self-customize

Modifies agent environment by installing packages, adding MCP servers, and delegating code changes to builder agents.

1|Updated May 22, 2026
One-click install
npx skills add https://github.com/roseDwayane/LocalizeAgenticSys --skill self-customize-rosedwayane
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: self-customize
Source: https://github.com/roseDwayane/LocalizeAgenticSys/tree/main/nanoclaw/container/skills/self-customize
Command: npx skills add https://github.com/roseDwayane/LocalizeAgenticSys --skill self-customize-rosedwayane

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Agents often need new capabilities—system tools, MCP servers, or source code changes—but making those changes ad hoc risks unreviewable edits and broken environments. This Skill provides a structured decision tree and safe workflows for modifying your own agent environment with proper approval gates and reviewable change boundaries. ## Core Features & Use Cases - Decision-tree routing: Directs each change type (workspace files, apt/npm packages, MCP servers, source code) to the correct workflow with the right approval level. - Builder agent delegation: Spins up a dedicated builder agent via create_agent for source code and Dockerfile changes, enforcing diff size limits (200 new / 150 modified lines) so every change stays human-reviewable. - Approval-gated installs: Routes system package and MCP server additions through install_packages and add_mcp_server, which require admin approval and trigger automatic image rebuilds or container restarts. - Use Case: A user asks the agent to transcribe audio. The agent checks for ffmpeg, requests it via install_packages with a stated reason, waits for admin approval, and tests transcription after the automatic container restart. ## Quick Start Ask the agent to add a new capability, such as "install ffmpeg so you can transcribe audio" or "add an RSS reading tool", and it will follow the appropriate self-customization workflow.

Frequently Asked Questions about self-customize

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I add a new tool or capability to my agent?

Describe the capability you need and the agent follows a decision tree: workspace files are edited directly, system packages go through install_packages with admin approval, MCP servers use add_mcp_server, and source code changes are delegated to a builder agent.

How do I add an MCP server to my agent?

Use the add_mcp_server command with the server name, command, and arguments. The request requires admin approval; once approved, the container restarts with the new server wired up, with no image rebuild needed since bun runs TypeScript directly.

Why should code changes be delegated to a builder agent instead of editing directly?

Delegating to a builder agent creates a reviewable boundary and keeps the main session focused. Builder agents enforce diff limits of 200 new or 150 modified lines per task, forcing large changes into small, human-reviewable chunks.

Can I edit CLAUDE.md directly to change agent behavior?

No, the composed CLAUDE.md is read-only and regenerated on every spawn. Write persistent customizations to CLAUDE.local.md in the workspace instead, which is persisted on the host and requires no approval.

When should I not use self-customization workflows?

Avoid them for one-off tasks that can be done in the workspace, ambiguous requests that need clarification first, or unproven ideas. Prototype in the workspace with pnpm install before promoting to a container-level package install.