If a personcan use it,your agent can too.
Automate work across APIs, browsers, desktops, and legacy portals with agents that can see, act, verify, recover, and ask for approval before consequential changes.
Live automation run
RUN-1843 · Inventory reconciliation
Detected discrepancy
SKU-1048 inventory mismatch
API value
0
Storefront
In stock
Correction approved
Set available inventory to 18
Inspect
Shopify inventory read
Compare
Storefront visual check
Approve
Inventory correction
Verify
Read-back matched
Use the best path
Call an API when one exists, run deterministic code when the rules are fixed, and use vision only when the work lives behind a screen.
MCP + APIs • code • computer use • model judgment
Verify the work
Check that every click, form submission, and data mutation produced the intended result before the agent marks the task complete.
Visual diff • read-back • retry • checkpoint
Govern every outcome
Put sensitive writes behind human approval and preserve screenshots, decisions, tool calls, cost, and results as run evidence.
Approvals • evidence • audit • permissions
Performance you can inspect, not a demo claim.
The published graded web benchmark runs real browser tasks against known answers. Release governance turns benchmark and session evidence into a regression gate, then certifies the exact policy, runtime, audit state, and result.
95.8%
Graded web accuracy
23 of 24 independently graded tasks
$0.094
Average model cost
Measured per benchmark task
3
Execution engines
NSR, Claude, and OpenAI
4
Isolation tiers
Container, gVisor, microVM, and Kata
Everything between a click and a trusted outcome
StateSet adds routing, verification, recovery, isolation, approvals, and release governance around computer-use models so they can run real operations.
See what operators see
Read dashboards, tables, fine print, dialogs, and legacy applications through screenshots and zoomed visual inspection.
The same agent can move from a structured API response to a portal screen without handing the workflow to another bot.
Act across every surface
Click, type, scroll, drag, use keyboard shortcuts, run code, edit files, and call typed MCP tools from one reasoning loop.
API-first execution keeps routine work fast; computer use closes the gaps where no reliable integration exists.
Recover instead of stalling
Visual verification, stuck detection, retries, circuit breakers, checkpoints, and self-healing keep long tasks moving.
The runtime distinguishes a completed action from a click that landed but changed nothing.
Scale in isolation
Run concurrent jobs in separate sandboxes with warm pools, tenant limits, scale-to-zero, and configurable isolation.
Each job gets its own desktop and execution context so parallel automation does not mean shared-session risk.
Keep humans in control
Require approval for consequential writes, show the proposed change, and verify the result after execution.
Tool guards, permissions, prompt-injection defenses, and fail-closed production checks enforce the operating boundary.
Deploy in your cloud
Run the platform in Azure, AWS, GCP, or your own Kubernetes environment with your identity, secrets, and data controls.
Terraform, Helm, keyless workload identity, pluggable storage, and multi-provider inference support enterprise deployment.
Understand, route, act, verify, and govern
Every task follows the same accountable path, even when it crosses APIs, local code, browser tabs, and desktop applications.
Turn an operating request into a bounded objective.
1. Understand
The agent reads the task, gathers only the context it needs, identifies the systems involved, and defines what a successful outcome must look like.
- Route commerce, support, onboarding, and general tasks to reusable skills.
- Inspect current desktop and application state before acting.
- Preserve attention with just-in-time retrieval and structured memory.
The task starts with an explicit objective and completion condition, not an open-ended instruction to click around.
Automate the work between your systems
The strongest computer-use workflows combine clean integrations with the screens, portals, and exception paths APIs cannot reach.
Catalog operations
Operational problem
Product data is spread across APIs, spreadsheets, admin screens, and storefront pages, so API-only checks miss what customers actually see.
How StateSet solves it
Audit catalog hygiene through APIs, use code for batch analysis, inspect storefront reality with computer use, and route proposed fixes through approval.
Outcome profile
- Find corrupt SKUs, missing categories, junk tags, and inventory mismatches.
- Normalize brand taxonomy and structured product attributes.
- Reconcile system-of-record values against the live storefront.
Build around outcomes, not brittle click scripts
Traditional RPA works when the path never changes. StateSet combines adaptive reasoning with the operational controls required for dynamic interfaces and exception-heavy work.
Put one manual workflow on an evidence-backed automation path.
Choose a repetitive, measurable workflow. StateSet maps the API and GUI steps, defines the approval boundary, runs a graded pilot, and gives you the evidence to decide what should scale next.
Pilot success path
Define the outcome and a human-reviewed gold set
Map APIs, deterministic steps, and GUI-only gaps
Set approval, permission, cost, and isolation policy
Grade accuracy, evidence quality, and cost per outcome