Validation status (2026-10-02): BindCraft's binder-hallucination step on ColabDesign 1.1.3 and AlphaFold-Multimer v3 ran on the NVIDIA GB10 (aarch64, cc 12.1, driver 580.173.02) under the optimization kit's
exactmode, on jax/jaxlib 0.6.0 with CPython 3.10. It needs no alignment and contacts no external service at request time: the design step feeds AlphaFold the target's backbone as a template and a single sequence for the binder, so there is nothing to search or fetch. Only the design step runs — no MPNN redesign, reprediction, PyRosetta relaxation or filtering — so the sequences it returns are a hallucination's output, not filtered final designs. Nobody holds the model: upstream clears the device and rebuilds it for every trajectory, so what a held worker keeps is the imports, the CUDA context and the compile cache. This skill is UNVERIFIED: the acceptance ladder, the 400-token row and the example run through the live endpoint were not completed (stopped by the host memory guard). Measured so far: the smoke case is byte-identical between the kit's stock design body, the resident worker underoffand underexact(design.pdb sha256 d2a404a3...), peak_pool 4.0 GiB at 100 and 200 tokens. Public publication, cross-user isolation, concurrency, recovery and scientific accuracy are unvalidated; keep this deployment private.
Prerequisites
- Run from the ToolUniverse repository root on Linux with Python 3.12.3.
- A CUDA GPU is required. This provider holds its worker resident between requests.
- The model environment needs Python 3.10 (
uv venv --seed --managed-python --python 3.10.18 ~/colabdesign-venv). - Keep provider data, weights, caches, and credentials outside Git.
- Bind to loopback. A non-loopback bind requires TOOLUNIVERSE_API_TOKEN; never put it in arguments or results.
Run the standard-library contract check before downloading large dependencies:
python scripts/remote_validation/setup_skill_preflight.py --implementation colabdesign
After exporting provider resources, add --check-provider-env. After the
server starts, add --live to verify the exact MCP tool set without running
the model. Before sharing, add --check-connect-prereqs; this reports only
whether a key is set and never prints its value.
Create the two environments
This provider is on the remote-tool standard, so it has two environments, not one: the MCP server runs in ToolUniverse's, and the model runs in its own, built from a pinned lock. They cannot import each other, and that is the point.
python3 -m venv .venvs/colabdesign
. .venvs/colabdesign/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .
python -m pip install fastmcp
The model half lives at the path TOOLUNIVERSE_COLABDESIGN_PYTHON names, and
nothing above creates it. Create a Python 3.10 interpreter there first, and
tu remote setup --only environment fills it from
src/tooluniverse/remote/colabdesign/environment/requirements.lock with
--no-deps, so no version is silently resolved to something else.
The kit pins its stack as conda packages, which do not install on aarch64, so the
lock is that stack expressed as PyPI distributions at the kit's own versions. The
index was checked per file: jaxlib 0.6.0 publishes a
cp310-cp310-manylinux2014_aarch64 wheel, as do jax-cuda12-plugin 0.6.0 and
(as one py3-none file) jax-cuda12-pjrt 0.6.0, and every CUDA library the kit
pins has an aarch64 file at that version except NCCL, whose conda .1 build
suffix has none (2.30.7 is used; one device never runs a collective).
BindCraft's DSSP executable is an x86-64 binary, and it cannot run on aarch64.
run.sh install fetches it from BindCraft's repository at the pinned commit
and digest-checks it; on this architecture it fails with Exec format error. The
worker probes it at start-up and, when it cannot run, sets BindCraft's own
optimise_beta setting false (the setting that calls DSSP on the stage-1
structure). The reply says optimise_beta. On a host where DSSP runs, BindCraft's
default is kept.
The rest is built by tu remote setup, which reads
src/tooluniverse/remote/colabdesign/provider.toml and runs five steps:
preflight, environment, kit, weights, smoke. Each refuses by name and carries
its fix.
export TOOLUNIVERSE_COLABDESIGN_PYTHON=~/colabdesign-venv/bin/python
export COLABDESIGN_PARAMS_DIR=~/.colabdesign
tu remote setup colabdesign
Run one step at a time with --only preflight|environment|kit|weights|smoke.
The smoke step is the one that matters: it holds the worker, sends one design
request, asserts the provider produced what its manifest says a successful run
produces (a design.pdb plus a non-empty sequences), and records the measured
peak VRAM to ~/.cache/tooluniverse/measured/colabdesign.json. A host schedules
on that file.
Package/network-dependent commands must be rerun in a clean environment before marking this skill complete.
The input
The tool takes the target as PDB text (target_pdb), its chain, the
binder_len and optionally hotspot residues in BindCraft's own grammar: 56,
31,33,57, A31,A33 or A1-10, in the PDB's own residue numbering. A hotspot
that names a residue the PDB does not hold is refused by upstream
(positions 999 and chain A not found) and the request fails; nothing is written.
The ceiling is tokens, and the worker enforces it. Target residues plus binder length may not exceed 400, the row the VRAM budget was measured at. Peak memory and time both grow with it, so crop the target to the region around the site you want to bind, which is also BindCraft's own advice.
Every distinct (target residues, binder length, recycles) triple is a new set of
XLA programs, and the first request of one compiles for minutes. The reply says
first_of_engine_key. A trajectory that fails BindCraft's own pLDDT gates is
still a completed request: it returns its best sequence with
accepted_by_stage_gates: false and the gate named in gate_fired. Judge the
design, not the reply.
The optional iteration counts (soft_iterations, temporary_iterations,
hard_iterations, greedy_iterations) are BindCraft's own advanced settings,
defaults 75 / 45 / 5 / 15, bounded here. Upstream hard-codes the first 50 logit
iterations, so soft_iterations below 50 is not accepted.
Obtain credentials, data, and model weights
- No credentials.
COLABDESIGN_PARAMS_DIR(default~/.colabdesign) holds 1.7 GiB in five declared entries,params/params_model_{1..5}_multimer_v3.npz, each with a size and a digest. They are DeepMind's AlphaFold parameters (CC BY 4.0), fetched by the kit's ownpython -m colabdesign_opt.weightsfromhttps://storage.googleapis.com/alphafold/alphafold_params_2022-12-06.tar(a HEAD request answers 200, 5,587,968,000 bytes), and never from a mirror. - No alignment, no MMseqs2 server and no template database is needed or hidden. Nothing in the request path touches the network.
- The only other thing
tu remote setupfetches is BindCraft's DSSP executable, once, at install, digest-checked. - Nothing is fetched at request time.
Authorize once, then share with one short command
After installing dependencies, exporting the provider resources above, and installing the pinned relay SDK described under Connect below, run from the repository root. Log in only once per machine (and again after key rotation):
tu remote login
# Or import an existing protected 0600 file without sourcing it:
tu remote login --env-file /path/to/tooluniverse-service.env
Then each private share is one short command:
tu remote share colabdesign
By default, 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
copy/paste is required. On a headless machine, add --no-browser and open the
printed link elsewhere. The CLI exchanges the approval for a computer-only key,
verifies /remote-servers/preflight, stores it in a local 0600 config file, and
never displays it.
The share command runs environment and TU Platform preflights, starts or reuses the exact loopback endpoint, validates discovery, and keeps the relay in the foreground until Ctrl-C. Override the defaults only when needed:
tu remote share colabdesign --name my-colabdesign-remote --workers 1
Use tu remote check colabdesign for a non-sharing readiness check and
tu remote run colabdesign for a local-only foreground server.
In an interactive terminal, sharing automatically starts the same browser flow
when the key is missing, expired, or revoked. A malformed or revoked explicit
TOOLUNIVERSE_SERVICE_KEY fails fast instead of being silently replaced; unset
or correct it, then run tu remote login. Non-interactive jobs also fail fast.
Use tu remote logout to remove only the local copy. Use tu remote logout --revoke to revoke the computer-only platform connection first; the server record remains offline for owner inspection.
Start and verify locally
python -m tooluniverse.remote.colabdesign.colabdesign_tool
The Streamable HTTP endpoint is http://127.0.0.1:8054/mcp. In a second activated shell run:
python - <<'EOF'
import asyncio
from fastmcp import Client
async def main():
async with Client("http://127.0.0.1:8054/mcp") as client:
print([tool.name for tool in await client.list_tools()])
asyncio.run(main())
EOF
Confirm discovery contains colabdesign_design; stop on empty, duplicate, or schema-drifted discovery.
Connect to ToolUniverse Connect
The tuplatform-connect relay is not yet published on PyPI. Install the reviewed public wheel below; its SHA-256 is pinned. Interactive sharing uses browser device authorization, so no key copy/paste or GitHub access is required.
python -m pip install fastmcp pyyaml "tuplatform-connect @ https://connect.aiscientist.tools/downloads/tuplatform_connect-0.3.0-py3-none-any.whl#sha256=3fad5eee5ecf7887a693d93ccd1aa112dc0955617a885d1fc3daded0030f9ae0"
tu doctor --forward http://127.0.0.1:8054/mcp --json
tu serve --share --forward http://127.0.0.1:8054/mcp --name validation-colabdesign --workers 1
Prefer browser device authorization. For CI or migration, supply
TOOLUNIVERSE_SERVICE_KEY only through a protected environment or use
tu remote login --manual-key; never put a key in shell arguments.
This implementation was added after the authenticated 2026-08-16 Platform matrix and was not part of it. That matrix imported each implementation it
covered as unpublished owner drafts and invoked them through
/expert-sessions/{id}/test; this provider was imported as no such draft and
invoked through no such route. It has not been exercised through the Platform
at all: the evidence above is local, on one card, through the loopback
endpoint. Public publication, independent-caller authorization/isolation,
saturation, and persistent supervision are untested.
Run a verified example
Operation: colabdesign_design
The example is a deliberately tiny case so it answers in a minute or two: a 20-residue backbone-only crop of the PD-L1 target in BindCraft's own bundled example (chain A, residues 46-65, hotspot 56) and a 12-residue binder, with only the 50 logit iterations upstream hard-codes. It exercises the whole path; it is not a design anyone should expect to pass BindCraft's gates.
{"target_pdb": "ATOM 223 N LYS A 46 42.677 -51.097 -27.336 1.00 30.62 N \nATOM 224 CA LYS A 46 43.737 -51.751 -28.084 1.00 33.55 C \nATOM 225 C LYS A 46 43.367 -53.198 -28.413 1.00 34.94 C \nATOM 226 O LYS A 46 43.812 -53.733 -29.424 1.00 36.17 O \nATOM 232 N GLN A 47 42.542 -53.793 -27.559 1.00 35.11 N \nATOM 233 CA GLN A 47 42.044 -55.155 -27.717 1.00 38.92 C \nATOM 234 C GLN A 47 40.532 -55.115 -27.597 1.00 35.82 C \nATOM 235 O GLN A 47 40.013 -54.557 -26.651 1.00 37.51 O \nATOM 241 N LEU A 48 39.839 -55.740 -28.541 1.00 34.86 N \nATOM 242 CA LEU A 48 38.396 -55.871 -28.486 1.00 33.75 C \nATOM 243 C LEU A 48 37.991 -56.747 -27.270 1.00 32.17 C \nATOM 244 O LEU A 48 38.291 -57.937 -27.261 1.00 34.34 O \nATOM 249 N ASP A 49 37.377 -56.147 -26.247 1.00 29.17 N \nATOM 250 CA ASP A 49 36.835 -56.894 -25.104 1.00 29.89 C \nATOM 251 C ASP A 49 35.384 -57.302 -25.342 1.00 27.26 C \nATOM 252 O ASP A 49 34.469 -56.551 -25.039 1.00 25.90 O \nATOM 257 N LEU A 50 35.166 -58.512 -25.836 1.00 27.29 N \nATOM 258 CA LEU A 50 33.796 -58.946 -26.135 1.00 26.96 C \nATOM 259 C LEU A 50 32.886 -59.041 -24.905 1.00 27.30 C \nATOM 260 O LEU A 50 31.676 -58.839 -25.015 1.00 27.10 O \nATOM 265 N ALA A 51 33.446 -59.312 -23.726 1.00 26.91 N \nATOM 266 CA ALA A 51 32.610 -59.468 -22.535 1.00 28.07 C \nATOM 267 C ALA A 51 31.873 -58.183 -22.187 1.00 27.80 C \nATOM 268 O ALA A 51 30.744 -58.228 -21.689 1.00 28.66 O \nATOM 270 N ALA A 52 32.463 -57.045 -22.559 1.00 27.21 N \nATOM 271 CA ALA A 52 31.887 -55.734 -22.245 1.00 25.64 C \nATOM 272 C ALA A 52 31.019 -55.140 -23.350 1.00 23.78 C \nATOM 273 O ALA A 52 30.189 -54.278 -23.054 1.00 26.12 O \nATOM 275 N LEU A 53 31.174 -55.625 -24.576 1.00 20.26 N \nATOM 276 CA LEU A 53 30.510 -55.050 -25.758 1.00 19.00 C \nATOM 277 C LEU A 53 29.093 -55.580 -25.988 1.00 18.32 C \nATOM 278 O LEU A 53 28.859 -56.795 -26.014 1.00 18.18 O \nATOM 283 N ILE A 54 28.177 -54.653 -26.220 1.00 17.55 N \nATOM 284 CA ILE A 54 26.876 -54.957 -26.764 1.00 17.55 C \nATOM 285 C ILE A 54 26.633 -54.103 -28.000 1.00 16.09 C \nATOM 286 O ILE A 54 26.825 -52.894 -27.947 1.00 16.82 O \nATOM 291 N VAL A 55 26.199 -54.738 -29.086 1.00 15.13 N \nATOM 292 CA VAL A 55 25.818 -54.058 -30.315 1.00 15.27 C \nATOM 293 C VAL A 55 24.373 -54.374 -30.630 1.00 14.36 C \nATOM 294 O VAL A 55 24.024 -55.547 -30.721 1.00 14.45 O \nATOM 298 N TYR A 56 23.555 -53.331 -30.811 1.00 13.69 N \nATOM 299 CA TYR A 56 22.116 -53.468 -30.995 1.00 13.66 C \nATOM 300 C TYR A 56 21.723 -52.695 -32.246 1.00 13.92 C \nATOM 301 O TYR A 56 21.989 -51.521 -32.334 1.00 14.70 O \nATOM 310 N TRP A 57 21.157 -53.377 -33.224 1.00 13.95 N \nATOM 311 CA TRP A 57 20.607 -52.725 -34.414 1.00 14.19 C \nATOM 312 C TRP A 57 19.111 -52.999 -34.500 1.00 13.68 C \nATOM 313 O TRP A 57 18.671 -54.159 -34.333 1.00 13.85 O \nATOM 324 N GLU A 58 18.334 -51.967 -34.832 1.00 13.55 N \nATOM 325 CA GLU A 58 16.912 -52.113 -35.054 1.00 13.97 C \nATOM 326 C GLU A 58 16.495 -51.300 -36.260 1.00 15.21 C \nATOM 327 O GLU A 58 17.196 -50.380 -36.684 1.00 14.76 O \nATOM 333 N MET A 59 15.322 -51.651 -36.797 1.00 16.36 N \nATOM 334 CA MET A 59 14.682 -50.807 -37.834 1.00 19.55 C \nATOM 335 C MET A 59 13.403 -50.356 -37.132 1.00 19.31 C \nATOM 336 O MET A 59 12.612 -51.231 -36.759 1.00 17.59 O \nATOM 341 N GLU A 60 13.194 -49.056 -37.006 1.00 19.50 N \nATOM 342 CA GLU A 60 12.153 -48.499 -36.124 1.00 21.07 C \nATOM 343 C GLU A 60 12.346 -49.170 -34.761 1.00 19.54 C \nATOM 344 O GLU A 60 13.380 -48.905 -34.121 1.00 18.62 O \nATOM 350 N ASP A 61 11.407 -50.016 -34.319 1.00 18.18 N \nATOM 351 CA ASP A 61 11.535 -50.673 -33.010 1.00 17.05 C \nATOM 352 C ASP A 61 11.775 -52.169 -33.169 1.00 16.99 C \nATOM 353 O ASP A 61 11.821 -52.926 -32.165 1.00 16.77 O \nATOM 358 N LYS A 62 11.916 -52.618 -34.418 1.00 16.15 N \nATOM 359 CA LYS A 62 12.090 -54.069 -34.696 1.00 17.08 C \nATOM 360 C LYS A 62 13.570 -54.450 -34.624 1.00 15.39 C \nATOM 361 O LYS A 62 14.352 -53.924 -35.408 1.00 14.38 O \nATOM 367 N ASN A 63 13.904 -55.370 -33.723 1.00 15.36 N \nATOM 368 CA ASN A 63 15.298 -55.736 -33.568 1.00 15.15 C \nATOM 369 C ASN A 63 15.815 -56.537 -34.741 1.00 15.86 C \nATOM 370 O ASN A 63 15.115 -57.419 -35.233 1.00 16.08 O \nATOM 375 N ILE A 64 17.013 -56.187 -35.189 1.00 16.06 N \nATOM 376 CA ILE A 64 17.704 -56.923 -36.236 1.00 16.96 C \nATOM 377 C ILE A 64 18.882 -57.673 -35.660 1.00 17.31 C \nATOM 378 O ILE A 64 18.998 -58.856 -35.915 1.00 18.04 O \nATOM 383 N ILE A 65 19.735 -56.991 -34.886 1.00 17.16 N \nATOM 384 CA ILE A 65 20.955 -57.579 -34.324 1.00 17.58 C \nATOM 385 C ILE A 65 20.987 -57.359 -32.817 1.00 16.65 C \nATOM 386 O ILE A 65 20.706 -56.261 -32.339 1.00 16.26 O \nEND\n", "chain": "A", "binder_len": 12, "hotspot": "56", "seed": 0, "name": "example", "soft_iterations": 50, "temporary_iterations": 0, "hard_iterations": 0, "greedy_iterations": 0}
Invoke it through the live local MCP endpoint:
python - <<'EOF'
import asyncio
import json
import re
from fastmcp import Client
text = open("skills/setup-colabdesign-remote-tool/SKILL.md").read()
arguments = json.loads(re.search(r"~~~json\n(.*?)\n~~~", text, re.DOTALL).group(1))
async def main():
async with Client("http://127.0.0.1:8054/mcp") as client:
print(await client.call_tool("colabdesign_design", arguments))
asyncio.run(main())
EOF
The example has NOT been executed through the live endpoint yet; no result is quoted here.
Tune GPU and concurrency
- Use one worker. The resident worker serves one request at a time by design; a second in-flight request queues on its lock rather than doubling peak VRAM.
- Peak pool reads 4.0 GiB at 100 and 200 tokens (peak in use 2.26 and 3.24 GiB); the 400-token row is not yet measured.
idle_timeout_sis 300: with the kit'sexactpersistent cache a restarted worker's first call took 16-26 s where a cold one took 175-252 s.XLA_PYTHON_CLIENT_PREALLOCATE=falseis set by the launch and is load-bearing for scheduling: left at its default JAX claims most of the card up front and every peak this provider reported would measure the preallocation instead of the model.- Measure cold start, then at least four consecutive identical requests, then a mixed-shape sequence (small, small, large, small), then a request that changes the recycle count and back. A repeat and a shape change are different tests.
- Then parallel levels 1, 2, 4, 8, and only 16 if memory permits.
- Record successes/errors, p50/p95, peak RAM/VRAM, utilization, queueing, cancellation cleanup, and recovery.
- Increase workers only after single-flight initialization and sanitized recoverable OOM/timeout behavior are proven.
Troubleshoot and clean up
- Read the refusal. Every step of
tu remote setupnames what is wrong and carries the fix; exit 3 is a refusal and exit 1 is a failure. - The worker's own transcript is at
~/.cache/tooluniverse/remote-logs/colabdesign-resident.log. grep -c "ready device"that log before believing a warm number. From the supervisor's side a worker's death between requests is invisible, so a "warm" timing taken across one is not warm.
| refusal | what it means | what to do |
|---|---|---|
env_python_unset |
TOOLUNIVERSE_COLABDESIGN_PYTHON is not exported |
the refusal prints the recipe |
lock_install_failed |
the pinned stack did not install | read the tail; check the interpreter is Python 3.10 |
kit_install_failed |
run.sh install failed: pip, the DSSP fetch, or the kit's pin check |
the printed line names which |
weights_off_pin |
a file is absent, short, or a different digest | the named file is left in place for inspection; remove it by hand and re-run --only weights |
smoke_workerstarterror |
the worker died before binding | read the transcript; the kit's own NOT ACTIVE: ... names an inherited COLABDESIGN_OPT value |
jax reports no GPU device |
the model interpreter sees no card | check the driver and jax.default_backend() in that interpreter |
<N> tokens ... exceeds this provider's 400 |
target residues + binder length is past the budget's row | crop the target; the limit is the measured envelope |
refused by the kit: ... |
the kit's own named refusal before any model is built | the text names it (settings, DSSP, params file) |
positions <N> and chain <C> not found |
the hotspot names residues the PDB does not hold | use the PDB's own residue numbering |
- Import/executable failure in the model half: re-run
tu remote setup colabdesign --only environment. It rebuilds from the lock with--no-deps. - 401/403 on deliberate network binding: configure matching TOOLUNIVERSE_API_TOKEN bearer auth; prefer loopback plus relay.
- Stop server/relay with Ctrl-C. If installed, run
tuplatform-service uninstall --name validation-colabdesign. - A worker outliving its server is a known shape: it has an idle timeout and an orphan watchdog, and it exits when its parent becomes init. Check with
nvidia-smi --query-compute-apps=pid,used_memory --format=csvbefore assuming the card is free. - Revoke temporary keys. After confirmation, remove only .venvs/colabdesign and the provider's own cache directory; never use a broad recursive target.
Use only official upstream documentation linked by the implementation README; do not substitute third-party model mirrors.