Skip to content

README.md

Latest commit

 

History

History
86 lines (66 loc) · 3.92 KB

README.md

File metadata and controls

86 lines (66 loc) · 3.92 KB
Raw

AgentFox plugin for Claude Code

The operator plugin for the AgentFox control plane. It packages the product's knowledge as skills, commands, subagents and safety hooks, so a person can say "get my support agent governed" or "get me ready for the SOC 2 audit" and a coding agent can do it correctly. Without it, they'd have to learn 13 CLI commands and their subcommands, 200+ API routes and 40-odd docs pages first.

Install

Claude Code, from GitHub:

claude plugin marketplace add architsharm/agentfox
claude plugin install agentfox@agentfox

Claude Code, from a local clone (no install, for trying it or developing it):

claude --plugin-dir ./plugins/claude-code

Any other agent (Codex, Cursor, Gemini CLI, Aider, …). Point it at plugins/shared/AGENTS.md, the runtime-neutral copy of this plugin's rules, skills and reference. Everything is plain markdown with relative links, and the skills follow the SKILL.md folder convention.

The plugin drives the agentfox CLI, so the product has to be installed where the agent runs:

pip install "git+https://github.com/architsharm/agentfox.git"

What you get

Type the command What happens
/agentfox:tour Safe, offline tour of the product in a scratch database
/agentfox:start [path] Scan your code, add agentfox.auto() in observe mode, show what it sees
/agentfox:status Read-only posture: doctor, open findings, policy modes, agent states
/agentfox:findings [severity] Grouped triage with a recommended action for each finding
/agentfox:policy <what you want> Draft, validate, lint and simulate a policy; promotion only on your say-so
/agentfox:guardrail "<business rule>" Plain English → an executable business guardrail, in observe mode
/agentfox:gate [suite] Eval regression gate and a CI workflow for your repo
/agentfox:redteam <agent> Adversarial probes, results explained, fixes proposed
/agentfox:evidence [agent] Verify the audit chain and export an auditor-ready package
/agentfox:contain <agent> [reason] Incident response: quarantine (with confirmation), blast radius, evidence
/agentfox:proposals [id or status] Review what the improvement loop wants to change; applying anything needs your say-so
/agentfox:plugin-check Check the plugin against the live CLI and repo (for maintainers)

MCP server. The plugin also starts the MCP server (its .mcp.json uses the older spelling agentfox mcp serve, an alias of agentfox serve mcp), which gives any MCP client 27 read-only tools: posture, findings and how often each one has recurred, the improvement loop's change proposals, policy validate and simulate, guard a piece of text, analyse a SQL/shell/HTTP action, audit verify, compliance status, guardrail tests and more. Nothing there decides, applies or rolls back a change. Other MCP clients can run it directly:

agentfox serve mcp

Subagents (the model delegates to them, or you ask for one by name):

  • governance-auditor: a read-only posture review that ends in a written report.
  • policy-author: drafts and simulates policy. It never promotes one.
  • integration-engineer: wires guardrails into your code in observe mode.

The safety hook turns every command that changes what gets blocked into a permission prompt with a plain-language reason. That covers policy enforce, agents kill, demo, proposals apply, proposals rollback, --mode enforce and --submit.

How it's organised

See STRUCTURE.md: four layers of markdown, one home per fact, and a drift checker that keeps it honest. AGENTS.md, skills/ and reference/ here are copies of plugins/shared/, because an installed Claude Code plugin cannot read files outside its own directory; edit the shared originals, then refresh the copies:

uv run python scripts/check/plugins.py --write