Supported agents

The agents Cefense writes instructions for, and how to wire up one that is not listed.

3 sections2 min read

cf skill install detects the agents configured in a repository and writes to each of them, in the format that tool expects, at the path it actually reads.

Where each one is written

AgentPathFormat
Claude Code.claude/skills/cefense/SKILL.mdSkill frontmatter with a name and description
Cursor.cursor/rules/cefense.mdcMDC with alwaysApply: false
GitHub Copilot.github/instructions/cefense.instructions.mdapplyTo: "**"
Google Antigravity.agents/rules/cefense.mdPlain markdown
Windsurf.windsurf/rules/cefense.mdA model decision trigger
Devin Desktop.devin/rules/cefense.mdA model decision trigger
Cline and Roo Code.clinerules/cefense.mdPlain markdown
Gemini CLIGEMINI.mdA delimited section, full guide at .cefense/SKILL.md
Codex, Amp, OpenCode, JulesAGENTS.mdA delimited section, full guide at .cefense/SKILL.md
Choose, or take them allbash
cf skill install
cf skill list
cf skill install claude cursor
cf skill install --all
cf skill install --global

--global writes for your user rather than for the repository, and works for the agents that have a user-wide location: Claude Code, Gemini CLI, and AGENTS.md. Asking for it with any other agent is refused with global_unsupported rather than written somewhere arbitrary.

Anything else

There is nothing special about the listed agents. Any tool that can run a shell command and parse JSON can drive Cefense. Give it the skill document and let it use --agent.

Get the document, put it wherever your tool readsbash
cf skill show > my-agent-rules.md
curl -fsSL https://cefense.com/skill.md

What a custom integration needs to know

  • Pass --agent to every command, and --repo owner/name to every repository-scoped one.
  • Name the organization with --org <slug>, or set CEFENSE_ORG once. Skip it only when the account belongs to exactly one.
  • Parse one line of JSON from stdout. Branch on ok, then on error.code.
  • Follow next rather than composing commands, where one is offered.
  • Never add --yes without a person having agreed to that specific action.

The full contract is in Agent mode, and the payload shapes are in Agent commands.

Have a correction or need help?

Email the docs team