Level 01
Supervised
Each action waits for your review first.
- Reads
- Asks
- Edits
- Asks
- Commands
- Asks
Engineering help · for your business backlog
Work your backlog with Nx IDE.
Change requests, integrations and upgrades wait in your engineers' queue. Inside VS Code, their code editor, Nx IDE helps plan, write and check that work. When the Orchestrator (a team-lead mode) runs it, other specialists test and review the change. It runs on the AI model service you select; by default, every edit in the editor waits for an engineer's approval.
31 modes in total Over 20 AI providers No usage reporting
Within VS Code it is listed as the Nxgyle Code extension, which needs a Nxgyle license key to run.
How it fits
Nx·Studio produces digital workers for the processes your teams run. Nx IDE supports the engineers who extend the systems around those workers. The Cognitive Engine settles routine decisions once its rules have passed an exam on previously unseen cases.
Compare the three products In short
AI help for the engineers you rely on.
Your engineers describe a change request in ordinary language. It studies the code, proposes a plan, makes the edits and runs the tests. In the editor, each change waits for an engineer's approval by default.
What you gain
Added engineering capacity.
Dedicated modes for planning, coding, testing and review take on routine steps in change requests, integrations and upgrades. The Orchestrator keeps testing and review apart from the coder.
Best suited to
Teams waiting on IT delivery.
Functions such as claims, finance operations or customer service whose requests wait in an engineering queue, especially where audits, security reviews or aging systems add work. Approvals and inspectable controls come built in.
task walkthrough
One change request, start to finish.
Here is what happens to a change request in five steps. How much runs without an engineer's approval is a setting your team controls.
01
01
Describe the request
An engineer writes the change in ordinary language. Work opens in Code mode, and typing / switches among the 31 built-in modes, such as Plan, Debug or Chat. Typing @ attaches files, error messages, terminal output, PDFs, documents or spreadsheets.
02
02
A plan comes first
On bigger changes, Plan mode drafts a plan document and revises only the parts reviewers comment on. Building begins after an engineer approves the plan or, at higher autonomy levels, a countdown runs out.
03
03
Engineers approve each step
Out of the box, each edit and command in the editor requires an engineer's approval. With written rules and the Delegation Dial, a setting with five autonomy levels, teams can let chosen actions go ahead. Agents are told to stop before code is committed, published or deployed. A deny rule, which blocks an action outright, turns that stop into a hard block whatever the dial says.
04
04
Checkpoints you can roll back
Checkpoints (snapshots of the files) are saved as work proceeds, and any earlier one can be compared or restored with one click. A meter tracks the task's projected cost and the fill level of the model's working memory (context).
05
05
Self-checked, then summarized
By default, after a task changes files, the agent checks its own work and runs your automated build and tests. Only then does it report what changed.
model services
Model choice stays with you.
More than 20 cloud and local AI providers are supported, so you can stay with services your company already buys. Your IT team checks each one's terms and data handling.
Models on your own hardware
Where data must stay in-house, models can run on your own machines through a compatible local model server, subject to your IT team's data rules.
Draw on current subscriptions
Supported subscription integrations and models offered through VS Code can be used, subject to each provider's access rules and terms.
Match models to modes
Every specialist mode remembers which model it uses, letting each step, from planning to review, run on a model suited to the job and its cost.
Keys in the vault
Access keys for AI services are held by the computer's built-in credential vault, never written into settings files.
data handling
Direct model-provider connections.
Requests travel from the editor straight to the model service you select; no Nxgyle server sits in the path. No usage data is reported back, and even the license key is verified offline.
No usage reports
The extension contains no telemetry (usage reporting). Code, prompts and usage figures stay with you; none reaches Nxgyle.
Point-to-point calls
Traffic runs directly between the editor and the provider you picked; with a local model, it never leaves the machine.
An edition without restricted providers
Where the US Trade Agreements Act applies, a separate TAA build omits restricted providers and model families at build time, and they cannot be re-enabled.
Files the agents skip
Add files and folders to an ignore file, and no agent will read them.
code insight
See what a change could affect first.
Before an engineer approves a change, Cortex shows which connected services it could affect and grades the risk from low to critical. The analysis runs on the engineer's machine, with no network access.
Knock-on effects shown
Lists the flows, entry points and parts of the system a change would touch, so they can be reviewed before release. It is computed offline on the engineer's machine.
Migration support
For modernization projects, Translate mode rewrites each source file as a same-named file in the target language and highlights every difference for engineers to verify.
Session-to-session memory
Project notes carry over between sessions, and earlier chats can be searched for past decisions.
Documents as input
Specifications in PDFs, documents or spreadsheets, and web pages, can inform a task. If enabled, it can also study other code repositories.
built-in limits
Limits held by the tool itself.
Approvals are one layer. Nx IDE also refuses incomplete file writes and can fence off where agents edit. An optional sandbox, an isolated area on macOS and Linux, limits what their commands can change to your project.
Checkpoints
On by default, checkpoints capture your files as the agent proceeds; a single click brings back any earlier state. Effects of terminal commands are not reversed.
Folder-level limits
Written rules mark out the folders agents can change. Automated tasks can be confined to specific folders, and the tools enforce it.
Optional command sandbox
Off by default; supported on macOS and Linux. Commands may change files only in your project and folders you allow, and can be denied network access.
Incomplete edits refused
If a write is interrupted, or skips code with “rest of the file unchanged”, it is rejected, so no half-finished file is saved.
specialists
Specialists for defined jobs.
Nx IDE has 31 built-in modes: ten core roles such as Plan, Code and Reviewer, plus 21 specialists, each limited to one kind of work. Your team can add its own agents, skills and commands in simple files, and the built-in marketplaces offer further skills and tool integrations.
Ready-made experts
Alongside the ten core roles, 21 prepared specialists handle routine engineering work, from refactoring to database migrations.
Team-built agents
A custom agent is a short text file, so your team versions and reviews it like any other code change.
Reusable skills
Install a skill (a reusable set of instructions) or a tool integration from the marketplace, or write one once and use it everywhere.
House rules on file
An instruction file kept with the code tells every specialist how your team works.
the Orchestrator
Builder, checker and approver kept apart.
With the Orchestrator, one specialist or a small team writes the change and two others test and review it independently. At the default setting, a person makes the release call; at higher autonomy levels, a deny rule can keep that decision with your team.
01 One coder or a team
Implement
A single specialist, or 2–5 working together
02 Tester · Reviewer
Test + Review
In parallel, each on its own
03 You
Ship?
Put to you as a question
04 Shipper
Ship
Saves work on a side branch and pushes it
Fresh-context review
Tester and Reviewer see only the requirement and the edited files. They never see the writer's reasoning, and they have no editing tools.
Retry limit
When tests fail or changes are requested, a fresh implementer tries again, at most twice. After that, the decision returns to you.
Single stop control
A single Stop button ends the whole chain at the same moment; the agent map tracks who is working on what.
Several agents
Agent Manager
Run several AI chats from one screen and see, per agent, how work is progressing, what it costs and which questions are waiting.
Several agents
Agent Teams
For bigger jobs, 2–5 agents share the work, for instance one on the back end and one on the front end. They message each other over a shared task list you can watch.
Several agents
/batch
A broad change can become up to 30 smaller tasks, several running at once, each in its own separate workspace. If you choose, every task produces a draft pull request for your team to review.
Several agents
/review
Up to five reviewers read your changes in parallel without editing anything. Fixes begin only for the findings you select.
By instruction, Shipper keeps work on its own branch and leaves other people's changes alone. “Ship?” is an approval question. At Supervised and Trust Reads a person must answer it; at higher levels, add a deny rule if every release must wait.
built-in controls
Controls your auditors can trace.
Approvals, written rules, automated checks and split duties come built in, so your IT and audit colleagues can match each to existing policy.
01
Review gate
In the editor, a person approves every action unless settings say otherwise.
02
Rule files
Rules allow, query or deny each kind of action. Whenever a rule approves something automatically, the approval shows which rule it was.
03
Release check
Agents are told to pause before commits, pushes, deployments, migrations or deletions; a deny rule turns that pause into a block at every dial level.
04
Hooks
Your scripts run at fixed points in a task; one can, for example, require a clean code-quality check before work is marked done.
05
Locked settings
Agents cannot alter permission, hook or plugin files; the lead agent needs consent to edit its own rule files.
06
Split duties
Testing and review agents have no way to edit files.
Tool access, rules, hooks and sandbox limits are enforced by code, not left to the AI's discretion.
security review
What security will ask you.
Expect your security chief to ask about approval, spend and access before engineers rely on it. Below is how Nx IDE responds to each.
Q1
Which approval covered it?
An engineer gave approval, or the action ran under a dial setting or rule that someone chose. By default nothing runs unapproved in the editor, and every automatic approval shows its rule.
Q2
How much did it spend?
A live meter estimates each task's cost as it runs, while the Usage & Cost dashboard shows spend by day, model and project, with budget alerts. If you set them, per-task limits halt the agent until a person decides.
Q3
Which files were in scope?
Only what settings and rules permitted. Your allow, ask and deny rules define the scope, and the deny rules apply to sub-agents too. Permission and hook files stay locked, and testing and review agents cannot edit.
Most guardrails live in settings and rule files. Your security colleagues can inspect them, store the rule files with your code under version control and link each to a control they already run.
Read our Trust Center autonomy setting
Raise autonomy as evidence builds.
Your engineering lead sets how far agents act alone, using one control below the chat box. Move up a level only after results at the current one hold up.
Note: checkpoints remain on whatever the level. Once the dial is above Supervised, sub-agents (agents that another agent starts, such as the Orchestrator's specialists) receive all auto-approvals, but deny rules still apply to them.
Closer review More independence
Level 02
Trust reads
Reading project files needs no approval; edits and commands still ask.
- Reads
- No prompt
- Edits
- Asks
- Commands
- Asks
Level 03
Trust edits
Reading and editing need no approval. Routine agent questions, plan approval among them, take the first suggested answer once a 60-second countdown ends. Commands still ask.
- Reads
- No prompt
- Edits
- No prompt
- Commands
- Asks
Level 04
Trust execution
Reading, editing, commands and tool connections marked always-allowed need no approval. Launching sub-agents, edits to protected or outside-project files and a few always-ask actions still ask; deny rules still block.
- Reads
- No prompt
- Edits
- No prompt
- Commands
- No prompt
Level 05
Autonomous
Nothing waits for approval except in three situations. Deny rules continue to block, edits to protected or outside-project files still ask, and a few actions, for example fetching web pages, also ask.
- Reads
- No prompt
- Edits
- No prompt
- Commands
- No prompt
built-in roles
Core roles at a glance.
Every core role has a single focus, and several can work together on one task under the Orchestrator.
Plan
Lays out the approach ahead of any code.
Code
Writes, edits and restructures code.
Chat
Explains code and answers questions.
Debug
Traces faults and fixes them methodically.
Orchestrator
Leads a team of writer, tester, reviewer and shipper.
Tester
Checks a change on its own, without editing.
Reviewer
Reviews code from a fresh context.
Shipper
Pushes approved work on a side branch.
Spec Bootstrap
Writes a project wiki after interviewing your team.
Spec Sync
Keeps the wiki current as commits arrive.
21 more specialists for everyday engineering jobs
RefactorTest GenDoc GenSecurityAPI ModDep UpgradeIssue FixerPR FixerMerge ResolverDocs ExtractorIssue InvestigatorIssue WriterDB MigratorDB ArchitectCode ArchitectDevOpsAPI TesterCode ReviewerTranslateUI DesignerDev
Need another? A short text file defines it, and the new specialist appears in the list for one project or for all.
scope of use
Where it fits in delivery.
Use this list with your engineering lead to choose which stages a pilot covers. Autonomy at each stage remains a setting.
Scoping work
Agree the approach before any coding.
- A plan document with comments on any section
- An architecture draft at the start
- Project wiki, maintained
Building and updating
Write new code and bring older code forward.
- New code, edits and refactoring
- API modernization and dependency upgrades
- Translate mode for code migration
Quality checks
Faults caught before shipping.
- New tests for every change
- A self-review and test pass at task end
- A separate tester in each Orchestrator run
Release control
Review first, then a managed release.
- /review with as many as five parallel reviewers
- Engineer approval on edits, by default
- A release question first, then a feature branch
Codebase context
Awareness of the whole repository.
- Cortex: change impact, flows and routes, computed locally
- Notes across sessions and searchable chat history
- Meaning-based search after one setup step
System connections
Links to the systems you already use.
- Connections to further systems
- Marketplace installs for tools and skills
- Optional: runs the finished app on the machine to confirm it works
Control settings
Limits you set and can verify.
- Editor approval by default, then the Delegation Dial
- Allow / ask / deny rule files and hooks
- Opt-in sandbox on macOS and Linux
- Usage & Cost dashboard with spend alerts
Scaling up
From a single chat to parallel agents.
- Agent Manager for several chats at once
- Agent Teams of 2–5 sharing one task list
- /batch: up to 30 tasks in isolated copies of the code
Everyday fit
Adapts to existing team habits.
- Checkpoints with one-click restore
- Dictation and autocomplete on request
- Everyday words → terminal command
- VS Code chat, or a browser tab
latest changes
This release, in brief.
- Parallel
Several agents at once
Agent Manager, Agent Teams and /batch run work in parallel: small teams coordinate, and big changes are split into tasks, each able to produce a draft pull request.
- Review
Plan review and code review
Comment on a plan document before anything is built, then have up to five reviewers read the code at once.
- Safeguards
Stronger guardrails
Allow, ask and deny rules, your own checks at set moments (hooks), and an opt-in sandbox on macOS or Linux.
- Insight
Better context, clearer costs
Memory that persists across sessions, research in web pages and documents, and budget alerts on the Usage & Cost dashboard.
- AI models
Wider model selection
The model list now includes recent releases from major providers.
defaults
Decide what to enable.
Switched on by your team
- Sandbox for commands (macOS and Linux)
- Per-task limits on requests and cost
- Researching the web and other repositories
- Opening the finished app in Chrome or Edge to test it
- Project memory that takes its own notes
- Autocomplete, read-aloud and voice input (voice input requires FFmpeg and a transcription key)
One setup step first
- Codebase search by meaning
- Tool connections, added by your team or installed from the marketplace
Straight answers
Settle these before a pilot.
Plain answers, caveats included.
Ask us something elseCould it act unapproved?
Not in the editor's default state, where every action needs approval.
The command-line tool approves most actions automatically unless approval is required at launch. Beyond that, the Delegation Dial and your rules set the limits; above Supervised, sub-agents act without further approval, yet deny rules still restrict them. Check these settings before connecting enterprise repositories.
Could costs run away?
Estimated cost is visible as work runs.
Daily and monthly budget alerts give warning but never stop work; per-task caps, if enabled, pause work until a person chooses what happens next.
Which AI models are allowed?
Over 20 supported providers, or models that run entirely on your hardware.
The TAA edition excludes providers of restricted origin. Confirm each choice against your data-boundary rules.
Is our code shared with Nxgyle?
No.
The extension sends no telemetry and has no Nxgyle server in the path; code goes only to the model service you select.