Know what ships. Run it on your stack.
For each capability: the component your engineers receive, its default setting and what your team still configures, from the design scorer to the exported project code.
Up to five designs, scored and ranked.
The Nx·Studio analyst starts from your written context or one of eight starter packs. It proposes up to five designs, one per archetype: conservative, balanced, ambitious, compliance-first and cost-optimal. Each one lists its steps, integrations and sign-off gates, with an AI-estimated monthly run cost and estimated hours saved.
Rules you can rely on
Stated policies become fixed rule checks with no model involved, so identical input gives identical output. Each rule keeps its source document and stays a draft until a reviewer you name approves or amends it.
The Foundation, in detailGenerated from your context
Your systems, policies, data shapes and objectives feed each design; there is no fixed catalog of workers.
What the analyst reads
A 10-step wizard, chat, uploads of common document and spreadsheet files (up to 25 MB) and sample datasets of up to 200,000 rows. Existing process diagrams import as editable workflows, and a Discover view reads process-mining logs. What you share is used only to design your workers, once data-processing terms are in place.
Six-dimension scoring
A review step scores every design on six weighted dimensions, safety and compliance among them, and marks one as recommended. Any design failing validation is regenerated or dropped.
Requested changes
Optional: describe a change, such as “add a sign-off before payment”. The assistant drafts it on the canvas for you to accept or reject, and the editor keeps 100-step undo and named checkpoints.
Your architects compare the scored options, their integrations and their estimated run cost before anything is exported.
Mock mode before launch.
Each worker can get synthetic records fitted to its own data, or your imported samples. Mock mode answers every tool, connector and worker-to-worker call from four response packs (happy path, empty, error, timeout) without network access. It ships in the export too. Run one step, the full flow with real sign-off pauses, or many workers on an accelerated clock.
Rehearsal before reality
Studio test runs execute as a separate process limited to allowlisted settings, on synthetic or imported sample data, with external systems mocked. An optional hardened sandbox is available.
The Foundation, in detailPer-worker synthetic datasets
AI-generated records that match each worker's input fields, or imported samples of up to 200,000 rows.
Costed traces
Each test run writes a trace with token and cost totals per event, plus scores for accuracy, speed, robustness and coverage.
Version A/B runs
Compare two or three versions on one scenario. Deterministic record and replay of a run is available as an option.
Attack and multi-worker drills
Red-team testing is off by default and returns a scorecard without blocking export. The simulation lab, on by default, exposes conflicting updates and broken hand-offs between workers.
Your engineers read every trace and its per-event cost before any credential for a live system is issued.
A blueprint you can diff.
A versioned workflow graph defines each worker and ships as the workflow definition alongside the code. Engineers fork, tag, diff and three-way merge it; process owners see the same graph as swimlanes. With formal verification switched on, four properties of the graph are checked and a failure blocks export.
Blueprints you can read
Steps, sign-off gates and hand-off points live in the workflow definition, not in prompt text, so a code review covers them. Each agent's instructions are fingerprinted in the AI bill of materials, which ties every agent to the exact version it shipped with.
The Foundation, in detailFork, tag, diff, merge
Versions support fork, branch, tag and diff, a three-way merge and a genealogy view. Every test score stays linked to its source version.
Gates per step
Several independent approval gates per workflow, each assigned by role, team or variable. Any tool can require approval, and maker-checker blocks self-approval.
Traces and audit log
Every execution writes a run log with token and cost totals, with optional export to your monitoring stack. A tamper-evident, signed audit log is optional.
Read the audit-log specificationConfidence floors
A step scoring under its confidence floor logs a warning, or, if you configure it, parks the case for a person. The Health Analyzer flags weak approvals before export, and paused runs persist and resume exactly once.
Code review covers the logic itself: the graph, its rules and its sign-off gates, versioned like any repository.
A complete codebase you can audit.
Each export is a ZIP of the complete project. It holds a standard web service with its workflow runtime, agent and tool source, tests, a container definition, an example configuration file and a README. It must pass a build check before download and exposes run, health, readiness and metrics endpoints; access to each worker sits behind your own network controls. The code contains no Nxgyle endpoint or license check.
Review the sample export on the Trust CenterProof in every export
Every export carries a signed AI bill of materials listing models, agents, connectors, tool code and dependencies. The signature covers a hash of every exported file, and your team checks it offline with the included key and verifier. Each package also includes a component-level threat model plus a template for EU AI Act Art. 53 documentation.
The Foundation, in detailA reference catalog spanning thousands of APIs
Curated starter connectors (SAP, Oracle EBS, Workday, Salesforce, ServiceNow and more) ship without resolved endpoints, so your engineers complete each call. Catalog operations with a complete web-call definition run once credentials are set. API-description upload, a generic web-call step and an AI tool-server call step cover the rest.
Verifiable file hashes
Each exported file's hash sits inside the signed record, so a changed file fails offline verification.
Reviewer inbox to wire
Exports with approval steps add an inbox app (API, UI, migrations). Your engineers connect request creation and the resume call; Slack works, email is limited to one supported service, and Teams is not live yet.
Optional hardening
Personal-data masking is an optional module; a default export runs unmasked, with a warning. Injection guards, policy packs, a policy-based authorization step and drift detection are opt-in.
Targets and models
Optional deployment templates and CI files, plus deploy-target code for Snowflake, AWS, container clusters, Azure, Databricks or Google Cloud that Nxgyle never runs. Eleven model providers, local models included, chosen per step.
Your platform team builds, deploys and verifies the worker without us, and can verify from the source that nothing calls home.
Three architectures for digital workers.
Each route moves different work onto your engineering team. Here is a plain comparison, ours included.
Wire open frameworks together yourself.
Full control, but your team builds orchestration, approvals, tracing, testing and security review before the first worker ships.
Run agents on a vendor platform.
Fast to start, but the runtime, logs and billing meter usually stay on the vendor's side for as long as you pay.
Generate the code, then own the repo.
A generated, build-checked code project with tests and a signed manifest, hosted and maintained by your team. This is the option Nxgyle delivers.
Artifacts for your code review.
Two illustrated examples of what lands in your repository and review queue: a blueprint diff and an evaluation report.
- Intake supplier records posted by the procurement API
- Check tax ID, IBAN and sanctions list, as coded rules
- Approval one buyer clears every new supplier
- Approval suppliers above the spend limit add a finance gate
- Route possible sanctions hits go to the compliance queue
Illustrative only: one policy change, visible as a diff and recorded against its version.
How blueprints are versioned-
Scenarios run against mocks
Single steps or the full flow, with every tool, connector and sub-worker call answered by a response pack.
-
Red-team probes, when switched on
A scorecard of known attack attempts; it never blocks export.
-
Gates held for an approver
The test run holds at the gate until an approver decides on the actions page.
-
Two or three versions compared
Same scenario, side by side; deterministic replay is available as an option.
Generated from a mock-mode run; the traces and cost totals behind it stay attached.
See how Nx·Studio runs testsDelivery catalog, counted.
Figures reflect the delivery catalog published on this site. A digital worker is software that carries one business process from intake to outcome, following your rules. You receive full source code, not a repackaged bot or script.
Same demo at kickoff. Different stack at exit.
With owned code, the runtime, logs and secrets stay in your stack; a rented platform usually keeps them on its side.
Usually the vendor's tenancy.
Typically a per-call or per-seat meter.
Usually a vendor console over opaque logic.
Access and logs typically end with it.
Your clusters, inside your network perimeter.
No Nxgyle meter; your AI provider bills tokens.
A run log per execution, optional export to your monitoring.
Workers keep running from your repository.
Nxgyle bills design and delivery; nothing in the code meters usage.
Managed Delivery, if you want it, is scoped as a separate service.