MyClaws Logo MyClaws

Agent Configuration

Create, edit, and manage Agents; understand system prompts, tool whitelists, built-in tools, prompt boosts, and visibility.

Updated Β· Mon Jun 22 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

On this page 9
  1. When to create a new Agent
  2. Creating a minimum-viable Agent
  3. Writing the system prompt (three-block style)
  4. Tool whitelist β€” keep open or tighten
  5. Agent Prompt Boost
  6. Public / private / built-in visibility
  7. Home-screen Agents and personal tags
  8. Risk and boundaries
  9. Related

Agent Configuration

In a nutshell: an Agent is a combination of "role + tool whitelist + model preferences" that determines what it can do in a session, what it can see, and how it responds.

When to create a new Agent

  • A dedicated Agent for Code Review, monitoring, or copywriting.
  • Restrict an Agent from using bash or network access.
  • Share an Agent with your team without exposing the system prompt.

Creating a minimum-viable Agent

  1. Settings β†’ Agent β†’ New.
  2. Fill in name, description, and icon.
  3. Write the system prompt (recommended structure: role / objectives / output format).
  4. Choose enabled tools (all selected by default).
  5. Set visibility (public / private).
  6. Save β€” available immediately in new conversations.

Writing the system prompt (three-block style)

  • Start with "You are …" to define the role.
  • Follow with "Core objectives" and "Output format".
  • Avoid hard-coding machine configuration in the prompt β€” use Memory and Skill instead.

Tool whitelist β€” keep open or tighten

Built-in tools include but are not limited to:

  • bash: execute shell commands (supports attached desktop mode, see Desktop Power)
  • write / read / edit: file read/write
  • web_search: user-level search (see LLM & Credentials)
  • todo_write: maintain the Agent's own todo list
  • memory_recall / memory_write / memory_update: access long-term memory
  • skill: load the instruction body of a host-local Skill
  • ask_user_question / notify_user: interactive questions and notifications

All tools are enabled by default; restrict per task to reduce accidental use.

Agent Prompt Boost

Admins maintain a "global prompt boost template" in Admin; all enabled Agents share it. Users can also define their own global template (appended after the admin template). Each Agent has an explicit opt-in toggle, enabled by default.

Public / private / built-in visibility

  • Public: visible to all users.
  • Private: visible only to the creator.
  • Built-in: created by admins; regular users can use the Agent but cannot see the system prompt.

Home-screen Agents and personal tags

Each user can maintain personal tags and a home-screen toggle for any visible Agent. The conversation entry point shows only "home-screen Agents" by default, with a one-tap switch to "all Agents".

Risk and boundaries

Do not put credentials in the system prompt

Credentials should be configured in LLM Connection. The prompt is visible to the LLM β€” storing credentials in plain text is a security risk.

Disabling ask_user_question prevents the Agent from talking back

This is the trade-off for "unattended" mode. Trigger-driven sessions can use it this way, but keep it enabled for human-in-the-loop scenarios.

Built-in Agent system prompts are hidden but editable

Admin changes take effect immediately for all users β€” evaluate the blast radius carefully.

Continue reading