Level 01
Supervised
Every action is reviewed by a person before it runs.
- Reads
- Asks
- Edits
- Asks
- Commands
- Asks
Nx IDE · governed AI engineering
More engineering. Same control.
Nx IDE gives your engineering organization specialist AI agents inside VS Code. They plan, write and check changes, and in Orchestrator mode separate specialists test and review them. It runs on the model provider you choose, or a fully local model, and by default no change lands without an engineer's approval.
31 specialist modes 20+ providers, or run locally No reporting to us
Nx IDE ships as a VS Code extension named Nxgyle Code and is activated with a Nxgyle license key, checked offline.
Where it fits
Nx·Studio gives each business process its own digital worker, owned by your enterprise. Nx IDE raises the output of the engineers who maintain and extend your systems. The Cognitive Engine settles routine decisions after its rules have proven themselves.
Explore the platform The capability
Engineering capacity, inside VS Code.
Specialist AI agents that read your codebase, plan the work, change code and run tests, in the editor most engineers already use. By default, no change lands without an engineer's approval.
Why it matters
Throughput without losing control.
Most AI coding tools leave a developer checking every suggestion by hand. Nx IDE divides the labor: under Orchestrator mode a single agent drafts the change, and a second and third agent test and review it. The agents are designed to pause for approval before anything ships.
The deciders
A decision for the CTO and the CFO.
The CTO weighs delivery capacity and security posture; the CFO weighs model spend, visible task by task. Both keep the freedom to change model providers without changing tools.
How work is governed
Control at every step of a task.
Five steps, each with a control your engineering leaders set: what the agents may do alone, and what waits for a person.
01
01
The request is stated
An engineer states the requirement in plain language and attaches what matters: files, error output, PDFs, or Word and Excel documents. Work starts in Code mode, and any of the 31 built-in modes is one command away.
02
02
Approval before the build
Bigger changes start with a written plan from Plan mode. Reviewers annotate it section by section, and building waits until the plan is approved: by a reviewer, or after a countdown at higher dial levels.
03
03
Autonomy is set by policy
By default, every change and command waits for approval. Written policy and the five-level Delegation Dial determine what may run unattended. The agents are designed to ask before code is committed, pushed or deployed, and deny rules block an action outright at every level.
04
04
File history at every step
Checkpoints, snapshots of your files, are saved as the agents work; any earlier state can be compared with today's or recovered in a single click. A live meter shows each task's estimated cost while it runs.
05
05
Tests run before the summary
By default, when a task changes files, Nx IDE reviews its own edits and runs your build and tests, then reports a summary of every change it made.
vendor independence
No lock-in to one AI vendor.
Choose from more than 20 model providers across the major clouds and open-model hosts, or run models on hardware you control. Changing providers never means changing tools.
Keep models in-house
Models can run on hardware you control, through any compatible model server, so source code never has to leave your network.
Subscriptions you already pay for
Engineers can sign in with an AI assistant subscription the company already holds, or use the models its VS Code environment provides.
Spend matched to the task
Each specialist keeps its own model setting, so planning, coding and review can each run on the model that balances cost and quality.
Keys never in plain files
Provider API keys sit in the operating system's secure credential store, not in settings files.
data exposure
Your code bypasses Nxgyle entirely.
Code travels directly to the model provider you selected, and no Nxgyle server sits on the path. The extension reports nothing to us, and license checks happen offline.
Nothing reported back
The product no longer contains usage reporting (telemetry), so no code, prompts or usage data ever come to Nxgyle.
No hop through Nxgyle
Each request travels only to your selected provider; local models keep it on the machine.
A TAA edition for public sector
Organizations subject to the US Trade Agreements Act can use a separate TAA edition, which excludes providers and model families of restricted origin at build time. They cannot be switched back on.
Sensitive folders fenced off
Anything named in an ignore file (.nexgileignore) stays out of reach for every agent.
change risk
Know a change's risk in advance.
Cortex, a private map of your codebase, shows which code calls which services, what a change could break and how serious that would be. It is built on the developer's machine, stays current during the work and never leaves that machine.
Consequences shown in advance
Flows, routes, subsystems and a change's likely impact are visible before the change is made, computed locally with no network calls.
Legacy code, modernized
Translate mode converts older code one file at a time and flags each difference, so modernization stays reviewable.
Decisions remembered
Project notes carry forward between sessions, and earlier decisions stay findable in past conversations.
Reads PDFs, Word and Excel
Documents and web pages are usable sources and, when enabled, so are other code repositories.
fixed boundaries
Controls enforced in software.
Beyond approvals, Nx IDE refuses unsafe file writes and can restrict where agents may edit. An optional sandbox, available for macOS and Linux, lets the operating system itself block commands from changing anything outside the project.
Checkpoints
On by default and saved as the agents work, so the files can return to any earlier state in one click.
Where agents may work
Rules can restrict where agents edit, and automated work can be fenced into chosen folders; the tools themselves enforce both.
Sandboxed commands, optional
For macOS and Linux only: commands may change files just within the project and folders you approve, and can lose network access.
Partial writes refused
Incomplete files are never saved, and writes that were cut off, or that skip code with “rest of the file unchanged”, are turned down.
specialists
Specialists, not a single assistant.
31 built-in modes, from Plan and Code to Tester, Reviewer and Shipper, each focused on one part of the job. Teams can add their own agents, skills and commands as simple files, or install them from built-in marketplaces.
A bench of 31 built-in modes
Ten core roles and 21 specialists cover day-to-day engineering, refactoring and database migrations included.
Your own agents, reviewed
A custom agent is one short Markdown file, versioned and reviewed like any other change in your repository.
Skills from a marketplace
Skills (reusable instructions) and tool connections come from built-in marketplaces, or are written once and shared.
Standards every agent reads
Instruction files in the project, such as AGENTS.md, tell every specialist how your organization builds software.
separation of duties
No change ships on one agent's word.
In Orchestrator mode, one specialist or a small team writes the change, two others check it independently, and a person is asked whether it ships.
01 Code · or a team
Implement
One agent, or a team of 2–5
02 Tester · Reviewer
Test + Review
Simultaneous, separate
03 You
Ship?
A person is asked to sign off
04 Shipper
Ship
Commits to its own branch
Unbiased checks
The Tester and Reviewer start fresh with only the requirement and the changed files, never the author's reasoning. Editing is not available to either of them.
At most two retries
If tests fail or changes are requested, a fresh implementer tries again, no more than twice, before the decision returns to a person.
One stop for the whole chain
A single Stop brings the entire chain to a halt, and an agent map tracks each agent's current task.
Parallel work
Agent Manager
One screen runs several AI sessions in parallel, with each agent's progress, cost and open questions in view.
Parallel work
Agent Teams
For larger jobs, 2–5 agents split the work, such as back end and front end, sharing a task list and messaging each other in plain sight.
Parallel work
/batch
Large changes break into up to 30 small tasks that run in parallel, each in a separate copy of the code. Any of them can become a draft pull request on GitHub if you choose.
Parallel work
/review
A panel of up to five reviewers checks your changes simultaneously and edits nothing; fixes wait until you select them.
The Shipper is instructed to commit on a branch of its own and leave others' work untouched. “Ship?” is an approval question: at the Supervised and Trust Reads levels a person always answers it, and higher up a deny rule keeps it that way.
Change governance
Change discipline, by default.
The change-management disciplines your organization already runs, built into how the agents work.
01
Gatekeeping
Nothing is edited or executed by default until someone signs off.
02
Rule sets
Policies turn into rules that permit, question or block each action, and each automatic approval names the rule that allowed it.
03
Ask-before-release
The agents are designed to pause and check with you before they commit, push, deploy, migrate or delete. Under a deny rule, the pause becomes mandatory at every level.
04
Hooks
Your scripts run at set points, for example a code-quality check that must pass before a task completes.
05
Guarded settings
No agent can touch the files that hold permissions, hooks or plugins, and the main agent must ask before it edits its own rule files.
06
Separation of duties
Whoever tests or reviews a change has no ability to edit files.
Tool access, rules, hooks and the sandbox are enforced in code, not left to instructions the AI may or may not follow.
Ahead of approval
Three questions before sign-off.
Any AI coding tool raises three executive questions: who approved a change, what it cost, and how far the agents could reach. Here is how Nx IDE answers each.
Q1
Who said yes to the change?
Either a person, or a setting or policy someone put in place. Approval is on by default, and written rules and the Delegation Dial define what may run unattended. The agents are designed to check in before code is committed, pushed or deployed.
Q2
What did the AI spend come to?
A live meter shows each task's estimated cost as it runs, and a Usage & Cost dashboard breaks spend down by day, model and project, with budget alerts. Optional per-task limits pause work until someone decides.
Q3
What was off limits?
Anything your settings and rules did not permit. Rules sort each action into allowed, needs approval, or denied, and deny rules also bind sub-agents. No agent may edit permission or hook files, and the tester and reviewer cannot edit at all.
Nearly every control is a setting or a rule file. Your security team can read them, version the rule files alongside your code and map each one to controls it already runs.
Review the security evidence the delegation dial
Set the agents' freedom by policy.
A single setting under the chat box decides how much the agents may do unattended. Raise it one level at a time, as results justify it.
Note: checkpoints remain on at all five levels. Above Supervised, sub-agents (agents started by another agent, such as the Orchestrator's specialists) run with every automatic approval switched on, and they remain subject to your deny rules.
More oversight More autonomy
Level 02
Trust reads
Reads within the project happen without approval, while edits and commands keep asking.
- Reads
- Automatic
- Edits
- Asks
- Commands
- Asks
Level 03
Trust edits
Reads and edits run unattended. After a 60-second countdown, routine questions such as plan approval default to their first suggestion. Commands still ask.
- Reads
- Automatic
- Edits
- Automatic
- Commands
- Asks
Level 04
Trust execution
Reads, edits and commands run unattended, as do tool connections marked always allowed. Starting sub-agents, editing protected files or files outside the project, and certain always-ask actions still require approval, and deny rules keep blocking.
- Reads
- Automatic
- Edits
- Automatic
- Commands
- Automatic
Level 05
Autonomous
Everything proceeds unprompted, with three exceptions. Deny rules still block, and edits to protected or outside-project files still require approval, as do a few actions such as fetching web pages.
- Reads
- Automatic
- Edits
- Automatic
- Commands
- Automatic
The roster
Clear duties, by role.
Ten core roles, each answerable for a single duty. For one task, the Orchestrator brings several together.
Plan
Scopes the work and designs it before code is written.
Code
Writes new code and reworks old code.
Chat
Explains the system and answers questions about it.
Debug
Finds and fixes defects, step by step.
Orchestrator
Runs a team of author, tester, reviewer and shipper.
Tester
Tests each change independently.
Reviewer
Reviews code it did not write.
Shipper
Commits approved work to its own branch.
Spec Bootstrap
Interviews your team, then writes a project wiki.
Spec Sync
Updates the wiki with every commit.
Plus 21 specialists for day-to-day jobs
RefactorTest GenDoc GenSecurityAPI ModDep UpgradeIssue FixerPR FixerMerge ResolverDocs ExtractorIssue InvestigatorIssue WriterDB MigratorDB ArchitectCode ArchitectDevOpsAPI TesterCode ReviewerTranslateUI DesignerDev
Need a role that isn't listed? Write it as a brief Markdown file, for one project or all of them.
Coverage
Every stage of software delivery.
From planning through release, Nx IDE has a role at each stage, and your policy decides how much it does unattended.
Design up front
Settle the approach before money is spent on code.
- A plan reviewers comment on, section by section
- Architecture drafted before the build
- A project wiki, written and kept current
Build & update
Deliver new work and retire technical debt.
- Write, change and restructure code
- Updating APIs and dependencies; translating code
- Legacy modernization through Translate mode
Testing & checks
Find defects before customers find them.
- Tests written for each change
- It reviews itself and runs tests before reporting
- An independent tester on every Orchestrator run
Review & handover
Independent review, then a controlled release step.
- /review: up to five reviewers at once
- Every edit held for approval by default
- Asks before shipping; ships to a feature branch
Whole-codebase context
Decisions made with the whole codebase in view.
- Cortex, a private map of flows and change impact
- Notes kept across sessions; past chats searchable
- Semantic search after a one-time setup
Connected tooling
Fits the systems your organization already runs.
- Links into the systems you run
- Add-ons from built-in marketplaces
- Optional: opens the app it built to confirm it runs
Inside your limits
Pace set within limits you define.
- Approval by default, the Delegation Dial for exceptions
- Rules that permit, ask or block, plus hooks
- An optional sandbox on macOS and Linux
- A cost view per day, model and project
Parallel by design
From a single session to many agents at once.
- Agent Manager: many sessions at once
- Agent Teams: 2–5 agents that coordinate
- /batch: up to 30 parallel tasks, each isolated
Familiar workflow
Fits how your engineers already work.
- Checkpoints at every step, restored in one click
- Optional autocomplete and dictation
- Plain-language requests → terminal commands
- A chat panel, in the editor or a browser tab
Latest release
What the latest release adds.
- Parallel
Many agents, working in parallel
Many agents can work at once and small teams can coordinate, while large changes break into parallel tasks, each with an optional draft pull request.
- Reviews
Reviewable plans, and /review
A plan document stakeholders comment on before anything is built, plus code review by up to five reviewers at once.
- Rules
Guardrails: rules, hooks, sandbox
Rules that permit, ask or block, custom checks at key moments (hooks) and an optional operating-system sandbox on macOS and Linux.
- Notes
Memory and a cost dashboard
Project memory across sessions, reading of web pages and documents, and a Usage & Cost dashboard with budget alerts.
- Models
Newest provider models
The model list includes current releases from leading AI providers.
Settings
Defaults versus options.
Disabled until switched on
- Sandboxing for macOS and Linux
- Request and cost caps per task
- Web lookups and research across repositories
- In-browser checks of what it built (Chrome, Edge)
- Automatic memory notes
- Autocomplete, spoken replies and dictation by voice (needs FFmpeg and a speech-to-text key)
Needs setup once
- Semantic codebase search
- Tool connections: add the ones your teams use, or install them from the marketplace
Can it act on its own?
In the editor, only if you allow it.
Approval is the default there, and your rules plus the Delegation Dial set how much it may do alone. The command-line version approves automatically by default, and at higher dial levels sub-agents also act without asking, within your deny rules.
Could AI spend run away?
Spend is visible as it happens.
Budget alerts by day and month warn but never stop work, and optional per-task limits pause an agent until someone decides.
Is there AI vendor lock-in?
No.
Choose among more than 20 providers, including fully local models, and the TAA edition removes providers of restricted origin. Code never has to leave your network.
Does Nxgyle ever see our code?
No.
There is no telemetry and no Nxgyle service in between; the model provider you choose is the only recipient of your code.