Neural InverseRequest access
Neural Inverse Build / CLI

An AI coding agent for the terminal, under rules you control.

It reads your code, runs commands and makes changes with the model of your choice, interactively or in CI.

Part of
Neural Inverse Build
Command
neuralinverse, alias ni
Platforms
macOS, Linux and Windows
Install
npm or a checksum-verified binary
Models
Gateway, provider key or local

Overview

The CLI works in your project folder. Every tool call is checked before it runs: first by your organisation's governance, then by the permission mode and rules you and your team set. It runs the same way in a terminal and in a pipeline.

What it does Workflow Governance Permission modes Technical facts Get started

What it does

What the CLI does.

A coding agent in the shell, with the controls a security team expects.

An agent in your terminalAsk in plain language, attach files with @, run your own commands with ! and keep typing while it works. The conversation stays in your scrollback.
Permission modesask, auto-edit, read-only and full-auto, refined by allow and deny rules for commands, paths, hosts and MCP tools. Deny rules always win, in every mode.
SandboxShell commands can run in an operating-system sandbox, bubblewrap on Linux and sandbox-exec on macOS, with the file system read-only outside the workspace and the network cut.
Secrets kept outProcesses the agent starts do not inherit variables that look like secrets, and secrets in tool output are redacted before it reaches the model.
Sessions and checkpointsSessions are saved and every prompt is a checkpoint you can return to. Project instructions live in AGENTS.md at the repository root and in subdirectories.
Scripts and CIneuralinverse -p runs one prompt to completion and exits, with JSON output, turn, cost and time limits, and documented exit codes. A GitHub Action is included.
Firmware and HDL environmentsOptional project environments add firmware tools for Cortex-M builds with Arm GCC, Arm Compiler 6 or IAR, and HDL tools for lint and simulation.
MCP and checksManage MCP servers, and check changes against your organisation's rules with the same findings as the Editor and CI.

Key workflow

From install to a reviewed change.

  1. Install
    With npm, or as a standalone binary verified against the release checksums.
  2. Connect a model
    Set a provider key, run neuralinverse setup, use a local model, or sign in with neuralinverse login.
  3. Start in your project
    Run neuralinverse in the project folder and describe the task.
  4. Approve or deny
    A call that needs you shows what it will do, with the diff for an edit.
  5. Review and commit
    Inspect the diff, undo to any checkpoint, then commit or open a pull request.
  6. Automate
    Run the same work non-interactively in scripts and CI.

Who uses it

Built for every role on the team.

Terminal usersAsk in plain language, approve each call that needs it, and undo to any checkpoint.
Build engineersRun governed AI steps in CI, with cost and time limits and a work item for each run.
Platform teamsShip a managed policy that fixes providers and models and requires sign-in.
Security teamsDeny secret files and risky commands, require the sandbox and forward the audit log to the SIEM.

How it is governed

Organisation governance comes first.

Before the CLI's own rules and mode, every tool call is decided by the same decision point as the Editor and the Agents app.

Deny wins, approval is a personA governance deny is refused. A call that needs approval asks you in every mode; no rule, hook or 'always allow' can pre-approve it.
Managed policyA managed settings file locks every key it sets. Users can add deny rules but never remove the organisation's, and full-auto can be switched off.
Remote signed policyThe managed file can point to a policy served over HTTPS and signed with Ed25519. Unsigned or wrongly signed policies are rejected, and the CLI can refuse to start without one.
Work itemsIf policy requires a work item, an interactive session asks for it; headless runs are refused without one.
No silent approvals in CIIn a non-interactive run, anything that would ask is denied. A governance approval counts only if a signed approval of exactly that action is in the repository.
Hardware needs a personFlashing always needs a person's approval, in every mode, and non-interactive runs cannot flash. Nothing is flashed if the image changed after approval.
Audit logAn append-only JSON Lines log of decisions, tool calls with a SHA-256 of their input, mode changes and MCP connections, ready for your log forwarder. File contents are not recorded.

Permission modes

Four modes. Rules refine each one.

Switch with --mode, /mode or Alt+M. Deny rules apply in every mode.

ask (default)Reads in the workspace are allowed. File edits, commands and network ask, except recognisably read-only commands.
auto-editEdits in the workspace are allowed; commands and network ask. Git internals, shell start-up files and the CLI's own settings always ask.
read-onlyEdits, commands and network are denied. The agent plans instead.
full-autoEverything is allowed unless a deny rule or your organisation's governance says otherwise. Administrators can switch this mode off.

Technical facts

For your IT and security review.

Installnpm install -g @neuralinverse/cli with Node.js 22.19 or newer, or a standalone binary that needs no Node.js. The installer refuses any checksum mismatch and does not edit your shell profile.
PlatformsmacOS on x64 and arm64; Linux on x64 and arm64, and musl-based distributions on x64; Windows on x64.
Sign-inOptional unless your organisation requires it. Browser sign-in with PKCE, or a device code, used by itself over SSH and on Linux with no display.
ModelsAnthropic, OpenAI, Google Gemini, AWS Bedrock, Claude on Vertex AI, Azure OpenAI and any OpenAI-compatible server. Ollama is detected automatically. Keys are kept in the operating system's keychain.
Without internet accessHost the release assets on an internal mirror, route model traffic through an internal gateway or a local model, and turn off the daily version check. Proxies and custom CA certificates are honoured.
TelemetryOff by default, with no built-in destination. It can be sent only to an OpenTelemetry collector you configure, and tool inputs and outputs are never exported.
Licence and statusProprietary, under your agreement with Neural Inverse; source review under NDA. Pre-release: interfaces and data formats may change between versions.

Get started

Install, sign in and read the documentation.

Talk to our team.

Neural Inverse is available by request. We work with your engineering, security and procurement teams from the first conversation.

Request accessContact us
Strictly necessaryAlways on

Storage the website needs to work, such as remembering these settings. It contains no identifier and is never used to track you.

Google Analytics counts visits and shows which pages are useful, so we can improve the website. It sets cookies and sends your IP address, browser details and the pages you visit to Google in the United States. Google Signals and advertising features are switched off. Off unless you turn it on.

Read the cookie policy

CLI · Neural Inverse Build · Neural Inverse