> ## Documentation Index
> Fetch the complete documentation index at: https://braintrust.dev/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Cloudflare Think

> Trace @cloudflare/think agent runs in Braintrust to debug multi-turn reasoning, LLM calls, and tool use on Cloudflare Workers

If you are a coding agent, prefer the Braintrust [`bt` CLI](/docs/reference/cli/quickstart) for repeatable, scriptable work: running evals, instrumenting code, querying logs, syncing data, managing functions, and configuring coding agents. Use the MCP server for reasoning over Braintrust data in conversation, such as ad-hoc lookups and exploration from your IDE.

[`@cloudflare/think`](https://developers.cloudflare.com/agents/harnesses/think/) is an agent framework for Cloudflare Workers built on Durable Objects. Each agent extends the `Think` class and runs multi-turn reasoning loops, delegating inference and tool execution to the AI SDK. Braintrust traces every `runTurn` call as a task span and nests the AI SDK calls and tool invocations beneath it.

<Note>
  For Cloudflare Workers deployments, use [manual instrumentation](#manual-instrumentation-typescript) with `wrapCloudflareThink()`. The `--import` auto-instrumentation hook only runs under Node, not in the Cloudflare Workers runtime (`workerd`). See the [Cloudflare setup guide](/docs/sdks/typescript/install-and-instrument#cloudflare) for enabling `nodejs_compat` and flushing traces with `ctx.waitUntil()`.
</Note>

<View title="TypeScript" icon="https://img.logo.dev/typescriptlang.org?token=pk_BdcHD9e5SCW3j1rnJkNyMQ">
  <h2 id="setup-typescript">
    Setup
  </h2>

  Install Braintrust alongside `@cloudflare/think`, then set your API keys. Requires `@cloudflare/think` v0.13.0 or later.

  <Steps>
    <Step title="Install packages">
      <CodeGroup>
        ```bash pnpm theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
        pnpm add braintrust @cloudflare/think
        ```

        ```bash npm theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
        npm install braintrust @cloudflare/think
        ```
      </CodeGroup>
    </Step>

    <Step title="Set environment variables">
      ```bash title=".env" theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
      BRAINTRUST_API_KEY=<your-braintrust-api-key>

      # For organizations on the EU data plane, use https://api-eu.braintrust.dev
      # For self-hosted deployments, use your data plane URL
      # BRAINTRUST_API_URL=<your-braintrust-api-url>
      ```
    </Step>
  </Steps>

  <h2 id="manual-instrumentation-typescript">
    Manual instrumentation
  </h2>

  Manual instrumentation is the recommended approach for Cloudflare Workers. Call `wrapCloudflareThink()` on the `@cloudflare/think` module at module scope, then extend the wrapped `Think` class so every agent instance receives Braintrust tracing without any changes to your agent logic. Initialize the logger inside `fetch` with your `env` bindings, then pass `logger.flush()` to `ctx.waitUntil()` so buffered traces ship after the response returns.

  ```typescript title="cloudflare-think-manual.ts" theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
  import * as cloudflareThink from "@cloudflare/think";
  import { createOpenAI } from "@ai-sdk/openai";
  import { initLogger, wrapCloudflareThink } from "braintrust";
  import { tool } from "ai";
  import { z } from "zod";

  interface Env {
    BRAINTRUST_API_KEY: string;
    OPENAI_API_KEY: string;
    THINK_AGENT: DurableObjectNamespace<MyAgent>;
  }

  // Wrap the module at module scope so every agent instance is instrumented
  const { Think } = wrapCloudflareThink(cloudflareThink);

  export class MyAgent extends Think<Env> {
    getModel() {
      return createOpenAI({ apiKey: this.env.OPENAI_API_KEY }).chat("gpt-5-mini");
    }

    getSystemPrompt() {
      return "You are a helpful weather assistant.";
    }

    getTools() {
      return {
        lookup_weather: tool({
          description: "Get the weather for a city.",
          inputSchema: z.object({ city: z.string() }),
          execute: async ({ city }) => ({ city, condition: "sunny" }),
        }),
      };
    }
  }

  export default {
    async fetch(request: Request, env: Env, ctx: ExecutionContext) {
      const logger = initLogger({
        projectName: "cloudflare-think-example", // Replace with your project name
        apiKey: env.BRAINTRUST_API_KEY,
      });

      try {
        const agent = env.THINK_AGENT.getByName("default");
        const result = await agent.runTurn({
          input: "What is the weather in Vienna?",
          mode: "wait",
        });
        return Response.json(result);
      } finally {
        ctx.waitUntil(logger.flush());
      }
    },
  };
  ```

  `wrapCloudflareThink()` accepts the module namespace (`import * as cloudflareThink from "@cloudflare/think"`) and patches the `Think` prototype in place, so all subclasses inherit tracing without modifying your agent definitions.

  Because `Think` agents are Durable Objects, your `wrangler.jsonc` needs the `nodejs_compat` compatibility flag, a Durable Object binding for your agent class, and a migration that registers it. Store `BRAINTRUST_API_KEY` as a Wrangler secret.

  ```json title="wrangler.jsonc" theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
  {
    "main": "src/cloudflare-think-manual.ts",
    "compatibility_flags": ["nodejs_compat"],
    "durable_objects": {
      "bindings": [{ "name": "THINK_AGENT", "class_name": "MyAgent" }]
    },
    "migrations": [{ "tag": "v1", "new_sqlite_classes": ["MyAgent"] }]
  }
  ```

  See the [Cloudflare setup guide](/docs/sdks/typescript/install-and-instrument#cloudflare) for the full deployment configuration.

  <h2 id="auto-instrumentation-typescript">
    Auto-instrumentation
  </h2>

  Auto-instrumentation patches the SDK at runtime without modifying your application code, but the `--import` hook only runs under Node (local development or tests), not in the Cloudflare Workers runtime. For a deployed Worker, use manual instrumentation above.

  <Steps>
    <Step title="Initialize Braintrust and define your agent">
      <CodeGroup>
        ```javascript title="cloudflare-think-auto.js" theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
        import { createOpenAI } from "@ai-sdk/openai";
        import { initLogger } from "braintrust";
        import { Think } from "@cloudflare/think";

        initLogger({
          projectName: "cloudflare-think-example", // Replace with your project name
          apiKey: process.env.BRAINTRUST_API_KEY,
        });

        export class MyAgent extends Think {
          getModel() {
            return createOpenAI({ apiKey: process.env.OPENAI_API_KEY }).chat(
              "gpt-5-mini",
            );
          }

          getSystemPrompt() {
            return "You are a helpful weather assistant.";
          }
        }
        ```
      </CodeGroup>
    </Step>

    <Step title="Run with the import hook">
      ```bash theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
      node --import braintrust/hook.mjs cloudflare-think-auto.js
      ```

      <Warning>
        The `--import` hook only patches the SDK when your code runs under Node, such as local development or tests. It does not run in the Cloudflare Workers runtime (`workerd`), so a Worker deployed with Wrangler stays uninstrumented. To trace a deployed Worker, use [manual instrumentation](#manual-instrumentation-typescript) with `wrapCloudflareThink()`.
      </Warning>

      The auto-instrumentation example uses plain JavaScript so `node --import` can run the file directly. The Braintrust APIs work the same in TypeScript projects — compile your TypeScript to JavaScript, then run the compiled file with the import hook.

      <Note>
        If you're using a bundler, see [Trace LLM calls](/docs/instrument/trace-llm-calls#auto-instrumentation) for plugin and loader setup.
      </Note>
    </Step>
  </Steps>

  <h2 id="what-traced-typescript">
    What Braintrust traces
  </h2>

  Braintrust captures:

  * Task spans for each `Think.runTurn` call, with the turn input messages as input and the final assistant response as output.
  * Model name, provider, and token usage as metadata on the task span.
  * Nested LLM spans for each AI SDK call issued during the turn, with messages, parameters, response content, and token usage.
  * Tool call spans for each tool invoked during the turn, with tool name, arguments, and result.
  * Errors captured on the span if the turn fails or the stream encounters an error.

  <h2 id="resources-typescript">
    Resources
  </h2>

  * [`@cloudflare/think` documentation](https://developers.cloudflare.com/agents/harnesses/think/)
  * [`@cloudflare/think` on npm](https://www.npmjs.com/package/@cloudflare/think)
  * [Trace LLM calls](/docs/instrument/trace-llm-calls)
</View>