Boost (The Optimizer)

Audit and refine existing assistant configurations by comparing each agent's role (system prompt and description) to its attached tools.

Instead of just blindly adding missing tools, boost operates as an Optimizer that prevents Tool Bloat by recommending both additions (Add) of missing essential tools and removals (Remove) of unnecessary or conflicting tools.

Not This Skill

Skill	Use for
recruit	Create a new specialist when none exists
tool-installer	Register or import MCP servers that are not in inventory yet
skill-deployer	Install or manage bundled skills (docx, deep-research, etc.)
agent-init	Generate workspace agents.md
libragent-harness-reference	LibrAgent prompt layers / session isolation facts

Critical API Rules

externalMcpServers replaces the full list

agent__updateAgent replaces the externalMcpServers array entirely — it does not merge it.

When proposing updates:

• Add: Add the new tool IDs to the existing list.
• Remove: Filter out the bloated tool IDs from the existing list.

Compute the final list precisely:

next_external = (current_externalMcpServers + recommended_add_ids) - recommended_remove_ids

Never send only the added tools, or you will accidentally delete all pre-existing tools.

Server IDs, not display names

Resolve names like github or search to cuid2 server IDs via tool__listServers before making updates.

---

Workflow (Audit & Optimize)

1. Agent Audit

Scan the current state of assistants and tools:

1. Retrieve Configurations:

   agent__listAgents({ "type": "configs", "verbose": true })
   

   For each assistant, record its id, name, description, systemPrompt, externalMcpServers, and builtinCapabilities.

2. Retrieve Tool Inventory:

   tool__listServers({ "availability": "inventory" })
   

   Get all available tool servers in the environment.

2. Optimization Logic (Add & Remove)

For each assistant, perform a semantic check of its role against its current toolset:

• Add (Augmentation): If the assistant has a clear functional need (e.g., Finance Specialist) and a highly relevant tool is available in the inventory (e.g., Market Search API) but not attached, propose adding it.
• Remove (Pruning / Tool Bloat Prevention): If a tool is attached to the assistant but is completely unrelated to its domain (e.g., github attached to a Finance Agent, or finance-api attached to Coding Expert), propose removing it.
  • Reasoning: Too many tools increase cognitive load, waste token context, and lead to tool selection errors. Propose removals under the banner of minimizing cognitive load.
• Keep (Retention): Retain core tools that are aligned with the assistant's primary focus.

Explicit Calculation Step: Calculate the final tool list precisely using a mathematical union and difference to prevent accidental tool deletion:

final_list = (current_list + add_list) - remove_list

Double-check that all tools intended to be retained (Keep) are explicitly present in the resulting final_list payload for agent__updateAgent.

3. Draft Diagnostic Report

Before applying any changes, present a structured Audit & Optimization Report to the user:

### [Assistant Name] (`[Assistant ID]`)
* **Role Summary:** [1-sentence summary of description/systemPrompt focus]
* **Proposed Optimization:**
  * **Keep:** `[tool-label-1]`, `[tool-label-2]`
  * **Add (Recommended):** `[tool-label-3]` (Reason: [brief explanation of domain relevance])
  * **Remove (Prune):** `[tool-label-4]` (Reason: High cognitive load / Irrelevant to core role)
* **Action:** Update `externalMcpServers` to `[final merged & pruned ID list]`

Provide a high-level summary of the overall status (e.g., "3 assistants audited, 1 needs additions, 2 need pruning").

4. Apply Updates

Once the user approves the optimization proposal:

For each assistant requiring changes:

agent__updateAgent({
  "id": "<assistant-id>",
  "externalMcpServers": ["<final-resolved-id-1>", "<final-resolved-id-2>"],
  "builtinCapabilities": ["<final-resolved-builtins>"]
})

Note: You cannot update the configuration of the assistant you are currently running as. Use a root/parent session or another agent (e.g., App Wizard) to run the update.

5. Verification & Handoff

Verify that the changes were correctly applied:

1. Check the updated configurations:

   agent__listAgents({ "type": "configs", "verbose": true })
   

2. Report the successful optimization to the user, highlighting the added and pruned tools. Tell them that existing sessions will pick up these changes upon restart or next configuration resolution.

---

Guidelines

• Prevent Tool Bloat Proactively: Treat tool removal with equal importance to tool addition. Do not let assistants become "jack of all trades" with bloated tool lists.
• Role-Tool Alignment: Every tool attached to an assistant must be explicitly justified by its description or systemPrompt.
• Atomic updates: Ensure you calculate the union and difference correctly before calling agent__updateAgent. Never send an incomplete list that deletes required tools.
• English Prompt Preservation: Boost only modifies tool configurations (externalMcpServers and builtinCapabilities). Do not modify the systemPrompt text unless separately requested.

Builtin Tools

Step	Tool
List assistants	agent__listAgents({ "type": "configs", "verbose": true })
Inventory MCP	tool__listServers({ "availability": "inventory" })
Apply changes	agent__updateAgent({ id, externalMcpServers, builtinCapabilities? })
Create new specialist	recruit (separate workflow)