Tool limits

What to do when your client caps how many tools it loads.

https://api.whippy.co/mcp advertises all 256 Whippy tools. Most clients handle this fine, but some cap how many tools they send to the model, and long tool lists use up the model's context.

There are two ways to keep the tool list small.

Option 1: Turn off tools you don't need

Most clients let you switch individual tools on or off for a server. Keep the read-only tools you use most (for example search_contacts, filter_contacts, search_conversations, search_messages, list_channels, list_tags) and turn off whole areas you don't use, such as AI Agent testing or developer webhooks.

Option 2: Use the compact endpoint

https://api.whippy.co/mcp/agent uses the same sign-in and the same tools, but advertises about 30 tools instead of 256:

  • The most-used tools, registered directly: search_contacts, filter_contacts, show_contact, list_channels, show_channel, list_agents, search_conversations, show_conversation, search_messages, list_tags, list_contact_lists, list_campaigns, list_sequences, list_teams, search_users, show_organization, list_templates, list_forms, show_form, list_form_field_types, list_organization_variables, list_segments, create_segment, list_custom_objects, and list_distinct_custom_object_property_values.
  • Discovery tools that reach every other tool:
    • search_tools finds tools by keyword.
    • describe_tools returns a tool's full description and input schema.
    • run_tool runs any tool from the tool reference by name.
  • list_skills and get_skill, which load Whippy's step-by-step playbooks for tasks such as building WQL segments or reviewing calls.

To use it, enter https://api.whippy.co/mcp/agent as the server URL in place of https://api.whippy.co/mcp. Everything else in the client guides stays the same.

The compact endpoint also lists search_web and fetch_web_page. They only work inside Whippy's own AI Command chat and return an error when called from other clients.

The trade-off: when a tool isn't registered directly, the model takes an extra step or two (search_tools, then run_tool) to call it, and some clients show those steps as generic run_tool approvals rather than the underlying tool name. Approve run_tool calls with care, because it can run destructive tools.


Did this page help you?