NightVision Skills Are Now Available in Codex

A skill is a written procedure an agent loads when the work in front of it matches. Not a description of a product, and not a prompt: the actual sequence of commands, what a successful run looks like, and what to do when one fails. The four NightVision skills have been available to Claude Code for some time. They now install into Codex as well, from the plugin directory, with no configuration file to hand-edit.

The skills themselves did not need to be rewritten for a second agent so much as unlearned from the first. Their source is now agent-neutral, which means the procedure a Codex user gets is the same procedure a Claude Code user gets, from the same file, rather than a fork that drifts.

Why the delivery mechanism matters

Applications and APIs now ship from every corner of the enterprise, and a growing share of them are written with an agent in the loop. That has changed who is sitting in front of the code when a security decision gets made. Increasingly it is an agent, working on behalf of a developer who is not an application-security specialist.

An agent in that seat does more than read. Given the findings, it prioritizes them and writes the fix, with no one hand-reading scanner output or hand-patching the code in between. What it cannot do is get there from a severity label alone. Severity ranks the work, which is why NightVision reports it and why triage starts there, but it tells you how much to care rather than what to change. The fix gets built from the rest of the finding: the request that triggered it, the response that came back, and, where the evidence supports it, the line of source that produced it.

That is the shape of the loop today. The agent does the work and arrives with a pull request. A person still reads it and decides whether it merges.

That is the layer NightVision provides, and coverage improves when a control lives inside a workflow instead of beside it. The same argument applied when these skills arrived for Claude Code, in NightVision Skills for Claude Code Are Now Available. It applies to any agent a developer actually works in, which is the reason to stop shipping this per-agent.

None of it shrinks the security team. It is the only thing that scales one. The agents absorb the translation work, application security stops being the throughput limit, and the judgment stays where it belongs: setting the standard, judging the fix, approving the merge.

The four skills

API discovery drives nightvision swagger extract, which reads your repository and generates an OpenAPI specification through static analysis in under 20 seconds. No running application and no compilation step, because it is reading source rather than watching traffic. Each recognized route retains the file and line that defines it, which is what makes it possible to trace a finding back to code later. The skill also covers framework detection, unresolved variables, and running swagger diff against the specification you already publish. That last one is usually the uncomfortable part, because the gap between the two is your shadow API surface.

Scan configuration covers everything that has to exist before a scan runs: projects, targets, scope exclusions, and authentication. Authentication is where most scan setups quietly fail, and a scanner that is not logged in tests the login page and reports that your application is clean. The skill handles Playwright-scripted logins including MFA, header and cookie credentials, and recorded traffic for applications with workflows a crawler will not reach on its own.

Scan triage takes the findings the other way. It reads SARIF or CSV output, explains each finding in terms of the affected endpoint and what an attacker would do with it, uses the source annotations to open the responsible code, and builds a curl command so a human can reproduce the finding by hand before anyone writes a patch. Reproducing it first is the point: it is the difference between fixing a vulnerability and editing code until an alert disappears.

CI/CD integration moves the whole thing into the pipeline across GitHub Actions, GitLab CI, Azure DevOps, Jenkins, Bitbucket, and JFrog. Token handling, SARIF export into your code-scanning view, and detecting breaking API changes between builds. This is where a scan stops being something someone remembers to run.

How the skills actually work

Each skill is a markdown file with front matter. The front matter declares a name, a description that tells the agent when the skill applies, and the tools the skill is allowed to use. The body is the procedure. These four declare shell access, plus file reading for triage, because they work by running the NightVision CLI and reading what it writes.

That decision is the whole design. The skills do not reimplement NightVision and they do not call a hosted API on your behalf. They run nightvision on your machine, under your credentials, in the repository you already have checked out. Three things follow from that:

• Private targets work. The scan is initiated from your environment, so localhost, an internal staging host, and a service that never faces the public internet are all reachable.
• The repository is the context. API discovery is static analysis over source you already have, which is why it needs no running application and no traffic sample.
• Nothing new holds your credentials. The skills use the CLI session you established with nightvision login. No token is copied into a config file for an agent to read.

The tradeoff is that the CLI has to be installed and authenticated first. Each skill opens by checking that nightvision --help responds and tells the agent how to recover when it does not, because the alternative is an agent confidently narrating a scan that never ran.

The MCP server, and when you want it

The skills teach an agent to operate the CLI. NightVision also publishes an MCP server, which is a different shape: it exposes the platform as tools the agent calls directly, so the agent can start a scan, poll it to completion, and read structured findings back without parsing terminal output. That suits a workflow where the agent runs the scan itself rather than handing you commands.

The two are complementary, and the server installs from the repository into either agent:

codex plugin marketplace add https://github.com/nvsecurity/nightvision-mcp
codex plugin install nightvision@nightvision

Getting started

Install NightVision from the Codex plugin directory, then authenticate the CLI the skills drive:

nightvision login

From there, ask for what you want in plain language. "Discover APIs in this repository with NightVision." "Configure a NightVision scan for this application." "Triage these NightVision scan findings." The description on each skill is what routes the request, so you do not need to name the skill or remember a command.

Frequently asked questions

What are NightVision skills for Codex?

They are four written procedures a coding agent loads when it is asked to do application security work: API discovery, scan configuration, CI/CD integration, and finding triage. Each one tells the agent which NightVision CLI commands to run in what order, what a successful run looks like, and how to recover from a failure. They install from the Codex plugin directory.

Do I need a NightVision account to use them?

Yes. The skills drive the NightVision CLI, which authenticates against your NightVision account with nightvision login. You can start free without talking to anyone.

What is the difference between the skills and the NightVision MCP server?

The skills teach an agent to operate the NightVision CLI, so the work happens through shell commands you can read. The MCP server exposes NightVision as tools the agent calls directly, so it can start a scan, wait for it, and read structured findings back without parsing terminal output. The skills install from the Codex plugin directory; the MCP server installs from the GitHub repository.

Can the agent scan an application on localhost or a private network?

Yes. The skills run the NightVision CLI in your environment rather than calling a hosted service on your behalf, so anything your machine can reach is a valid target, including localhost, internal staging hosts, and services that never face the public internet.

Do the skills send my source code anywhere?

API discovery is deterministic static analysis performed by the NightVision CLI against the repository you have checked out. The skills add no new credential store and no new network destination: they use the CLI session you already established, which talks to the NightVision API your CLI is configured for.

Are these the same skills that work in Claude Code?

Yes, from the same agent-neutral source. Codex installs them from the plugin directory, and Claude Code installs the same four from the NightVision repository.