Nexpath CLI
Before a prompt reaches Claude Code or Cursor it adds checks, tests and rollback steps, and you edit the result
- 358 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- Hooks into Claude Code, Cursor and Windsurf settings and captures every prompt before sending
- Stores prompts in ~/.nexpath and sends parts of them to OpenAI or the Nexpath service for classification
- Keeps the OpenAI key or Nexpath token in the system keychain or in config.json
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/nexpathDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Hooks into Claude Code, Cursor and Windsurf settings and captures every prompt before sending
- Stores prompts in ~/.nexpath and sends parts of them to OpenAI or the Nexpath service for classification
- Keeps the OpenAI key or Nexpath token in the system keychain or in config.json
This is third-party code. Review the repository files before installing.
What it does
Nexpath sits between you and an AI agent and improves the prompt before it is sent. It installs as nexpath-cli from npm, and nexpath install connects it to the agents it finds: Claude Code, Cursor, Windsurf and, through Chrome and Firefox extensions, Replit, Lovable and Bolt.new. A VS Code extension covers Cursor and Windsurf. For each prompt Nexpath looks at the work stage and at signs of what is missing: no regression test, no acceptance criterion, no bug reproduction. From that it builds one editable prompt with sections for scope, constraints and verification. Risky actions get confirmation, backup or rollback steps, and a complex task is split into an ordered sequence of prompts. Your original text always stays inside, and you can return to it before sending. Data is stored locally in ~/.nexpath, and API keys, tokens such as sk- and ghp_ and PEM blocks are stripped before storage. For classification Nexpath calls a model. If a call fails it quietly falls back to local detection, so run nexpath status when behaviour looks odd.
Who it is for. For people who build fast with agents, such as solo developers and founders, and do not want tests and checks to slip.
Good fit when
- You ask an agent to fix a bug and the request has no reproduction steps and no check that the bug stays fixed
- The agent is about to delete files or change a database and you want a confirmation and a rollback plan first
- A large task goes out as one prompt although it is better split into a sequence of steps
- You work in Replit, Lovable or Bolt.new in the browser and want to see the improved prompt before sending
- You work in Cursor or Windsurf and want to see the improved request right in the editor through the VS Code extension
Not a fit when
- You need protection from dangerous agent commands at execution time: Nexpath works on the prompt before sending and blocks nothing
- Code and prompts must not go to an external API: classification needs OpenAI or the Nexpath service
- You use an agent outside the list: support is declared for Claude Code, Cursor, Windsurf, Replit, Lovable and Bolt.new
- You want ready-made skills or MCP servers: Nexpath improves requests and adds no tools to the agent
Example request
Fix the payment handling bug that sometimes charges twiceLimitations
- Requires Node 18 or newer.
- Part of the prompt goes to a model for classification. Telemetry is off until you turn it on with `nexpath config set telemetry.enabled true`.
- The author does not promise bug-free code. In the author's 40-task test 29 tasks were solved against 27, and a difference of two tasks is too small to treat as reliable.
- Nexpath does not work without local capture, so nexpath store disable effectively switches it off.
- Gemini as a second provider is only planned.
How to disable. nexpath uninstall disconnects Nexpath from detected agents and offers to clear the stored key. Then run npm uninstall -g nexpath-cli. Local data lives in ~/.nexpath, and nexpath store delete -y removes saved prompts. Remove the browser and VS Code extensions in their own managers.
FAQ
How do I install Nexpath for Claude Code?
Run npm install -g nexpath-cli, then nexpath install. It finds Claude Code, Cursor and Windsurf, asks which credential to use and asks for your role so its guidance fits the work you do. Check the install with nexpath --version.
Do I need an OpenAI key?
You need one of two options: your own OpenAI key or a Nexpath token from a free account at parseos.tech/nexpath. If both are set, the OpenAI key is used. The key lives in the macOS keychain, Secret Service on Linux or Credential Manager on Windows, with ~/.nexpath/config.json at mode 0600 as a fallback.
Why does Nexpath show nothing?
Run nexpath status. It reports which credential is in effect and the last model call that failed: a key out of credit, revoked or replaced. On a failure Nexpath quietly falls back to local detection, which looks like nothing happening.
What does Nexpath send out?
Targeted requests to a model that classify the prompt and prepare guidance. Keys such as sk- and ghp_, bearer tokens and PEM blocks are stripped before anything is written to disk.
Is it proven that Nexpath improves results?
The author gives their own test on 40 SWE-bench Verified tasks: Claude Code solved 27, with Nexpath 29, and the gain on two tasks came from Nexpath. The sample is small, so treat it as a signal rather than proof. Per-task data sits in the Nexpath_benchmark folder.
Discussion
No questions or reviews yet.
Related
CLI toolsA teaching terminal coding agent in Python, built as a "nanoGPT for agents": read it, take it apart and fork it
A plugin that gives an agent the IntelliNode skill and MCP server: ask any model, search, review code, tests, SQL, images and speech
A memory and coordination layer for coding agents: specs and issues live in git while agents work in parallel
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent
A kit of 138 agents, 296 skills and 74 hooks that turns Claude Code into a full software development team
A skill for coding agents that pushes them to finish the job: keep trying after failures, switch approaches and prove the result
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| Nexpath (this entry) | no | 358 | |
| Superpowers | no | 307,845 | development process skills inside the agent, with no prompt review before sending |
| HOL Guard | yes | 1,012 | checks agent actions at execution time and blocks dangerous commands |
README in short
The project explains why a prompt quality layer is needed and walks through how it works in four steps. Separate sections cover privacy and troubleshooting.