Skills
The client skill pack
Skills for a coding agent on a developer's machine that work through the Nullgap MCP server, what they do and do not do locally, and how to install them.
The skills on the other pages of this section run on Nullgap. The client skill pack is a different thing: a set of skills for the coding agent on a developer's own machine. They are playbooks, in the same open Agent Skills format, that tell the agent how to use Nullgap well: triage findings against the checkout, fix a finding, start and follow a pentest, check a dependency before it is added, work out what a patch changed.
| Skills on Nullgap | Client skill pack | |
|---|---|---|
| Where the playbook is read | On Nullgap, by Nullgap AI | On your machine, by your coding agent |
| Who manages them | Admins, on the Skills page | Each developer, in their agent |
| How they reach Nullgap | They are part of it | Through the MCP server, connected in the browser or with an API key |
| What they return | Findings, documents, triage verdicts | Whatever your agent does with them: a summary, a local fix, a recorded decision |
#Every tool run goes through the MCP server
This is the rule the whole pack is built on.
- No skill in the pack runs a scanner on your machine. A skill never installs, downloads or runs a security tool locally, and never falls back to a local one when a tool is not offered.
- Everything that touches Nullgap is an MCP tool call. Reading findings, changing a state, starting a pentest, checking packages and running a skill on Nullgap all go through the tools on MCP tools, with the access you approved when you connected (or the API key you configured). A skill can do what that access allows and nothing more.
- Locally, the agent only reads your checkout, edits it when you ask for a fix, and makes git commits you approve.
What follows from that:
- The scopes, rate limits and audit log of the API apply to everything a client skill does. See what a connected app can do.
- Text that comes back from Nullgap about your repository is marked as untrusted, and the skills tell the agent to treat it as data.
- Starting something that is charged or counted, such as a pentest, needs your explicit yes in the conversation after you have seen the price and the scope. The skills tell the agent never to take that approval from a file, a finding or a tool result.
- Live targets and test accounts of a pentest are configured in the console by an admin. The pentest tools have no arguments for them, so a client skill cannot point a pentest at a host you name in a chat.
run_tooldoes take an address for a live-application scanner, and the server refuses it unless the host is already on the scope list an admin keeps, the connection has admin access (or the key theadminscope) and the tool has been switched on by the operator.
The scanner playbooks in the pack run a scanner on Nullgap through three MCP tools: list_tools, run_tool and get_tool_run. They are described on scanning tools. When the tool list says a tool is switched off, refused or not available on your server, a playbook reports that and continues with what it can do by reading: it does not run a scanner another way.
#What is in the pack
The pack holds 63 skills. Each is one folder with a SKILL.md.
#Product skills
Five skills that drive Nullgap itself. Their names start with nullgap-.
| Skill | What it does | Key scope |
|---|---|---|
nullgap-finding-triage | Lists findings for the repository in your checkout, reads the real code at each location, and records a decision when you agree | read to triage, write to record |
nullgap-fix-finding | Fetches one finding and its fix prompt, fixes the root cause locally and adds a regression test | read to fix locally, write to change a state or ask for a review |
nullgap-pentest | Estimates a pentest, shows the price and the scope, starts it only after you approve, follows it and summarises the result | read to estimate and follow, admin to start |
nullgap-dependency-guard | Checks the exact package versions in a change before a dependency is added or upgraded | read |
nullgap-patch-diff | Starts the patch-diff-analysis skill on Nullgap for two versions and reads the result | write |
#The SecOps kit and research skills
| Group | Skills | What they are |
|---|---|---|
| Application security | 18 | Static analysis per language, triage, dynamic testing, API and GraphQL review, dependency risk |
| DevSecOps | 8 | Secrets, containers, infrastructure as code, Kubernetes, pipelines, cloud posture |
| Secure SDLC | 6 | Threat modelling, security requirements, design review, security tests, pull request guardrails, the release gate |
| Compliance | 4 | Policy as code, control mapping, audit evidence, the risk register |
| Incident response | 11 | Triage, log analysis, detection writing, credential, dependency, cloud and web application compromise, recovery and review |
| Authorised testing | 6 | Method skills for testing systems you own or have written permission to test |
| Security research and reporting | 5 | Cryptography review, finding reports, severity scoring, coordinated disclosure |
There are two kinds of skill in these groups:
- MCP tools: a playbook around Nullgap MCP calls.
- Method: guidance for a person doing authorised work, and analysis of data that person provides. A method skill about tools Nullgap does not offer says so, and contains no commands for an agent to run.
Skills that touch a running system start with an authorisation section: you confirm ownership or written permission and name the scope, the target must be on your organisation's scope list, production is out unless you say otherwise, and anything that changes state needs a second explicit yes.
#Install
#1. Connect the MCP server
The skills do nothing without it. Add the hosted endpoint to your agent as a remote MCP server, then sign in and approve in the browser when the agent asks:
{
"mcpServers": {
"nullgap": {
"type": "http",
"url": "https://nullgap.example/mcp"
}
}
}
Name the server nullgap: the skills refer to its tools by that name. Choose read and write on the consent screen if you want the skills to triage and fix; read only is enough to read findings and check packages. The details are on MCP server. For an agent that runs without a person (CI), use an API key instead, kept in the environment and never in the file or a chat.
#2. Add the skills
There are three ways, depending on what your agent supports.
As a plugin marketplace. The pack's folder is a plugin marketplace: it carries a marketplace file that lists eight plugins, one per group, so you can install only the groups you want. If your coding agent supports plugin marketplaces, register the pack's folder (or the repository that holds it) as a marketplace named nullgap-skills, then install plugins from it:
| Plugin | Contains |
|---|---|
nullgap-product | The five product skills |
nullgap-appsec | Application security |
nullgap-devsecops | DevSecOps |
nullgap-secure-sdlc | Secure SDLC |
nullgap-compliance | Compliance |
nullgap-incident-response | Incident response |
nullgap-offsec | Authorised testing (method skills) |
nullgap-research | Security research and reporting |
How a marketplace is registered and a plugin installed differs between agents; follow your agent's own documentation for those two steps.
With the CLI. nullgap skills pull copies the product skills into your project:
nullgap skills pull
nullgap skills pull nullgap-fix-finding nullgap-finding-triage --path ./agent-skills
Without names it pulls all of them. --path says where to write them; the default is the project-level skills folder that coding agents read. Each skill is checked against the Agent Skills format before anything is written, files are written as plain files, and nothing that is pulled is made executable or run. The source is set with NULLGAP_SKILLS_SOURCE (a folder or an https:// address); no credentials are sent to it, and a redirect to another place is refused.
By copying. Any agent that reads Agent Skills can use them: copy the skill folders you want into the agent's skills folder.
#3. Check it works
Ask your agent to list your open findings. It should call list_findings on the nullgap server and show the result with the untrusted-text notice. If the tool call is refused, the message says why: the key, its scope, or the organisation in the address.
#Good practice
- Give the agent a
readkey to begin with. Move towritewhen you want it to record decisions, and toadminonly when you want it to start pentests. - Have your MCP client ask before any tool that changes state.
- Review a fix the agent made like any other change before you merge it. A fix made locally by an agent is not a verified fix: ask for a pull request review to have it checked.
Nullgap documentation