Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Bentoml · mkurman | DecimalAI

mkurman/bentoml

SkillScore77Maintained

BentoML — model serving and deployment. Build prediction services from any ML framework with OpenAPI/Swagger. Containerize, deploy to Kubernetes, AWS, GCP, Azure. Adaptive batching and GPU support.

AI/MLby mkurman

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/mkurman-bentoml/SKILL.md

decimalai skills pull mkurman-bentoml --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/mkurman-bentoml/SKILL.md https://app.decimal.ai/s/mkurman-bentoml/SKILL.md

Raw file: /s/mkurman-bentoml/SKILL.md ↗ · machine-readable summary at /s/mkurman-bentoml.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 8d ago

Yes — +9% lift-19% tokens0% turns77% of whole cases pass with skill (17/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-03	✗→✓	▲ Improved	-53%	0%
case-06	✗→✓	▲ Improved	-56%	0%
case-17	✗→✓	▲ Improved	-73%	0%
case-05	✓→✗	▼ Worse	7%	0%
case-07	✓→✓	＝ Same ✓	20%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

mkurman / zorai

skills/scientific-skills/bentoml/SKILL.md

View on GitHub ↗

License MITImported 8/3/2026Version v1

Skill contents

SKILL.md

Overview

BentoML packages ML models with service definitions, dependencies, environment config, and deployment targets into a portable "Bento." Deploy to Kubernetes (Kserve, Seldon), AWS SageMaker, GCP Vertex AI, or as a standalone Docker container.

Installation

bash
uv pip install bentoml

Service Definition

python
import bentoml
from bentoml.io import JSON
import numpy as np

iris_clf = bentoml.sklearn.get("iris_model:latest")

@bentoml.service
class IrisClassifier:
    def __init__(self):
        self.model = iris_clf.to_runner()
        self.model.init_local()

    @bentoml.api(input=JSON(), output=JSON())
    def classify(self, input_data):
        result = self.model.run(np.array([input_data["features"]]))
        return {"class": int(result[0]), "probabilities": result[1].tolist()}

Build & Deploy

bash
bentoml build      # creates a Bento
bentoml containerize iris_classifier:latest  # Docker image
docker run -p 3000:3000 iris_classifier:latest

References

Want to see if mkurman/bentoml would activate on your phrasing? Try it before you install — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/21/2026, 08:55 UTC · 22 cases

Lift

+9%

77% vs 68%

Agent turns

1.0 vs 1.0

Tokens

-19%

1,385.45 vs 1,719.23

fail → pass · 3pass → pass · 14fail → fail · 4pass → fail · 1

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 I'm setting up a machine learning inference service with BentoML. Could you write a Python file that imports a saved sklearn model, configures a service class with local runner initialization, and sets up a JSON-in, JSON-out prediction route?	fail→fail	14,234	10,343	-27%	1	1	0%	2,812	2,471	-12%	0	0	—
case-02 How do I install BentoML using uv, and what Python code structure should I use to wrap my scikit-learn classifier into an API endpoint that returns model predictions and probabilities as JSON?	fail→fail	12,499	8,886	-29%	1	1	0%	2,660	2,079	-22%	0	0	—
case-03 I manage Python environments using uv and want to install the model serving framework. Standard pip is slow in my workflow. What command installs the package via uv?	fail→pass	7,039	2,162	-69%	1	1	0%	1,241	583	-53%	0	0	—
case-04 I saved a scikit-learn model in my local model store named 'classifier_model'. When writing my serving script, should I load it directly using joblib.load or use the framework's specialized model store getter?	fail→fail	11,648	8,948	-23%	1	1	0%	1,997	1,879	-6%	0	0	—
case-05 After fetching my saved sklearn model from the local model store, should I pass inference inputs directly to model.predict, or convert the model reference into a runner instance first?	pass→fail	12,711	11,289	-11%	1	1	0%	2,025	2,175	+7%	0	0	—
case-06 I am writing a Python class to encapsulate my inference logic. Rather than inheriting from an explicit Service base class, how should the service class be marked so the framework registers it?	fail→pass	11,785	3,407	-71%	1	1	0%	2,110	922	-56%	0	0	—
case-07 Inside the __init__ method of my service class, I assigned self.model = iris_clf.to_runner(). What additional method call on self.model is required right inside __init__ to prepare the runner locally?	pass→pass	3,755	2,600	-31%	1	1	0%	603	724	+20%	0	0	—
case-08 When decorating my service prediction method, I want to handle raw JSON inputs and outputs. Should I pass Pydantic BaseModel types to @bentoml.api, or use explicit I/O descriptor instances?	fail→fail	12,625	9,669	-23%	1	1	0%	2,342	2,056	-12%	0	0	—
case-09 Inside my API prediction method, the incoming input_data dictionary contains a 'features' key with a plain Python list. How must this list be converted before passing it to self.model.run?	pass→pass	8,329	4,959	-40%	1	1	0%	1,534	1,332	-13%	0	0	—
case-10 The runner's predict call returns a numpy array of probabilities alongside the class prediction. How should the probability numpy array be formatted before putting it into the return dictionary?	pass→pass	10,692	7,814	-27%	1	1	0%	2,072	1,600	-23%	0	0	—
case-11 I finished writing my service script and service config. What exact CLI command bundles these definitions into a deployable Bento archive?	pass→pass	3,988	3,090	-23%	1	1	0%	658	751	+14%	0	0	—
case-12 I built a Bento archive named 'iris_classifier:latest'. What CLI command converts this built Bento into an OCI-compliant Docker image?	pass→pass	2,888	2,739	-5%	1	1	0%	477	778	+63%	0	0	—
case-13 I containerized my service into a Docker image 'iris_classifier:latest'. What docker run command starts the container while exposing the default service port 3000 to local port 3000?	pass→pass	3,571	2,555	-28%	1	1	0%	610	738	+21%	0	0	—
case-14 We want to deploy our model services onto Kubernetes. Which open-source Kubernetes CRD orchestration framework is officially named as a supported deployment target?	pass→pass	10,768	5,247	-51%	1	1	0%	1,794	1,166	-35%	0	0	—
case-15 Our company uses AWS for cloud ML infrastructure. Which managed AWS inference service is officially supported as a Bento deployment target?	pass→pass	5,215	3,248	-38%	1	1	0%	713	808	+13%	0	0	—
case-16 Our ML team operates on Google Cloud Platform. Which managed GCP machine learning service supports direct Bento artifact deployment?	pass→pass	7,150	3,871	-46%	1	1	0%	1,136	1,017	-10%	0	0	—
case-17 I need to configure JSON inputs for my service API endpoints. From which submodule should the JSON descriptor class be imported?	fail→pass	14,503	1,904	-87%	1	1	0%	2,269	616	-73%	0	0	—
case-18 When retrieving a saved scikit-learn model 'iris_model' from the store without knowing its specific version string, what tag string fetches the most recently saved version?	pass→pass	6,510	2,577	-60%	1	1	0%	1,121	729	-35%	0	0	—
case-19 Besides Kubernetes or managed cloud AI platforms, what standard lightweight execution mechanism can run a Bento service directly on a bare VM?	pass→pass	14,733	7,207	-51%	1	1	0%	2,488	1,503	-40%	0	0	—
case-20 I am training a PyTorch neural network and want to log epoch loss, validation metrics, and model weights to a central server. How do I set up MLflow tracking URIs and start experiment runs?	pass→pass	12,878	9,717	-25%	1	1	0%	2,615	2,354	-10%	0	0	—
case-21 How do I create custom asynchronous HTTP middleware in a FastAPI application to calculate and attach request processing latency headers?	pass→pass	11,361	9,120	-20%	1	1	0%	2,184	2,362	+8%	0	0	—
case-22 How do I write a config.pbtxt protocol buffer configuration file for Triton Inference Server to host an ONNX model with dynamic batching enabled?	pass→pass	12,392	7,654	-38%	1	1	0%	2,362	1,837	-22%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +9 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 68.2%
Pass rate with the skill: 77.3%
Cases attempted: 22
Cases counted toward lift: 22
Cases that regressed: 1
Skill version measured: v1
Run completed: 8/21/2026, 08:55 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-03

Prompt

I manage Python environments using uv and want to install the model serving framework. Standard pip is slow in my workflow. What command installs the package via uv?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run77

Live evalNeeds ≥20 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns-13% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 8 days ago

Other measured skills in the registry, with their headline benchmark lift.