Morpheus Lumerin Node

The purpose of this software is to enable interaction with distributed, decentralized LLMs on the Morpheus network through a desktop chat experience.

v7.0.0 — Full TEE capability. The v7 release completes a two-hop Trusted Execution Environment (TEE) trust chain for any model registered on-chain with the tee tag:

• Phase 1 — consumer → P-Node. A consumer proxy-router (v6.0.0+) cryptographically verifies the provider's P-Node runs the exact official hardened -tee image inside a genuine Intel TDX SecretVM, with TLS pinning, at session open and on every prompt.
• Phase 2 (new in v7) — P-Node → backend LLM. The v7+ P-Node itself verifies the backend LLM it forwards inference to (CPU TDX quote, TLS pinning, workload RTMR3 replay of the backend's docker-compose.yaml, CPU-GPU nonce binding, and NVIDIA NRAS GPU attestation) at startup and on every prompt.

Because Phase 2 runs inside the attested P-Node, any v6+ consumer is forward-compatible with a v7+ provider and gains the Phase 2 guarantees automatically — no client-side upgrade required. See the TEE reference, the SecretVM quickstart, and the TEE backend verification developer reference.

What this fork adds

Everything above is the official Morpheus network. This branch (stake-duration) adds a consumer desktop experience on top of it, in two areas:

Session duration and cost, actually under your control.

• Type how long you want the session, in plain language — "1 day", "2 years" — instead of picking from fixed chips. That typed length sets the stake directly and opens as ONE block, up to the network's per-session cap (7 days by default, read live from the contract, not hardcoded).
• Only a length longer than that cap chains into multiple blocks automatically, and only then does a renewal choice appear: Seamless (the next block opens before the current one ends — inference never pauses, but two blocks' stake is briefly locked at once) or Sequential (waits for the block to end first — a short pause, same total cost either way). A normal-length session never sees this; there is nothing to renew.
• Pick a specific provider, not just the router's auto-pick (auto-select itself is upstream's own long-standing default, not something added here). A model is often served by several providers at different prices and uptimes; the picker shows each one's online/offline state and its price, greys out any you can't afford, and — once picked — every renewal in a chained session targets that same provider rather than whatever the router would auto-pick next.
• Stake even if you can only afford some of a model's providers. The old gate priced against the model's most expensive provider and blocked you outright; it now gates on the cheapest and tells you "covers N of M providers" so a comfortably-affordable session isn't refused over a provider you were never going to use anyway.
• See who's actually serving you — the provider's endpoint and the MOR you actually have locked, not an abbreviated on-chain address and a cost figure that used to be computed wrong.
• A warning before an early close locks your MOR, naming the exact figure and until when — closing before a block's EndsAt time-locks that stake for ~24h on-chain, which is the network's behavior, not a bug in this app; you just weren't told beforehand. Held stakes past their lock are auto-claimed back for you.

Your terminal, not just the app window.

• grok and opencode integration — pin the Morpheus models you use in Settings → OpenAI-compatible API, and they appear in both tools' own model pickers, session or no session.
• No model ever decides to spend. Ask a pinned model with no session open and your terminal says so; this app's window comes forward with the price, the provider, and the duration, and a session opens only once you confirm it here — never from inside the agent's own turn.
• One command from a clean clone to a running build — see Build the desktop app from source below.

Full detail on each release lives in the app itself (Settings → What's new in this version) and in ui-desktop/src/shared/release-notes.ts.

Documentation

The canonical documentation lives at nodedocs.mor.org. Source files are in /docs and built with Mintlify. The site replaces the previous 00-overview.md / 02-*.md / 04-*.md / 99-troubleshooting.md set of files; old paths still resolve via redirects in docs/docs.json.

The site is structured around role-based journeys (consumer / prosumer / provider tiers), with anti-hallucination AI knowledge pages and curated mirrors of the broader ecosystem (mor.org, tech.mor.org, active.mor.org, MyProvider, Everclaw, NodeNeo, app.mor.org).

What's in this repo

• Lumerin proxy-router — background process that monitors blockchain contract events, manages secure sessions between consumers and providers, and routes prompts and responses between them.
• ui-desktop/ (MorpheusUI) — Electron desktop app for consumers. Release assets are *-morpheus-app-* installers (.dmg / .AppImage / portable .exe); on first launch the app downloads the proxy-router, a local llama.cpp + tinyllama demo model, and an IPFS (kubo) node. There is no zip / mor-launch package in current releases.
• cli/ — CLI client (*-morpheus-cli-* release binaries; local builds produce mor-cli).
• Standalone *-morpheus-router-* release binaries for headless / provider deployments.

End-to-end picture

0. PreRequisites: BASE Layer 2 Blockchain, MOR and ETH on BASE for staking and bidding.
1. Existing, Hosted AI model available for inference via the Morpheus network.
2. The proxy-router talks to and listens to the blockchain; it routes prompts and inference between providers' models and consumers.
3. Providers register their models via bids on the blockchain.
4. The consumer node is the "client" that purchases bids, sends prompts via the proxy-router, and receives inference back from the provider's models.
5. Consumers stake MOR to open a session for the duration they intend to use.
6. Once the session is open, prompt and inference (ChatGPT-like) can start.

Tokens and contract information

Item	BASE Mainnet	BASE Sepolia (testnet)
Chain ID	8453	84532
Branch	main (MAIN-* releases)	test (*-test releases)
MOR Token	0x7431aDa8a591C955a994a21710752EF9b882b8e3	0x5C80Ddd187054E1E4aBBfFCD750498e81d34FfA3
Diamond Marketplace	0x6aBE1d282f72B474E54527D93b979A4f64d3030a	0xA328196f2438DADA5ab729E39388D86896c27c85
Block Explorer	https://base.blockscout.com/	https://base-sepolia.blockscout.com/

You will need both MOR (for stake / fees / session payment) and ETH on BASE (for gas) in your wallet.

Quickstart

The linked pages below are the official Morpheus docs — accurate for wallet setup, staking, and how the network works, regardless of which fork built your binary. The "download from GitHub releases" step they describe does NOT apply to this fork: this repo's only release (desktop-v7.5.0-cc.1) predates everything below, carries no desktop-app asset at all, and its proxy-router binaries predate the model-health and attestation work this branch just merged in from upstream. Build from source instead — it is the one path on this fork that is actually current.

Role	Start here
Consumer (chat)	yarn app, then the Consumer quickstart from "Onboard" onward
Provider (host your own model)	cd proxy-router && make build, then the Provider quickstart from "Configure" onward
TEE provider (SecretVM)	SecretVM quickstart
Resale provider	Resale overview
Prosumer / agent	Prosumer overview
Developer (API)	API overview

Build the desktop app from source

Requires Node >= 20 — nothing else. One line builds an installer for the machine you are on and puts it in your Downloads folder:

(git clone https://github.com/cowboycoderhq/Morpheus-Lumerin-Node.git || (cd Morpheus-Lumerin-Node && git fetch -q && git merge --ff-only -q @{u} || true)) && cd Morpheus-Lumerin-Node/ui-desktop && NODE_OPTIONS=--dns-result-order=ipv4first npx --yes yarn@1.22.22 app

npx runs the exact Yarn release pinned in ui-desktop/package.json (packageManager) without installing anything globally, so there is no separate npm install -g yarn step and no drift toward whatever Yarn happens to already be on your machine. That single yarn app then installs dependencies, creates .env from .env.example if you do not have one, picks the right build target for your OS and CPU, and copies the finished installer to ~/Downloads. Nothing to choose and nothing to remember.

Safe to run more than once in the same spot. git clone into a directory that already exists fails outright ("destination path already exists") — when it does, the fallback branch cds into it and fast-forwards it to the latest commit itself (git fetch && git merge --ff-only; || true so a diverged or offline checkout still falls through to building whatever is there rather than stopping). This lives in the command itself, not only in a script inside the repo: a checkout old enough to predate a fix has no way to run that fix's own self-update code, so the update has to happen before yarn app is ever invoked, from a copy-paste that's always current because it comes fresh from this page every time. yarn app also re-checks and fast-forwards on its own once inside a checkout that already has that logic — redundant on a checkout old enough to need the fallback above, but keeps working for anyone invoking yarn app directly without this wrapper.

NODE_OPTIONS=--dns-result-order=ipv4first works around a real, fairly common failure: some VPNs and mesh networks (Tailscale among them) advertise a default IPv6 route that isn't actually reachable, and Node tries that route first when a registry host offers both an A and AAAA record — surfacing as yarn install dying mid-fetch with EHOSTUNREACH. This flag makes Node try IPv4 first without disabling IPv6. If yarn install still fails on EHOSTUNREACH or a registry timeout after that, the network you're on can't reach the npm registry at all right now — check your VPN/proxy, not this repo.

The build is unsigned, because signing needs an Apple Developer ID that only the publisher has. On macOS the first open is refused with "Apple cannot check it for malicious software" — right-click the app and choose Open to run it anyway. On first launch the app downloads its own services (proxy-router, IPFS, a local model), about 2GB, so give it a few minutes before the window is usable.

Install with yarn, not npm: yarn.lock is the committed, tested dependency tree and CI runs yarn install --frozen-lockfile against it. npm install resolves a different tree from the one that is tested.

Building a specific target by hand

yarn install
yarn build:mac-arm64        # or build:mac-x64 / build:win-x64 / build:linux-x64

The installer lands in ui-desktop/dist/. scripts/release.sh is the signed and notarized path and works only for the publisher.

Run it without building

cd ui-desktop && yarn dev

The checks

The desktop app carries a verification suite that runs in seconds and needs no network, no wallet, and no built app:

cd ui-desktop
yarn typecheck                      # both tsconfigs

# The harness is a separate package and uses npm — it has no yarn.lock.
cd tools/ui-verify && npm install
npm run logic                       # pure-logic + source invariants
npm run isolate                     # mounts real components in a browser, screenshots
npm run openai                      # drives the OpenAI-compatible endpoint end to end
npm run frozen                      # catches colours that cannot theme-swap

Run these before opening a PR. logic and openai are the fastest way to tell whether a change broke something you did not touch. None of them need a network, a wallet, or a built app.

Contributing

PRs should target dev, not main. Changes promote dev → test → main.

See CONTRIBUTING.md for the branch model, PR checklist, and local docs/proxy-router tips. The GitHub PR template reminds you to pick the right base branch.

For AI agents reading this repo

Start with AGENTS.md — hard rules, quick lookup tables, and ingestion instructions.

To load the full documentation corpus in one fetch:

Resource	URL	Use
Full corpus (preferred)	llms-full.txt	Complete markdown export — primary ingestion path
Page index	llms.txt	Titles, descriptions, and per-page .md links
Page Markdown	append .md or non-browser fetch	Same page URL returns clean Markdown when Accept lacks text/html
Docs MCP	/mcp	Search / retrieve pages without scraping HTML

See AGENTS.md for priority reading slugs and anti-hallucination rules.