Protenix v1 remote tool

TOU validation and deployment status (2026-08-16)

Protenix 1.1.0 (ByteDance’s AlphaFold3 reproduction) ran on the NVIDIA GB10 (aarch64, cc 12.1) with the optimization kit OFF — its one prebuilt object is x86-64 and kit_foreign_objects refuses per directory, so the kit mode is unreachable here. Verified on FOUR CONSECUTIVE IDENTICAL REQUESTS plus a mixed-shape sequence and an ascending ladder: 4.4-5.1 s per 35-residue fold, peak reserved ~1.9 GiB at 100 residues rising to 9.65 at 800 — about 5.1x, the same shape protenix_v1 measures (5.3x) for the same architecture through a different reimplementation. One thing to know before deploying: stock’s fast-LayerNorm extension cannot run on this card and fails LATENTLY. It emits SASS for sm_70..sm_100 with no PTX, so on sm_121 the call returns, torch.cuda.synchronize() passes, and the NEXT unrelated torch operation raises no kernel image is available for execution on the device — with a LayerNorm output in hand that was never written. [kit.env] LAYERNORM_TYPE = "torch" is upstream’s own switch and is load-bearing; the worker refuses to start if an operator overrides it back on a card outside the extension’s built set. Public publication, cross-user isolation, concurrency, recovery and structural accuracy are unvalidated; keep this deployment private.

• Operation: protenix_v1_fold

• Start: python -m tooluniverse.remote.protenix_v1.protenix_v1_tool

• Endpoint: http://127.0.0.1:8046/mcp

• Provider configuration: this provider is on the remote-tool standard, so the model runs in its own environment built from a pinned lock, and the weights are declared with a size and a sha256 and verified before the worker starts. Keep provider data, weights, caches and credentials outside Git. Use one worker: the resident worker serves one request at a time by design, and a second in-flight request queues on its lock rather than doubling peak VRAM. The measured VRAM budget is 9.65 GiB; read the envelope it was measured at in the manifest before co-tenanting anything else on the card, because a larger request exceeds it.

• TOU check: tu doctor --forward http://127.0.0.1:8046/mcp --json

• Private relay: tu serve --share --forward http://127.0.0.1:8046/mcp --name validation-protenix_v1 --workers 1

Non-loopback binding requires TOOLUNIVERSE_API_TOKEN; otherwise keep the server on loopback. The relay requires TOOLUNIVERSE_SERVICE_KEY. Public publication and independent-caller testing were not run.

New-user check: python scripts/remote_validation/setup_skill_preflight.py --implementation protenix_v1. Add --check-provider-env before launch, --live after launch, and --check-connect-prereqs before sharing. Live preflight checks exact MCP discovery only; it does not run or validate a model.

The authenticated 2026-08-16 Platform matrix found all 30 private owner relays online and all 41 operations discoverable. It imported each implementation as unpublished owner draft(s), configured a 120-second timeout, and invoked it through /expert-sessions/{id}/test. This implementation was added after that matrix and was not part of it, so none of that evidence covers it: it has not been exercised through the Platform at all, and everything above was measured locally, on one card, through the loopback endpoint. Public publication, independent-caller authorization/isolation, broad saturation and persistent supervision are unvalidated here.

Authorize once, then share with one short command

Log in once per machine, and again after key rotation:

tu remote login

tu remote login requests a short-lived device code, opens the TU Platform approval page and polls until the signed-in user approves; no key is copied or pasted, and the CLI never displays the computer-only key it stores in a local 0600 file. On a headless machine add --no-browser. Then each private share is one command:

tu remote share protenix_v1

Override the reviewed name and worker count only when needed:

tu remote share protenix_v1 --name my-protenix_v1-remote --workers 1

In an interactive terminal, sharing starts the same browser flow when the key is missing, expired, or revoked. A malformed or revoked explicit TOOLUNIVERSE_SERVICE_KEY fails fast rather than being silently replaced; unset or correct it and run tu remote login again. Non-interactive jobs also fail fast. tu remote logout removes only the local copy; tu remote logout --revoke revokes the computer-only platform connection first and leaves the server record offline for owner inspection.

See the complete setup and verification guide. The skill is the whole guide for this provider – how to build the two environments, what tu remote setup’s five steps each prove, every refusal name with its fix, and the measured numbers a host schedules on. This file carries the deployment contract and deliberately does not restate it, because two copies of a procedure drift and only one of them gets corrected.

Getting the structures out

return_structures: true returns the mmCIF as text, and that works only for small complexes. A tool result is capped at 100,000 characters by smcp.py, which then falls back to RAW STRING truncation – it cuts the JSON mid-string and hands the caller something that does not parse, rather than a short answer or a named error. A 131-residue monomer measured 102,306 characters, so the very first thing most callers fold already exceeds it. The tool now refuses a structure above that limit with a message naming the way through, instead of letting the transport corrupt the reply.

The way through is out_dir: an absolute server-side directory the predictions are written to, returned as structure_paths. That is also the right shape when one provider’s output is another’s input – it keeps a megabyte of mmCIF off the wire entirely. apnovo_generate_designs already works this way, returning a run_dir its own evaluation reads, and the apnovo -> protenix_v1 -> apnovo chain measured on 2026-10-15 used out_dir for exactly this reason.

Overview

Protenix 1.1.0 all-atom structure prediction for a protein, nucleic-acid, ligand and ion complex

Weights and kit are pinned: src/tooluniverse/remote/protenix_v1/provider.toml declares the upstream kit by commit, the environment by lock file, and every weight file with its byte count and digest. Nothing is fetched at request time. The measured peak VRAM for this host is written to ~/.cache/tooluniverse/measured/protenix_v1.json, and that per-host file – not the number in the manifest – is what a scheduler should read.