Mantis: a sequential security skill suite for coding assistants

Who it's for

Mantis is for people who want to do security-oriented work with a coding assistant: a decoupled, sequential, security-centered skill suite built to work with Coding Agents. The main path is to install the skills, then enter slash commands one by one, not to hand the whole flow to auto-approval.

If your first step is planning rather than changing code at once, the repository's example is /mantis-plan. The docs require entering these commands separately, and not using --yolo or --dangerously-skip-permissions. That fits people who will stop at each step, read the output, and then decide whether to continue.

It is not an “auto-fix-the-bug tool” for an unisolated, unrestricted environment. The repository states that this suite generates and runs auto-generated code, which may be unstable and may take unexpected actions. It is also not for treating the output as a confirmed security conclusion: every finding must be checked in person by a security expert before it is reported outside.

First try in 3 minutes

The official install command is this one: npx skills add google/mantis. The repository URL is https://github.com/google/mantis Follow only this page; do not rewrite it into an undocumented installer.

Preconditions: you already have a usable coding-assistant environment, and this machine can run npx. The install location is that assistant's skills directory (written by this official command). After install, first confirm that slash commands starting with mantis appear in the command list, especially /mantis-plan.

Low-risk first task: enter only /mantis-plan, do planning, and do not turn on auto-approval at the same time. Explicitly do not add --yolo or --dangerously-skip-permissions. Before you trigger it, confirm the current workspace can accept a planning discussion, and that you have not authorized it to run unread generated code. The success signal is that the assistant starts asking questions or presenting a plan along Mantis planning steps, rather than editing files or starting unexplained execution.

The planning result is only a plan. If the assistant next wants to generate or run code, stop first and look at what it would change. The repository does not name every command after /mantis-plan, so the first try can stop once a plan is visible and nothing has auto-executed.

Use cases

Case 1: Interactive security planning. After install, enter /mantis-plan by itself, move step by step, and do not use --yolo or --dangerously-skip-permissions. Fits when you want to set the check scope first, then decide whether to continue. The docs do not give a full command list, only the /mantis-plan example; do not invent other slash commands. If the command does not appear, go back to https://github.com/google/mantis and check whether npx skills add google/mantis succeeded.

Case 2: Treat auto-generated check or patch code as a draft still needing verification. The repository states that this suite generates and runs auto-generated code, which may be unstable or take unexpected actions. That fits teams that already have an isolated environment and a security expert who will read the result. Every finding must be checked in person by a security expert before it is reported. The docs do not describe how to build the isolation environment, and they do not give a report template; until human verification is done, do not send the output outside or treat it as a formal vulnerability ticket.

Usage limits

Install limit: the official docs only give npx skills add google/mantis; do not use undocumented install methods. Operation limit: enter slash commands one at a time; do not treat --yolo or --dangerously-skip-permissions as a speed-up. Execution limit: the suite generates and runs code that may be unstable and may take unexpected actions. Conclusion limit: findings must be checked in person by a security expert before they are reported. Material limit: it is written as decoupled, sequential security skills for Coding Agents; it does not name which specific products are supported, and it does not promise coverage of every language or hardware platform.