Making Every CameraIntelligent.

Build, evaluate, and operate vision systems from versioned data, specialist models, VLMs, and production workflows.

Connect via MCP

Give your agents sight.

Score Studio is the computer vision layer for agents—automating data generation, annotation, training, evaluation, workflows, and deployment in one evidence-backed system.

Available now · API key required

npx add-mcp https://mcp.scorestudio.ai/mcp --name score --header "Authorization: Bearer $SCORE_API_KEY"

The installer detects supported agents and lets you choose where to connect Score.Create a connection key

01 · Orchestrate

Workflows

Compose data, specialist models, VLMs, decisions, and outputs on one canvas.

02 · Understand

Data & annotation

Version visual evidence, label it precisely, and expose the hard examples.

03 · Adapt

Models & VLMs

Use pretrained intelligence, fine-tune specialists, and compare every candidate.

04 · Execute

Workloads

Run durable annotation, evaluation, training, and multimodel jobs with full control.

05 · Operate

Production

Release to cloud or edge, observe live behavior, and route failures back into learning.

Built on visual evidence

Real scenes. Inspectable outputs.

See the world your cameras see, shape what they understand, and carry that intelligence into production.

Object detection

Urban interchanges

12.8k frames · 9 classes

Instance segmentation

Dock operations

4.2k frames · 7 zones

Tracking & counting

Storefront analytics

8 live cameras · monitored

Start where you are

Bring what you have. Build what has never existed.

Begin with proven data and models, or start from an unsolved vision problem. Both paths become one coherent production system.

Use what is ready

Move directly to system design.

Select an existing dataset and a model with runnable weights, verify its measured performance, then compose and deploy the workflow.

01Choose versioned data
02Select a runnable model
03Evaluate, build, and deploy

See the production loop

Create from first principles

Go beyond the model catalogue.

Upload or generate data, annotate it, train directly, or engage Model Foundry to develop a frontier model for a problem that off-the-shelf systems cannot solve.

01Create and version data
02Train directly or commission a frontier model
03Build, deploy, and monitor

Start building

The production loop

One project. Every stage. No lost context.

Score Studio organizes computer vision around the lifecycle of the system—not a collection of disconnected tools.

01Data

02Annotate

03Train

04Evaluate

05Workflow

06Deploy

07Monitor

Data

Upload or generate images, review annotations, and preserve immutable versions.

Datasets · annotation · health · versions

Models

Use ready weights, train a model, or engage Model Foundry against a defined eval contract.

Lab · training · Model Foundry · artifacts

Evidence

Inspect predictions, ground truth, failure modes, class metrics, and measured latency.

Evaluation · benchmarks · release gates

Systems

Compose processing logic visually, select infrastructure, and observe production behavior.

Workflow canvas · providers · deploy · monitor

Workflow canvas

Compose the entire vision system visually.

Configure inputs, select runnable weights, transform predictions, add logic, connect providers, and define outputs without losing the underlying technical contract.

Typed inputs and outputs
Ready-made and custom model blocks
Provider and deployment blocks

Image input

image

Vehicle detector

runnable model v3

Count objects

class · confidence · zone

Event output

provider destination

Built for CV engineers

Inspect the evidence behind every release.

Dataset structure, annotation quality, training behavior, evaluation evidence, and runtime infrastructure remain first-class technical surfaces—not hidden steps in a wizard.

Dataset intelligence · urban interchanges v3

Know what the model will learn.

Start with your data

images

601

annotations

100%

coverage

duplicates

Annotation area

n=601

Annotation centroid

n=601

Annotation dimensions

n=601

wider →↑ taller

Asset dimensions

n=60

wider →↑ taller

Class distribution4 classes

car364

truck164

bus39

motorcycle34

Resolution: 1.08 MP
Density: 7.0 median
Object scale: 0.36%
Orientations: 3

Standalone annotation studio

Boxes, polygons, class shortcuts, model assist, review queues, and annotation import/export in a viewport-safe workbench.

label · assist · review

Training telemetry

Configure presets and advanced parameters, then inspect live loss, per-class metrics, artifacts, logs, and evaluation gates.

train · compare · approve

Providers, edge, and SDK

Connect inference, compute, and storage; manage edge devices and keys; automate the same lifecycle from Python or REST.

connect · automate · deploy

Model Foundry · frontier vision on demand

When the right model does not exist, we build it.

Commission Score's frontier computer vision lab directly from the platform, 24/7—without waiting for an introduction, RFP, or lab slot. Define success upfront and approve only the production-ready model that clears it.

Start a vision brief See Foundry proof

Open 24/7

Submit the task, evals, and budget directly—no introduction or RFP

Frontier lab on demand

Specialists adapt, distill, or build what the task requires

Proof before release

Approve the benchmark result, then open-source the model or deploy it privately

Example development brief · complex vision system

Multimodal damage inspection

Detect subtle surface damage, read identifiers, reason across multiple views, and produce an auditable severity decision at the edge.

Frontier project

≥ 97.0%

Target recall

≤ 80 ms

Edge latency

Critical misses

Define the unsolved problem

Scene, failure modes, output schema, runtime, and commercial constraints.

Lock the evaluation contract

Ground truth, test cohorts, target metrics, latency, and delivery budget.

Develop frontier candidates

Architecture research, foundation-model adaptation, distillation, and task-specific training.

Approve the production artifact

Compare candidates on the same evals; release only the model that clears the contract.

Release principle

Nothing ships—and no development budget is released—until you approve a model that clears the agreed evaluation thresholds.

Customer-controlled release

Pretrained VLMs · domain fine-tuning

Start capable. Make it exact for your world.

Choose Satori, Qwen, InternVL, Cosmos, or bring your own pretrained VLM. Fine-tune it with reviewed, versioned multimodal examples using LoRA, QLoRA, or full SFT—then prove the specialist on your own evaluation set before it enters a workflow.

StartPretrained multimodal foundationAdaptDomain examples · LoRA / QLoRA / SFTReleaseEvaluated specialist · versioned and deployable

Inspect Satori

Need a capability no existing foundation can reach? Model Foundry remains the separate path for frontier architectures and from-scratch research programs.

Prompt

Which dock lane is obstructed, and why?

Satori · structured answer

Lane 03 · pallet inside the marked safety zone · confidence 0.94

Phrase grounding

Connect language to regions and objects in an image.

Visual question answering

Ask operational questions and retain the visual evidence.

Dense captioning

Describe scenes, relationships, changes, and anomalies.

Structured reasoning

Return typed answers that workflow blocks can act on.

Agent-ready infrastructure

Let your agents build vision systems—with evidence.

Connect agents to typed, scoped tools over MCP. Read-only evidence access is available now; tracked execution and release tools stay visibly unavailable until their side-effect and approval contracts ship.

Connect your agent Copy the connection command

Agent conversationEvidence mode

Is warehouse-detector v12 ready to release?

Inspected dataset version 8

Compared evaluation run ev_921

Checked deployment health

Release blocked

Small-object recall decreased by 4.8%. The current production revision remains healthier.

Available now

Understand your system

Inspect datasets, model lineage, evaluation runs, and deployment state through scoped read tools.

Coming soon

Execute real work

Start tracked training, evaluation, and workflow runs while preserving exact versions and durable run IDs.

Coming soon

Prove before release

Compare compatible evidence, expose regressions, and verify deployments before describing them as production-ready.

Trust is part of every tool call

Scoped access Immutable provenance Explicit side effects Measured evaluations Deployment verification No silent releases

Works where your team already works

Channels appear as available only after their package and hosted OAuth connection are public.

Remote MCP · API-key scoped

CClaude CodeAvailable

CCursorAvailable

MCPMCP clients with headersAvailable

CChatGPTComing soon

CClaudeComing soon

GGeminiComing soon

GGitHub CopilotComing soon

WWindsurfComing soon

Score Studio

Build the complete vision system.

From the first image to a measured, deployable, continuously improving production loop.

Open the studio Already have a workspace? Sign in

Making Every CameraIntelligent.

Give your agents sight.

Score Studio product capabilities

Workflows

Data & annotation

Models & VLMs

Workloads

Production

Real scenes. Inspectable outputs.

Urban interchanges

Dock operations

Storefront analytics

Bring what you have. Build what has never existed.

Move directly to system design.

Go beyond the model catalogue.

One project. Every stage. No lost context.

Data

Models

Evidence

Systems

Compose the entire vision system visually.

Inspect the evidence behind every release.

Know what the model will learn.

Annotation area

Annotation centroid

Annotation dimensions

Asset dimensions

Standalone annotation studio

Training telemetry

Providers, edge, and SDK

When the right model does not exist, we build it.

Multimodal damage inspection

Start capable. Make it exact for your world.

Let your agents build vision systems—with evidence.

Understand your system

Execute real work

Prove before release

Build the complete vision system.