag (LMX_ENABLE_SHELL / LMX_ENABLE_GUI / LMX_ENABLE_NET), both off by default. That is a security control, not a paywall — LMX is free.

Running a script

Three entry points, and the difference between the first two is the whole design:

import { runScript } from '@leumas/lmx';        // a WORKER: isolated, hard deadline, terminate()
import { runLocal }  from '@leumas/lmx/local';  // IN-PROCESS: no deadline, no bridges, streams to stdout
import { INTERPRETERS } from '@leumas/lmx/catalog';  // the catalogue alone — no express, browser-safe

runScript is right for a server running somebody else's script: the timeout is the only real bound on a while (true), and it never rejects — a syntax error is {ok: false, error}, which is what lets a caller tell a failed script from a request that never arrived.

runLocal is right for a command: a person running their own script on their own machine wants stdin, a TTY, an exit code and no five-second cap. It is also the only shape a single-file build can carry.

[warning] Import the leaves, not the barrel, unless you serve HTTP. . re-exports createLmxRouter, which imports express — free inside the API, and a whole web framework inside a CLI binary. @leumas/lmx/local and @leumas/lmx/catalog carry neither.

HTTP

The public pair is a separate router (createPublicLmxRouter), not the authenticated one with a softer guard: createLmxRouter turns every route into an open one when handed a non-function requireAuth, and making that conditional would arm exactly that trap. Its bounds are inside the router, not around it — a narrowed mode list, no bridges and no adapter(), a hard 3s deadline, a source cap and a concurrency cap. It is mounted only where the role declares publicSandbox, so a customer's appliance never carries it.

The bridges are FACTORIES, (user) => bridge, built per request. One bridge built at mount time and shared would hand every user the first user's permissions, key and wallet.

The CLI

lmx hello.lmx                  # the extension picks the interpreter
lmx -e 'print(6 * 7)' -m lmx   # a string
lmx modes                      # every interpreter and where it runs
lmx --studio https://api.leumas.tech --token <t> report.lmq

Shipped as one file per machine (pnpm lmx:binary), built by the same Node SEA lane as LeumasNode. Exit codes are a contract: 0 ran, 1 the script or the request failed, 2 you called it wrong.

Traps

• An lmai reply is not proof a model ran. providerRegistry.resolve() never fails — auto

falls back to an echo stub. A shaped ask on the stub echoes the prompt, which contains the SHAPE example, so a flawless-looking classification can be your own example handed back. stub: true and tokens: 0 are the only tells, and the runner prints a warning by itself.

• lmi must never answer "yes" by default. No entitlement engine wired means unknown, never

true. And entitled ≠ fulfillable: CAN and PROVIDERS FOR are different questions.

• {{ $C1 }}, not {{C1}}, in a template cell — the bare form renders the literal text.
• The in-script adapter() builtin is unmetered. lma's USE charges against the same

tool:<ref> target the HTTP route does; the older builtin does not, and changing it would alter existing scripts' behaviour.

Verify

pnpm smoke:lmx (every safe mode, every bridge, every gate, docs drift) · pnpm smoke:lmx-public (the anonymous lane: what it refuses) · pnpm smoke:lmx-cli --binary (the shipped binary, which is the only thing that sees the SEA defects) · node --test shared/engines/lmx/src/test/*.test.js.

Agent skill: leumas-lmx. Per-language references: docs/{lma,lmi,lmq,lmai,lmg}.md.