Tools

Think provides built-in workspace file tools on every turn, plus integration points for custom tools, code execution, and dynamic extensions.

Tool merge order

On every turn, Think merges tools from multiple sources. Later sources override earlier ones if names collide:

1. Workspace tools — read, write, edit, list, find, grep, delete, bash (built-in)
2. getTools() — your custom server-side tools
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(), refer to Agent Skills)
6. MCP tools — from connected MCP servers
7. Client tools — from the browser (refer to Client tools)

Tools belong to the agent running the turn. For parent-child orchestration, use Agents as tools instead of passing one-off tools through chat().

Built-in workspace tools

Every Think agent gets this.workspace — a virtual filesystem backed by Durable Object SQLite. Workspace tools are automatically available to the model with no configuration.

Tool	Description
read	Read text with line numbers; pass images and PDFs to multimodal models
write	Write content to a file (creates parent directories)
edit	Apply a find-and-replace edit to an existing file (supports fuzzy matching)
list	List files and directories in a path
find	Find files matching a glob pattern
grep	Search file contents by regex or fixed string
delete	Delete a file or directory
bash	Run a sandboxed Bash script against workspace files

The bash tool is enabled by default. It mounts workspace files into a just-bash virtual filesystem, runs with network access disabled, and writes created, updated, and deleted files and empty directories back to the workspace. Use it for shell-style workflows that combine multiple file operations; use the narrower tools for simple reads, writes, and edits.

To keep tool calls bounded, the Bash tool snapshots up to 1,000 workspace files by default and skips files larger than 1 MB. Skipped files are reported in the tool result and are treated as protected during write-back so the script cannot accidentally overwrite or delete content that was not mounted. You can tune maxWorkspaceFiles, maxWorkspaceFileBytes, maxOutputBytes, timeout, and network through workspaceBash.

Disable the default Bash tool for conservative deployments:

JavaScript

JavaScript

export class MyAgent extends Think {
  workspaceBash = false;

  getModel() {
    /* ... */
  }
}

TypeScript

TypeScript

export class MyAgent extends Think<Env> {
  workspaceBash = false;

  getModel() {
    /* ... */
  }
}

R2 spillover

By default, the workspace stores everything in SQLite. For large files, override workspace to add R2 spillover:

JavaScript

JavaScript

import { Think } from "@cloudflare/think";
import { Workspace } from "@cloudflare/shell";

export class MyAgent extends Think {
  workspace = new Workspace({
    sql: this.ctx.storage.sql,
    r2: this.env.R2,
    name: () => this.name,
  });

  getModel() {
    /* ... */
  }
}

TypeScript

TypeScript

import { Think } from "@cloudflare/think";
import { Workspace } from "@cloudflare/shell";

export class MyAgent extends Think<Env> {
  override workspace = new Workspace({
    sql: this.ctx.storage.sql,
    r2: this.env.R2,
    name: () => this.name,
  });

  getModel() {
    /* ... */
  }
}

This requires an R2 bucket binding:

wrangler.jsonc

JSONC

{
  "$schema": "./node_modules/wrangler/config-schema.json",
  "r2_buckets": [
    {
      "binding": "R2",
      "bucket_name": "agent-files"
    }
  ]
}

wrangler.toml

TOML

[[r2_buckets]]
binding = "R2"
bucket_name = "agent-files"

Custom tools

Override getTools() to add your own tools. These are standard AI SDK tool() definitions with Zod schemas:

JavaScript

JavaScript

import { Think } from "@cloudflare/think";
import { tool } from "ai";

import { z } from "zod";

export class MyAgent extends Think {
  getModel() {
    /* ... */
  }

  getTools() {
    return {
      getWeather: tool({
        description: "Get the current weather for a city",
        inputSchema: z.object({
          city: z.string().describe("City name"),
        }),
        execute: async ({ city }) => {
          const res = await fetch(
            `https://api.weather.com/v1/current?q=${city}&key=${this.env.WEATHER_KEY}`,
          );
          return res.json();
        },
      }),
    };
  }
}

TypeScript

TypeScript

import { Think } from "@cloudflare/think";
import { tool } from "ai";
import type { ToolSet } from "ai";
import { z } from "zod";

export class MyAgent extends Think<Env> {
  getModel() {
    /* ... */
  }

  getTools(): ToolSet {
    return {
      getWeather: tool({
        description: "Get the current weather for a city",
        inputSchema: z.object({
          city: z.string().describe("City name"),
        }),
        execute: async ({ city }) => {
          const res = await fetch(
            `https://api.weather.com/v1/current?q=${city}&key=${this.env.WEATHER_KEY}`,
          );
          return res.json();
        },
      }),
    };
  }
}

Custom tools are merged with workspace tools automatically. If a custom tool has the same name as a workspace tool, the custom tool wins.

Tool approval

Tools can require user approval before execution using the needsApproval option:

TypeScript

getTools(): ToolSet {
  return {
    deleteFile: tool({
      description: "Delete a file from the system",
      inputSchema: z.object({ path: z.string() }),
      needsApproval: async ({ path }) => path.startsWith("/important/"),
      execute: async ({ path }) => {
        await this.workspace.rm(path);
        return { deleted: path };
      },
    }),
  };
}

When needsApproval returns true, the tool call is sent to the client for approval. The conversation pauses until the client responds with CF_AGENT_TOOL_APPROVAL.

Note

Inside the code execution tool's sandbox, needsApproval behaves differently: it maps to the codemode runtime's durable pause/approve/resume flow, and a function-valued needsApproval always requires approval. Refer to Approvals (human-in-the-loop).

Per-turn tool overrides

The beforeTurn hook can restrict or add tools for a specific turn:

TypeScript

beforeTurn(ctx: TurnContext) {
  return {
    activeTools: ["read", "write", "getWeather"],
    tools: { emergencyTool: this.createEmergencyTool() },
  };
}

activeTools limits which tools the model can call. tools adds extra tools for this turn only (merged on top of existing tools).

MCP tools

Think inherits MCP client support from the Agent base class. MCP tools from connected servers are automatically merged into every turn.

Set waitForMcpConnections to ensure MCP servers are connected before inference runs:

JavaScript

JavaScript

export class MyAgent extends Think {
  waitForMcpConnections = true; // default 10s timeout
  // or: waitForMcpConnections = { timeout: 5000 };

  getModel() {
    /* ... */
  }
}

TypeScript

TypeScript

export class MyAgent extends Think<Env> {
  waitForMcpConnections = true; // default 10s timeout
  // or: waitForMcpConnections = { timeout: 5000 };

  getModel() {
    /* ... */
  }
}

Add MCP servers programmatically or via @callable methods:

JavaScript

JavaScript

import { callable } from "agents";

export class MyAgent extends Think {
  getModel() {
    /* ... */
  }

  @callable()
  async addServer(name, url) {
    return await this.addMcpServer(name, url);
  }

  @callable()
  async removeServer(serverId) {
    await this.removeMcpServer(serverId);
  }
}

TypeScript

TypeScript

import { callable } from "agents";

export class MyAgent extends Think<Env> {
  getModel() {
    /* ... */
  }

  @callable()
  async addServer(name: string, url: string) {
    return await this.addMcpServer(name, url);
  }

  @callable()
  async removeServer(serverId: string) {
    await this.removeMcpServer(serverId);
  }
}

Code execution tool

Let the LLM write and run JavaScript in a sandboxed Worker, recorded on a durable codemode runtime (abort-and-replay, human approvals, audit trail, reusable snippets). Requires @cloudflare/codemode and a worker_loaders binding.

npm install @cloudflare/codemode

The one-liner infers everything from the agent — state.* from this.workspace, the executor from env.LOADER, and a live browser (cdp.*) from env.BROWSER if bound:

JavaScript

JavaScript

import { Think } from "@cloudflare/think";
import { createExecuteTool } from "@cloudflare/think/tools/execute";

export class MyAgent extends Think {
  getModel() {
    /* ... */
  }

  getTools() {
    return {
      execute: createExecuteTool(this),
    };
  }
}

TypeScript

TypeScript

import { Think } from "@cloudflare/think";
import { createExecuteTool } from "@cloudflare/think/tools/execute";

export class MyAgent extends Think<Env> {
  getModel() {
    /* ... */
  }

  getTools() {
    return {
      execute: createExecuteTool(this),
    };
  }
}

Setup checklist:

wrangler.jsonc

JSONC

{
  "$schema": "./node_modules/wrangler/config-schema.json",
  "worker_loaders": [
    {
      "binding": "LOADER"
    }
  ],
  "browser": {
    "binding": "BROWSER"
  }
}

wrangler.toml

TOML

[[worker_loaders]]
binding = "LOADER"

[browser]
binding = "BROWSER" # optional — enables cdp.*

JavaScript

JavaScript

// worker entry — the runtime lives in a Durable Object facet, so the class
// must be exported (the @cloudflare/codemode/vite plugin does this
// automatically; the Think framework's generated entry already includes it)
export { CodemodeRuntime } from "@cloudflare/codemode";

TypeScript

TypeScript

// worker entry — the runtime lives in a Durable Object facet, so the class
// must be exported (the @cloudflare/codemode/vite plugin does this
// automatically; the Think framework's generated entry already includes it)
export { CodemodeRuntime } from "@cloudflare/codemode";

Each missing piece fails with an error naming the step.

Inside the sandbox the model sees typed namespaces plus the platform SDK:

• tools.* — your AI SDK tools (object args, validated against their schemas). Only tools with an execute function are exposed — client-side tools cannot run in the sandbox.
• state.* — the workspace filesystem (state.readFile({ path }), state.glob({ pattern }), state.planEdits(...), and so on).
• cdp.* — the browser, when a Browser Run binding is configured. The execute tool defaults to session: { mode: "dynamic" }: sessions are per-execution unless the model promotes one with cdp.startSession().
• codemode.search / codemode.describe / codemode.step / codemode.run — discovery, side-effect boundaries, and saved snippets.

Pass overrides for anything beyond the defaults — for example, custom tools.* alongside the agent-derived state:

JavaScript

JavaScript

execute: createExecuteTool(this, { tools: myDomainTools });

TypeScript

TypeScript

execute: createExecuteTool(this, { tools: myDomainTools });

Or fully explicit options (no agent inference):

JavaScript

JavaScript

import { createWorkspaceStateBackend } from "@cloudflare/shell";

createExecuteTool({
  ctx: this.ctx,
  tools: myDomainTools,
  state: createWorkspaceStateBackend(this.workspace),
  browser: this.env.BROWSER,
  loader: this.env.LOADER,
});

TypeScript

TypeScript

import { createWorkspaceStateBackend } from "@cloudflare/shell";

createExecuteTool({
  ctx: this.ctx,
  tools: myDomainTools,
  state: createWorkspaceStateBackend(this.workspace),
  browser: this.env.BROWSER,
  loader: this.env.LOADER,
});

Approvals (human-in-the-loop)

An AI SDK tool with needsApproval does not run immediately inside the sandbox — calling it pauses the run durably. The pause comes back as a normal tool output ({ status: "paused", executionId, pending }), the model tells the user what it needs, and the turn ends. This differs from the client-side approval flow for plain getTools() tools: inside the sandbox a function-valued needsApproval cannot be evaluated against the call's arguments ahead of time, so it conservatively always requires approval. Think ships built-in callables to resolve it:

• approveExecution(executionId) — resumes the run where it stopped. Already-done work is replayed, not re-executed. The outcome replaces the paused output in the transcript and the chat auto-continues.
• rejectExecution(executionId, reason?) — ends the run with { status: "rejected", reason } so the model can adapt.
• pendingExecutions() — pending actions (with full args) for rendering approval UI.

Note

Render approval cards from pendingExecutions(), not the transcript. The pending array in the paused tool output is a truncated preview — args are bounded (~2 KB each) so they do not blow up model context, but the full args (up to 1 MB) are what actually execute on approve. A human approving a gated call must see the authoritative args, so fetch them via pendingExecutions(executionId) before enabling the Approve button.

For a working approval card, refer to the assistant example ↗.

The runtime handle

createExecuteRuntime returns the moving parts when the host needs more than the tool — and the handle is also assigned to this.codemode when created from an agent:

JavaScript

JavaScript

import { createExecuteRuntime } from "@cloudflare/think/tools/execute";

const { runtime, connectors, tool } = createExecuteRuntime(this);
await runtime.executions(); // audit trail
await runtime.expirePaused(); // reclaim stale never-approved pauses (call from a scheduled task)
await runtime.saveSnippet("name", { executionId }); // promote a script for reuse

TypeScript

TypeScript

import { createExecuteRuntime } from "@cloudflare/think/tools/execute";

const { runtime, connectors, tool } = createExecuteRuntime(this);
await runtime.executions(); // audit trail
await runtime.expirePaused(); // reclaim stale never-approved pauses (call from a scheduled task)
await runtime.saveSnippet("name", { executionId }); // promote a script for reuse

Browser tools

Give your agent access to the Chrome DevTools Protocol (CDP) for web page inspection, scraping, screenshots, and debugging. Requires @cloudflare/codemode and a Browser Run binding.

JavaScript

JavaScript

import { Think } from "@cloudflare/think";
import { createBrowserTools } from "@cloudflare/think/tools/browser";

export class MyAgent extends Think {
  getModel() {
    /* ... */
  }

  getTools() {
    return {
      ...createBrowserTools({
        ctx: this.ctx,
        browser: this.env.BROWSER,
        loader: this.env.LOADER,
      }),
    };
  }
}

TypeScript

TypeScript

import { Think } from "@cloudflare/think";
import { createBrowserTools } from "@cloudflare/think/tools/browser";

export class MyAgent extends Think<Env> {
  getModel() {
    /* ... */
  }

  getTools() {
    return {
      ...createBrowserTools({
        ctx: this.ctx,
        browser: this.env.BROWSER,
        loader: this.env.LOADER,
      }),
    };
  }
}

wrangler.jsonc

JSONC

{
  "$schema": "./node_modules/wrangler/config-schema.json",
  "browser": {
    "binding": "BROWSER"
  },
  "worker_loaders": [
    {
      "binding": "LOADER"
    }
  ]
}

wrangler.toml

TOML

[browser]
binding = "BROWSER"

[[worker_loaders]]
binding = "LOADER"

This adds the durable CDP tool plus stateless Quick Action tools when a browser binding is present:

Tool	Description
browser_execute	Run JavaScript against a live browser over CDP (screenshots, DOM reads, JS evaluation).
browser_markdown	Read a page or raw HTML as Markdown.
browser_extract	Extract structured data from a page with AI.
browser_links	List links on a page.
browser_scrape	Scrape specific elements by CSS selector.

Pass quickActions: false to keep only browser_execute, or pass quickActions: { actions, maxChars, options } to configure the stateless tools. The Quick Action tools share the browser binding, need no Worker Loader, and resolve ctx from the current Agent automatically. To use only the stateless tools, import createQuickActionTools from @cloudflare/think/tools/browser.

The tool is backed by a codemode runtime with the cdp connector: the model writes async arrow functions that run in a sandboxed Worker isolate, with cdp.send(), cdp.attachToTarget(), cdp.spec() (the live, normalized protocol description), session helpers (cdp.startSession(), cdp.sessionInfo(), cdp.closeSession()), and debug-log helpers. Executions are recorded for abort-and-replay, so browser sessions survive approval pauses.

By default each execution gets a fresh browser session (one-shot), torn down when the run ends. Pass session: { mode: "dynamic" } to let the model promote a session with cdp.startSession() so later executions continue in the same browser, or session: { mode: "reuse", key } for a named long-lived session. Stale sessions are reclaimed by the connector's sweep() — call it from a scheduled task.

Note

The simplest setup is the unified execute tool in Code execution tool: createExecuteTool(this) already includes cdp.* alongside state.* and tools.* when env.BROWSER is bound — one tool, one durable history. Use createBrowserTools when you want a separate, browser-only tool.

For a custom Chrome endpoint, pass cdpUrl instead of browser:

JavaScript

JavaScript

createBrowserTools({
  ctx: this.ctx,
  cdpUrl: "http://localhost:9222",
  loader: this.env.LOADER,
});

TypeScript

TypeScript

createBrowserTools({
  ctx: this.ctx,
  cdpUrl: "http://localhost:9222",
  loader: this.env.LOADER,
});

For the full CDP connector API, refer to Browse the web.

Extensions

Extensions are dynamically loaded sandboxed Workers that add tools at runtime. The LLM can write extension source code, load it, and use the new tools on the next turn.

Extensions require a worker_loaders binding:

JavaScript

JavaScript

import { Think } from "@cloudflare/think";

export class MyAgent extends Think {
  extensionLoader = this.env.LOADER;

  getModel() {
    /* ... */
  }
}

TypeScript

TypeScript

import { Think } from "@cloudflare/think";

export class MyAgent extends Think<Env> {
  extensionLoader = this.env.LOADER;

  getModel() {
    /* ... */
  }
}

Static extensions

Define extensions that load at startup:

JavaScript

JavaScript

export class MyAgent extends Think {
  extensionLoader = this.env.LOADER;

  getModel() {
    /* ... */
  }

  getExtensions() {
    return [
      {
        manifest: {
          name: "math",
          version: "1.0.0",
          permissions: { network: false },
        },
        source: `({
          tools: {
            add: {
              description: "Add two numbers",
              parameters: { a: { type: "number" }, b: { type: "number" } },
              execute: async ({ a, b }) => ({ result: a + b })
            }
          }
        })`,
      },
    ];
  }
}

TypeScript

TypeScript

export class MyAgent extends Think<Env> {
  extensionLoader = this.env.LOADER;

  getModel() {
    /* ... */
  }

  getExtensions() {
    return [
      {
        manifest: {
          name: "math",
          version: "1.0.0",
          permissions: { network: false },
        },
        source: `({
          tools: {
            add: {
              description: "Add two numbers",
              parameters: { a: { type: "number" }, b: { type: "number" } },
              execute: async ({ a, b }) => ({ result: a + b })
            }
          }
        })`,
      },
    ];
  }
}

Extension tools are namespaced — a math extension with an add tool becomes math_add in the model's tool set.

LLM-driven extensions

Give the model createExtensionTools so it can load extensions dynamically:

JavaScript

JavaScript

import { createExtensionTools } from "@cloudflare/think/tools/extensions";

export class MyAgent extends Think {
  extensionLoader = this.env.LOADER;

  getModel() {
    /* ... */
  }

  getTools() {
    return {
      ...createExtensionTools({ manager: this.extensionManager }),
      ...this.extensionManager.getTools(),
    };
  }
}

TypeScript

TypeScript

import { createExtensionTools } from "@cloudflare/think/tools/extensions";

export class MyAgent extends Think<Env> {
  extensionLoader = this.env.LOADER;

  getModel() {
    /* ... */
  }

  getTools() {
    return {
      ...createExtensionTools({ manager: this.extensionManager! }),
      ...this.extensionManager!.getTools(),
    };
  }
}

This gives the model two tools:

• load_extension — load a new extension from JavaScript source
• list_extensions — list currently loaded extensions

Extension context blocks

Extensions can declare context blocks in their manifest. These are automatically registered with the Session:

TypeScript

getExtensions() {
  return [{
    manifest: {
      name: "notes",
      version: "1.0.0",
      permissions: { network: false },
      context: [
        { label: "scratchpad", description: "Extension scratch space", maxTokens: 500 },
      ],
    },
    source: `({ tools: { /* ... */ } })`,
  }];
}

The context block is registered as notes_scratchpad (namespaced by extension name).

Custom workspace backends

The individual tool factories are exported for use with custom storage backends:

JavaScript

JavaScript

import {
  createReadTool,
  createWriteTool,
  createEditTool,
  createListTool,
  createFindTool,
  createGrepTool,
  createDeleteTool,
  createWorkspaceTools,
} from "@cloudflare/think/tools/workspace";

TypeScript

TypeScript

import {
  createReadTool,
  createWriteTool,
  createEditTool,
  createListTool,
  createFindTool,
  createGrepTool,
  createDeleteTool,
  createWorkspaceTools,
} from "@cloudflare/think/tools/workspace";

Implement the operations interface for your storage backend:

JavaScript

JavaScript

const myReadOps = {
  readFile: async (path) => fetchFromMyStorage(path),
  stat: async (path) => getFileInfo(path),
};

const readTool = createReadTool({ ops: myReadOps });

TypeScript

TypeScript

import type { ReadOperations } from "@cloudflare/think/tools/workspace";

const myReadOps: ReadOperations = {
  readFile: async (path) => fetchFromMyStorage(path),
  stat: async (path) => getFileInfo(path),
};

const readTool = createReadTool({ ops: myReadOps });

Or create the full set from a Workspace, optionally disabling the Bash tool:

JavaScript

JavaScript

import { createWorkspaceTools } from "@cloudflare/think/tools/workspace";

const tools = createWorkspaceTools(myCustomWorkspace);
const toolsWithoutBash = createWorkspaceTools(myCustomWorkspace, {
  bash: false,
});

TypeScript

TypeScript

import { createWorkspaceTools } from "@cloudflare/think/tools/workspace";

const tools = createWorkspaceTools(myCustomWorkspace);
const toolsWithoutBash = createWorkspaceTools(myCustomWorkspace, {
  bash: false,
});