NeuroAPI: one-click installers that point Codex CLI and Claude Code at a Russian AI API
NeuroAPI: российский AI API для Codex CLI и Claude Code — безопасные one-click установщики для Windows и macOS.
At a glance
- What is it?
- NeuroAPI is a set of PowerShell and shell installers that wire the Codex CLI and Claude Code clients to neuroapi.host without ever writing your API key into a config file. This article covers what the installers create, how the key is stored on each platform, and where the approach breaks down.
- Who is it for?
- Adopt NeuroAPI if you already have Codex CLI or Claude Code installed, want them pointed at neuroapi.host, and are unwilling to leave a plaintext key in a TOML or JSON file. Do not adopt it if you need a headless or CI install path, or if you are not on Windows or macOS, because the repository ships no such route.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 47 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem is not the API, it is the credential handoff
Codex CLI and Claude Code both expect to talk to a model provider, and both let you override where that provider lives. Doing so by hand means editing a TOML profile for one client and a settings JSON for the other, then deciding where the API key goes. The obvious answer, pasting it into the config, is the one that ends up in a screenshot, a dotfile repository, or a support thread. NeuroAPI is the installer layer that avoids that answer. It is aimed at developers on Windows and macOS who already use Codex CLI or Claude Code and want them routed to neuroapi.host, a service the README describes as a Russian AI API offering unified access to OpenAI, Anthropic Claude, Google Gemini and DeepSeek models with payment in rubles, operated by a Russian LLC with infrastructure in Russia. The repository itself is MIT-licensed TypeScript, and the installers are the product.
The installer never holds the key, the helper does
The design separates three things that are usually collapsed into one config file: the client profile, the credential store, and the process that reads the credential. The setup script prompts for the key once, stores it in an OS-level secret store, and then writes a small helper whose only job is to print the token to stdout at the moment a client asks for it. Codex gets a dedicated profile at ~/.codex/neuroapi-host.config.toml on both platforms, with authentication handled by a command-backed auth helper. Claude Code gets an installer-owned JSON file passed via --settings, with apiKeyHelper resolving the key. On Windows the secret lives in DPAPI, scoped to the current user and machine, with no separately saved master key. On macOS it lives in the login Keychain as the item host.neuroapi.agents.api-key, written with the interactive -w flag. The installers do not call the API and do not transmit the key; network traffic begins only when Codex CLI or Claude Code sends your requests to https://neuroapi.host. The README states plainly that this protects against accidental publication of the key, not against malware already running as the same user, and points to docs/security.md for the full threat model. That is an honest boundary and worth reading before you decide the scheme is sufficient.
What lands on disk, and why existing configs survive
On Windows the installer creates %LOCALAPPDATA%\NeuroAPIAgents\ containing the helper, the Claude settings file, the DPAPI ciphertext and the launchers, plus %USERPROFILE%\.codex\neuroapi-host.config.toml and a single entry in the user PATH pointing at %LOCALAPPDATA%\NeuroAPIAgents\bin. On macOS it creates ~/.local/share/neuroapi-agents/ for the helper and Claude settings, the same ~/.codex/neuroapi-host.config.toml, and two launchers at ~/.local/bin/codex-neuroapi and ~/.local/bin/claude-neuroapi. The README states that existing configurations are not overwritten on either platform, and that every file the uninstaller removes carries a narrow ownership marker. If a target path is already occupied by a file the installer does not own, installation fails rather than overwriting. That failure mode is deliberate and it is the right call, but it also means a half-finished earlier attempt can block a clean run until you clear the path yourself. Neither installer requests administrator rights, and the macOS script explicitly does not use sudo and does not edit your shell profile.
Running it: four commands and two verification steps
Install Codex CLI and/or Claude Code first, then create an API key in the NeuroAPI dashboard. On Windows, download the ZIP from the agents branch, unpack it, double-click setup-windows.bat, and paste the key into the hidden PowerShell prompt. The BAT file does not accept the key as a command-line argument, which is the point. Open a new terminal and run codex-neuroapi or claude-neuroapi. On macOS, run chmod +x setup-macos.command followed by ./setup-macos.command from the unpacked directory, paste the key into the Keychain prompt, then run ~/.local/bin/codex-neuroapi or ~/.local/bin/claude-neuroapi. If ~/.local/bin is already on your PATH, the bare command names work. Verification differs per client. In Codex, run /debug-config and check for the neuroapi-host profile, the neuroapi provider and https://neuroapi.host/v1. In Claude Code, run /status and check the base URL https://neuroapi.host and the credential source apiKeyHelper. The README's examples use gpt-5.6-sol and claude-sonnet-4-5, and it warns that model IDs change: on a model not found error, pull the exact name from the NeuroAPI catalogue or GET /v1/models. To rotate a key, rerun the setup file. To remove everything, run uninstall-windows.bat or ./uninstall-macos.command; the README notes that a deleted key cannot be recovered through this package.
The gaps: no Linux, no headless path, and a thin release history
The installers cover Windows and macOS only. There is no Linux script, no Docker image, and no documented non-interactive mode, so provisioning a build agent or a remote dev box means reproducing the helper, the profile and the Keychain or DPAPI equivalent by hand from docs/manual-setup.md. The interactive prompt is a deliberate security property and also the thing that rules out automation. A second constraint is the release history: the listed releases are v2.1.1, v2.1.0 and v2.0.0, all dated December 2023 and all labelled NeuroGPT, while the most recent push to the repository is dated 2026-07-30. The tags do not describe the installer work visible in the README, so anyone pinning to a release tag should check what that tag actually contains rather than assuming it matches the branch. Third, the endpoint is a hosted commercial service, not something you run. If neuroapi.host is unreachable or a model is retired, the launchers still start but the requests fail, and the fix is a model ID or a base URL change rather than anything in this repository.
Against a plain provider block in your own config
The alternative most people reach for is editing the client configuration directly: add a provider block to the Codex TOML and an env or settings entry for Claude Code, with the key inline or exported from the shell. That approach is simpler, works on Linux, and needs no installer to maintain. The difference is where the secret rests. An inline key sits in a file that is easy to copy, sync or commit, and an exported environment variable is readable by any process that can read your environment. NeuroAPI's approach keeps the key in DPAPI or the login Keychain and exposes it only through a helper that prints it on demand, which narrows the accidental-exposure surface at the cost of an extra indirection layer and two more files to reason about. If your machine already has a secrets manager and your client can read from it, you get the same property without this package. NeuroAPI's value is that it does that wiring for two specific clients on two specific platforms, with CI checks for PowerShell and Bash syntax, a DPAPI smoke test with a test token, a mock-Keychain macOS test, ShellCheck, JSON and TOML parse checks, and a scan for accidentally committed secrets or unsafe key-passing patterns.
Maintenance cost and the MIT licence in practice
The update surface is small. The installers write a profile, a settings file and a helper, and the README says existing configs are untouched, so upgrading means rerunning the setup file rather than merging changes. The recurring cost is elsewhere: model IDs on the hosted service change, and the README tells you to resolve them from the catalogue or GET /v1/models when a request fails. Key rotation is a rerun of the same script. Uninstall is a single script per platform and it removes only installer-owned files plus the locally stored key. The repository is MIT-licensed, which permits use, modification and redistribution with the licence text retained; the API service itself is a separate commercial relationship governed by neuroapi.host's own terms, and nothing in the MIT grant covers that. This is not legal advice, and if you are deploying inside an organisation, the terms of the hosted service are the document to read, not the repository licence.
Editorial conclusion
Adopt NeuroAPI if you already have Codex CLI or Claude Code installed, want them pointed at neuroapi.host, and are unwilling to leave a plaintext key in a TOML or JSON file. Do not adopt it if you need a headless or CI install path, or if you are not on Windows or macOS, because the repository ships no such route. Before running setup-windows.bat or setup-macos.command, open docs/security.md and confirm that the command-backed helper and the apiKeyHelper path match your own threat model, since the README is explicit that the design does not protect against a malicious process already running as your user.
Community notes