Native Tool Mode
MCP clients are beginning to support Anthropic’s tool search protocol—defer_loading to mark tools as discoverable rather than always-loaded, and tool_reference content blocks to expand discovered tools into full definitions. We’ve added native mode to take advantage of this when clients support it.
From Meta-Execution to Native Tools
In meta mode, agents discover operations through find_api, learn interfaces through learn_api, and execute them through call_api. The execution path is uniform—every API call routes through the same meta-tool regardless of which operation it invokes.
Native mode preserves discovery and learning but changes execution. Operations become individual MCP tools that agents call directly. Instead of call_api("github_issues_create", { title, body }), agents call github_issues_create({ title, body }). The tool exists as a first-class entity in the agent’s tool list.
The find_api tool adapts to this change. In native mode, search results return as tool_reference content blocks. When the client receives these references, it automatically expands them into full tool definitions. The agent searches, the tools appear, and calling them is immediate.
Enabling Native Mode
Add ?tools=native to your MCP connection URL:
claude mcp add --transport http toolcog-github 'https://mcp.toolcog.com/github?tools=native'Without the parameter, connections default to meta mode. Client support for tool search remains limited—Claude Code supports it today, Claude.ai support is coming—so native mode is opt-in rather than automatic.
Tool Metadata
An LLM analyzes each operation during indexing and generates the metadata that MCP tools need. It produces a human-readable title that clients display during invocation, and behavioral annotations that classify whether the operation is read-only, destructive, or idempotent. Clients use these annotations to decide when to prompt for confirmation, whether failed calls can be safely retried, and how to sandbox execution. Publishers upload a spec; the metadata appears without manual annotation.
Operation Pinning
Not every operation should require discovery. Some are used constantly—the handful of tools an agent reaches for in most sessions. Pinning marks these operations as always-loaded.
Pinned operations have defer_loading: false. They appear in the agent’s context from connection time, no search required. Unpinned operations have defer_loading: true and surface through find_api when needed.
Pin and unpin operations from the ellipsis dropdown menu on operation list items in the web UI.
Primary Bridges
Each owner can designate one bridge as primary. The primary bridge serves as the default when connecting to mcp.toolcog.com/{owner} without specifying a resource.
For example, mcp.toolcog.com/github resolves to the github organization’s primary bridge. Set a bridge as primary from the ellipsis dropdown menu on API or catalog list items.
Bridge Identity
The server now presents itself using the bridge’s identity rather than generic Toolcog branding. The bridge’s title, summary, and description appear in your MCP client, and meta tools include the bridge name in their titles. When you have multiple MCP servers connected, you can tell them apart at a glance.