rezzminator/tool-visibility-controller

Claude Code plugin for per-agent tool visibility — hide and refuse subagents, skills, MCP and built-in tools per loop
Contents: Why · Quick start · Usage · How it works · Logs · Limits · FAQ · Help · Contributing · Built with Professor · License · Changelog
One frontmatter field decides, per loop, what the main chat and each sub-agent type sees in its listings and may call:
markdown
name: researcher
description: Searches the web and summarises sources.
visibility:
visible-to:
include: [research-lead] # only research-lead may see and spawn this agent
mcp:
include: [harvester/] # this agent's loop sees only these MCP tools
tools:
exclude: [Bash] # and every built-in tool but Bash
The main chat's agent listing, before and after:
diff
Available agent types for the Agent tool:
general-purpose: General-purpose agent for researching complex questions. (Tools: )
research-lead: Plans a research question and delegates it. (Tools: Agent, Read)
-- researcher: Searches the web and summarises sources. (Tools: Read, WebFetch)
If the main chat dispatches researcher by name anyway, the spawn is refused with:
text
Agent type researcher is not available to the main chat: visibility.visible-to in ~/.claude/agents/researcher.md admits only research-lead. Delegate the task to research-lead, or choose another agent type.
## 🤔 Why
Every loop gets the tools its job needs, not all of them. Claude Code lists every agent, skill, MCP tool and deferred tool to every loop. A helper agent, a dangerous MCP server or a heavy skill now shows up only where you allow it.
Hidden and refused, not just hidden. Each listing is filtered per loop, and a call by name from a loop that may not use the item is refused with a message naming the rule and what to do instead.
Parent → child chains keep working. permissions.deny blocks a tool in every loop, the orchestrator included. Here the rule is per caller: research-lead spawns researcher, the main chat cannot.
Things you do not own are covered too. Built-in agents (Explore), plugin agents and skills, MCP servers and built-in tools take their rules from a scope file, tool-visibility-controller.json.
Every other loop's context gets lighter, and the prompt cache holds. A hidden entry leaves the listing; every other byte stays identical.
It never blocks your work by accident. Any error is logged and the listing, the call and the spawn go on as they were; a loop or hook failure that leaves an action unchecked says which action went unchecked.
## 🚀 Quick start
Function hooks are experimental; turn them on before Claude Code starts:
bash
export CLAUDECODEENABLEFUNCTIONHOOKS=1
Install from its own marketplace:
bash
claude plugin marketplace add rezzminator/tool-visibility-controller
claude plugin install tool-visibility-controller@professor
One step, or from inside a session
Where claude plugin install --help lists --marketplace, one command adds the marketplace and installs:
bash
claude plugin install tool-visibility-controller --marketplace rezzminator/tool-visibility-controller
Inside a session:
text
/plugin marketplace add rezzminator/tool-visibility-controller
/plugin install tool-visibility-controller@professor