fix(mcp): avoid repeated tool schema materialization

.changeset/calm-mcp-catalogs.md

@@ -0,0 +1,6 @@
+---
+"agents": patch
+"@cloudflare/think": patch
+---
+
+Cache MCP JSON Schema conversion for the current catalog on each live connection, and let Think agents skip direct MCP AI-tool exposure when those tools are exposed through Code Mode or another mechanism outside Think's automatic tool set.

docs/agents/mcp-client.md

@@ -305,17 +305,20 @@ Once connected, access the server's capabilities:
 
 ### Getting Available Tools
 
+Use `listTools()` when you need to discover or inspect the raw MCP catalog without preparing tools for an AI SDK model call:
+
 ```typescript
-const state = this.getMcpServers();
+const tools = this.mcp.listTools();
 
-// All tools from all connected servers
-for (const tool of state.tools) {
+for (const tool of tools) {
   console.log(`Tool: ${tool.name}`);
   console.log(`  From server: ${tool.serverId}`);
   console.log(`  Description: ${tool.description}`);
 }
 ```
 
+`getMcpServers().tools` exposes the same raw tool shape as part of the full client state sent to connected applications. Neither API converts JSON Schemas to Zod.
+
 ### Resources and Prompts
 
 ```typescript
@@ -361,7 +364,13 @@ async function chat(prompt: string) {
 }
 ```
 
-> **Note:** `getMcpServers().tools` returns raw MCP `Tool` objects for inspection. Use `this.mcp.getAITools()` when passing tools to the AI SDK.
+> **Note:** Use `this.mcp.listTools()` or `getMcpServers().tools` for discovery and inspection. Call `this.mcp.getAITools()` only when preparing tools for the AI SDK because it converts each MCP input and output JSON Schema to Zod.
+
+### AI tool schema conversion lifetime
+
+`getAITools()` reuses converted schemas while a live connection keeps the same current catalog. Repeated filtered or unfiltered calls still return fresh tool records and execute closures. Connections excluded by a filter are not converted.
+
+The next call converts schemas again after discovery replaces a catalog or the live connection changes. If a custom integration mutates a schema, assign a new `inputSchema` or `outputSchema` object, or replace the catalog array, so `getAITools()` detects the change.
 
 ## Managing Servers
 

docs/think/index.md

@@ -857,11 +857,27 @@ path.
 | `workspaceBash`            | `true`                           | Include or configure the default workspace `bash` tool                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
 | `fetchTools`               | `false`                          | Opt-in allowlisted, read-only HTTP fetch tools (`fetch_url` + per-binding `fetch_<name>`). Set to a config object; see [Fetch tool](#fetch-tool)                                                                                                                                                                                                                                                                                                                                                                                                                 |
 | `messageConcurrency`       | `"queue"`                        | How overlapping submits behave — see [Client Tools](./client-tools.md)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
+| `includeMcpTools`          | `true`                           | Automatically convert connected MCP tools to AI SDK tools and merge them into model turns                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
 | `waitForMcpConnections`    | `false`                          | Wait for MCP servers before inference                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            |
 | `chatRecovery`             | `true`                           | Wrap turns in `runFiber` for durable execution, including sub-agent turns. Set to `{ maxAttempts, stableTimeoutMs, terminalMessage, onExhausted }` to tune bounded recovery.                                                                                                                                                                                                                                                                                                                                                                                     |
 | `chatStreamStallTimeoutMs` | `0` (off)                        | Opt-in inactivity watchdog: abort a turn whose model stream produces no chunk for this long (measures the gap between chunks, including tool execution — set above your slowest model TTFT + tool, e.g. `120_000`). Emits a `chat:stream:stalled` event; with `chatRecovery` on (the default) the stall routes into bounded recovery (see below) instead of an infinite spinner, and only terminalizes once the budget is exhausted. Override per-turn via `TurnConfig.chatStreamStallTimeoutMs` (returned from `beforeTurn`) for a turn with a known-slow tool. |
 | `contextOverflow`          | `undefined`                      | Opt-in mid-turn context-overflow handling: `{ reactive?, maxRetries?, proactive? }`. Requires `classifyChatError` + a session compaction function. See [Context-window overflow recovery](#context-window-overflow-recovery).                                                                                                                                                                                                                                                                                                                                    |
 
+### MCP tools exposed outside the harness
+
+Think normally calls `this.mcp.getAITools()` while assembling every turn. If you expose MCP tools through Code Mode or another mechanism outside Think's automatic tool set, set `includeMcpTools = false` to avoid direct JSON Schema-to-Zod materialization:
+
+```typescript
+export class MyAgent extends Think<Env> {
+  includeMcpTools = false;
+  waitForMcpConnections = true;
+}
+```
+
+This setting affects only Think's automatic AI SDK tool merge. Connections still register, restore, discover, and wait normally; raw tool listing and calls, Code Mode access, and explicit `this.mcp.getAITools()` calls are unchanged. Returning `activeTools: []` from `beforeTurn` does not provide the same behavior because tool conversion occurs before that hook.
+
+See [Tools](./tools.md#mcp-tools) for MCP setup and alternative exposure paths.
+
 ## Agent Skills
 
 Think supports [Agent Skills](https://agentskills.io/) as on-demand

docs/think/tools.md

@@ -11,7 +11,7 @@ On every turn, Think merges tools from multiple sources. Later sources override
 3. **Extension tools** — tools from loaded extensions (prefixed by extension name)
 4. **Session tools** — `set_context`, `load_context`, `search_context` (from `configureSession`)
 5. **Skill tools** — `activate_skill`, `read_skill_resource`, `run_skill_script` (from `getSkills()`)
-6. **MCP tools** — from connected MCP servers
+6. **MCP tools** — from connected MCP servers (when `includeMcpTools` is enabled)
 7. **Client tools** — from the browser (see [Client Tools](./client-tools.md))
 
 Tools belong to the agent running the turn. For parent-child orchestration,
@@ -182,7 +182,7 @@ beforeTurn(ctx: TurnContext) {
 
 ## MCP Tools
 
-Think inherits MCP client support from the `Agent` base class. MCP tools from connected servers are automatically merged into every turn.
+Think inherits MCP client support from the `Agent` base class. By default, tools from connected MCP servers are converted to AI SDK tools and merged into every turn.
 
 Set `waitForMcpConnections` to ensure MCP servers are connected before the inference loop runs:
 
@@ -197,6 +197,23 @@ export class MyAgent extends Think<Env> {
 }
 ```
 
+If you expose MCP tools through Code Mode or another mechanism outside Think's automatic tool set, disable direct AI SDK tool exposure:
+
+```typescript
+export class MyAgent extends Think<Env> {
+  includeMcpTools = false;
+  waitForMcpConnections = true;
+
+  getModel() {
+    /* ... */
+  }
+}
+```
+
+`includeMcpTools = false` suppresses only Think's automatic `this.mcp.getAITools()` merge. MCP registration, restoration, discovery, waiting, raw catalog access, direct calls, Code Mode connectors, and explicit application calls to `this.mcp.getAITools()` continue to work. This avoids converting every MCP JSON Schema to Zod when the model only uses the raw MCP connection through the connector.
+
+Use the class property instead of `beforeTurn({ activeTools: [] })` for this purpose. `activeTools` limits tools after turn assembly; MCP schema conversion happens while Think assembles the context, before `beforeTurn` runs.
+
 Add MCP servers programmatically or via `@callable` methods:
 
 ```typescript

packages/agents/src/mcp/client.ts

@@ -51,6 +51,28 @@ export type MCPAITool = {
  */
 export type MCPAIToolSet = Record<string, MCPAITool>;
 
+type MCPAIConvertedSchemas = {
+  inputSchema: z.ZodType;
+  outputSchema?: z.ZodType;
+};
+
+type MCPAISchemaSource = {
+  tool: Tool;
+  inputSchema: Tool["inputSchema"] | undefined;
+  outputSchema: Tool["outputSchema"] | undefined;
+};
+
+type MCPAISchemaCacheSlot = MCPAISchemaSource &
+  (
+    | { status: "converted"; converted: MCPAIConvertedSchemas }
+    | { status: "failed"; error: string }
+  );
+
+type MCPAISchemaCacheEntry = {
+  catalog: Tool[];
+  converted: Array<MCPAISchemaCacheSlot | undefined>;
+};
+
 /** Maximum length of a normalized MCP server id. */
 export const MCP_SERVER_ID_MAX_LENGTH = 64;
 
@@ -347,6 +369,11 @@ export type MCPServerFilter = {
  */
 export class MCPClientManager {
   public mcpConnections: Record<string, MCPClientConnection> = {};
+  /** Cache only the current catalog so old schema graphs are not retained. */
+  private readonly _aiToolSchemas = new WeakMap<
+    MCPClientConnection,
+    MCPAISchemaCacheEntry
+  >();
   private _didWarnAboutUnstableGetAITools = false;
   private _oauthCallbackConfig?: MCPClientOAuthCallbackConfig;
   private _connectionDisposables = new Map<string, DisposableStore>();
@@ -1789,70 +1816,130 @@ export class MCPClientManager {
   }
 
   /**
+   * Convert connected MCP tools for the AI SDK. Converted schemas are reused
+   * while a live connection retains the same catalog array and schema-source
+   * identities; tool records and execute closures are rebuilt on every call.
+   *
    * @param filter - Optional filter to scope results to specific servers
    * @returns a set of tools that you can use with the AI SDK
    */
   getAITools(filter?: MCPServerFilter): MCPAIToolSet {
     const connections = this.filterConnections(filter);
+    const entries: [string, MCPAITool][] = [];
 
-    for (const [id, conn] of Object.entries(connections)) {
+    for (const [serverId, conn] of Object.entries(connections)) {
       if (
         conn.connectionState !== MCPConnectionState.READY &&
         conn.connectionState !== MCPConnectionState.AUTHENTICATING
       ) {
         console.warn(
-          `[getAITools] WARNING: Reading tools from connection ${id} in state "${conn.connectionState}". Tools may not be loaded yet.`
+          `[getAITools] WARNING: Reading tools from connection ${serverId} in state "${conn.connectionState}". Tools may not be loaded yet.`
         );
       }
-    }
 
-    const entries: [string, MCPAITool][] = [];
-    for (const tool of getNamespacedData(connections, "tools")) {
-      try {
-        const toolKey = `tool_${tool.serverId.replace(/-/g, "")}_${tool.name}`;
-        const title = tool.title ?? tool.annotations?.title;
-        entries.push([
-          toolKey,
-          {
-            description: tool.description,
-            title,
-            execute: async (args) => {
-              const result = await this.callTool({
-                arguments: args,
-                name: tool.name,
-                serverId: tool.serverId
-              });
-              if (result.isError) {
-                const content = result.content as
-                  | Array<{ type: string; text?: string }>
-                  | undefined;
-                const textContent = content?.[0];
-                const message =
-                  textContent?.type === "text" && textContent.text
-                    ? textContent.text
-                    : "Tool call failed";
-                throw new Error(message);
-              }
-              return result;
-            },
-            inputSchema: tool.inputSchema
-              ? z.fromJSONSchema(
-                  tool.inputSchema as Parameters<typeof z.fromJSONSchema>[0]
-                )
-              : z.fromJSONSchema({ type: "object" }),
-            outputSchema: tool.outputSchema
-              ? z.fromJSONSchema(
-                  tool.outputSchema as Parameters<typeof z.fromJSONSchema>[0]
-                )
-              : undefined
+      const catalog = conn.tools;
+      let cache = this._aiToolSchemas.get(conn);
+      if (!cache || cache.catalog !== catalog) {
+        cache = { catalog, converted: [] };
+        this._aiToolSchemas.set(conn, cache);
+      }
+
+      for (const [index, tool] of catalog.entries()) {
+        const toolName = tool.name;
+        try {
+          const sourceInputSchema = tool.inputSchema;
+          const sourceOutputSchema = tool.outputSchema;
+          let slot = cache.converted[index];
+          if (
+            !slot ||
+            slot.tool !== tool ||
+            slot.inputSchema !== sourceInputSchema ||
+            slot.outputSchema !== sourceOutputSchema
+          ) {
+            try {
+              slot = {
+                status: "converted",
+                tool,
+                inputSchema: sourceInputSchema,
+                outputSchema: sourceOutputSchema,
+                converted: {
+                  inputSchema: sourceInputSchema
+                    ? z.fromJSONSchema(
+                        sourceInputSchema as Parameters<
+                          typeof z.fromJSONSchema
+                        >[0]
+                      )
+                    : z.fromJSONSchema({ type: "object" }),
+                  outputSchema: sourceOutputSchema
+                    ? z.fromJSONSchema(
+                        sourceOutputSchema as Parameters<
+                          typeof z.fromJSONSchema
+                        >[0]
+                      )
+                    : undefined
+                }
+              };
+            } catch (error) {
+              const errorText = String(error);
+              cache.converted[index] = {
+                status: "failed",
+                tool,
+                inputSchema: sourceInputSchema,
+                outputSchema: sourceOutputSchema,
+                error: errorText
+              };
+              console.warn(
+                `[getAITools] Skipping tool "${toolName}" from "${serverId}": ${errorText}`
+              );
+              continue;
+            }
+            cache.converted[index] = slot;
           }
-        ]);
-      } catch (e) {
-        console.warn(
-          `[getAITools] Skipping tool "${tool.name}" from "${tool.serverId}": ${e}`
-        );
+
+          if (slot.status === "failed") continue;
+
+          const toolKey = `tool_${serverId.replace(/-/g, "")}_${toolName}`;
+          const title = tool.title ?? tool.annotations?.title;
+          const description = tool.description;
+          entries.push([
+            toolKey,
+            {
+              description,
+              title,
+              execute: async (args) => {
+                const result = await this.callTool({
+                  arguments: args,
+                  name: toolName,
+                  serverId
+                });
+                if (result.isError) {
+                  const content = result.content as
+                    | Array<{ type: string; text?: string }>
+                    | undefined;
+                  const textContent = content?.[0];
+                  const message =
+                    textContent?.type === "text" && textContent.text
+                      ? textContent.text
+                      : "Tool call failed";
+                  throw new Error(message);
+                }
+                return result;
+              },
+              inputSchema: slot.converted.inputSchema,
+              outputSchema: slot.converted.outputSchema
+            }
+          ]);
+        } catch (error) {
+          console.warn(
+            `[getAITools] Skipping tool "${toolName}" from "${serverId}": ${error}`
+          );
+        }
       }
+
+      // Release removed slots when a public catalog array is spliced in place.
+      cache.converted.length = catalog.length;
     }
+
     return Object.fromEntries(entries);
   }