How to use

Wire up an LLM agent that reads and edits .docx files headlessly. Install the SDK, open a document, and run an agentic tool loop. Full working code below.

If you need real-time sync between the agent and a frontend editor, add collaboration. The SDK client joins the same Yjs room as the frontend: edits appear live.

Prerequisites

Node.js 18+
@superdoc-dev/sdk
An LLM provider API key (e.g., OPENAI_API_KEY)

Step 1: Install

OpenAI
Anthropic
Vercel AI

npm install @superdoc-dev/sdk openai

npm install @superdoc-dev/sdk @anthropic-ai/sdk

npm install @superdoc-dev/sdk ai @ai-sdk/openai

Step 2: Open a document

Create an SDK client and open a .docx file. client.open() returns a document handle you’ll pass to the dispatcher.

import { createSuperDocClient } from '@superdoc-dev/sdk';

const client = createSuperDocClient();
await client.connect();

const doc = await client.open({ doc: './contract.docx' });

Step 3: Load the toolkit

One call returns the tool definitions for your provider, the matching system prompt, and a pre-bound dispatcher. This guide uses the core preset (recommended); everything can be cached — it doesn’t change between requests.

OpenAI
Anthropic
Vercel AI

import { createAgentToolkit } from '@superdoc-dev/sdk';

const { tools, systemPrompt, dispatch } = await createAgentToolkit({
  provider: 'openai',
  preset: 'core',
});

import { createAgentToolkit } from '@superdoc-dev/sdk';

const { tools, systemPrompt, dispatch } = await createAgentToolkit({
  provider: 'anthropic',
  preset: 'core',
});

import { createAgentToolkit } from '@superdoc-dev/sdk';

const { tools: sdkTools, systemPrompt, dispatch } = await createAgentToolkit({
  provider: 'vercel',
  preset: 'core',
});

On the legacy preset instead? Pass preset: 'legacy' (or omit it — legacy is the default for backwards compatibility). The loop below is identical; what changes is the tool surface the model sees and that results are raw operation results rather than receipts. See the legacy preset page for its behaviors.

Step 4: Run the agent loop

The agent loop sends messages to the LLM, dispatches tool calls, feeds results back, and repeats until the model is done.

OpenAI
Anthropic
Vercel AI

import OpenAI from 'openai';

const openai = new OpenAI(); // uses OPENAI_API_KEY env var

const messages: OpenAI.Chat.ChatCompletionMessageParam[] = [
  { role: 'system', content: systemPrompt },
  { role: 'user', content: 'Find the termination clause and rewrite it to allow 30-day notice.' },
];

while (true) {
  const response = await openai.chat.completions.create({
    model: 'gpt-5.4',
    messages,
    tools,
  });

  const choice = response.choices[0];
  messages.push(choice.message);

  // Stop when the model has no more tool calls
  if (choice.finish_reason === 'stop' || !choice.message.tool_calls?.length) {
    console.log(choice.message.content);
    break;
  }

  // Execute each tool call and feed results back
  for (const toolCall of choice.message.tool_calls) {
    if (toolCall.type !== 'function') continue;

    try {
      const result = await dispatch(doc,
        toolCall.function.name,
        JSON.parse(toolCall.function.arguments),
      );
      messages.push({
        role: 'tool',
        tool_call_id: toolCall.id,
        content: JSON.stringify(result),
      });
    } catch (err: any) {
      // Return errors as tool results: the model will self-correct
      messages.push({
        role: 'tool',
        tool_call_id: toolCall.id,
        content: JSON.stringify({ error: err.message }),
      });
    }
  }
}

What’s happening:

The system prompt teaches the model how to use SuperDoc tools.
The while(true) loop calls OpenAI, checks for tool calls, dispatches them via the toolkit’s dispatch, and feeds results back.
When the model returns finish_reason: 'stop' (no more tool calls), the loop ends.
Errors are caught and returned as tool results so the model can see what went wrong and retry.

import Anthropic from '@anthropic-ai/sdk';

const anthropic = new Anthropic(); // uses ANTHROPIC_API_KEY env var

const messages: Anthropic.MessageParam[] = [
  { role: 'user', content: 'Find the termination clause and rewrite it to allow 30-day notice.' },
];

while (true) {
  const response = await anthropic.messages.create({
    model: 'claude-sonnet-5',
    max_tokens: 4096,
    system: systemPrompt,
    messages,
    tools,
  });

  messages.push({ role: 'assistant', content: response.content });

  // Stop when the model has no more tool calls
  if (response.stop_reason === 'end_turn' || !response.content.some((b) => b.type === 'tool_use')) {
    const textBlock = response.content.find((b) => b.type === 'text');
    console.log(textBlock?.text);
    break;
  }

  // Execute each tool call and feed results back
  const toolResults: Anthropic.ToolResultBlockParam[] = [];
  for (const block of response.content) {
    if (block.type !== 'tool_use') continue;

    try {
      const result = await dispatch(doc,
        block.name,
        block.input as Record<string, unknown>,
      );
      toolResults.push({
        type: 'tool_result',
        tool_use_id: block.id,
        content: JSON.stringify(result),
      });
    } catch (err: any) {
      // Return errors as tool results: the model will self-correct
      toolResults.push({
        type: 'tool_result',
        tool_use_id: block.id,
        content: JSON.stringify({ error: err.message }),
        is_error: true,
      });
    }
  }

  messages.push({ role: 'user', content: toolResults });
}

What’s happening:

The system prompt is passed via the system parameter (not as a message).
The loop calls Anthropic, checks for tool_use blocks, dispatches them, and collects tool_result blocks.
Tool results are sent back as a user message with an array of tool_result blocks.
When the model returns stop_reason: 'end_turn' (no more tool calls), the loop ends.
Errors use is_error: true so the model knows the call failed.

import { generateText, jsonSchema, stepCountIs } from 'ai';
import { openai } from '@ai-sdk/openai';

// Convert SDK tool definitions into Vercel AI tool objects with execute functions
// The vercel dialect is flat: { name, description, inputSchema }.
const tools: Record<string, any> = {};
for (const t of sdkTools as any[]) {
  tools[t.name] = {
    description: t.description,
    inputSchema: jsonSchema<Record<string, unknown>>(t.inputSchema),
    execute: async (args: Record<string, unknown>) => {
      try {
        return await dispatch(doc, t.name, args);
      } catch (err: any) {
        return { error: err.message };
      }
    },
  };
}

// generateText handles the agent loop internally
const { text } = await generateText({
  model: openai.chat('gpt-5.4'),
  system: systemPrompt,
  messages: [
    { role: 'user', content: 'Find the termination clause and rewrite it to allow 30-day notice.' },
  ],
  tools,
  stopWhen: stepCountIs(10),
});

console.log(text);

What’s happening:

SDK tool definitions are converted into Vercel AI tool objects: each with an execute function that calls the toolkit’s dispatch.
generateText handles the agent loop internally: it calls the model, executes tools, feeds results back, and repeats.
stopWhen: stepCountIs(10) sets a max iteration guard.
No manual while(true) loop needed: Vercel AI manages it for you.

Step 5: Save and clean up

await doc.save({ inPlace: true });
await doc.close();
await client.dispose();

Full example

A complete, copy-pasteable script that opens a document, runs an agent, saves, and exits:

OpenAI
Anthropic
Vercel AI

import OpenAI from 'openai';
import { createSuperDocClient, createAgentToolkit } from '@superdoc-dev/sdk';

// 1. Open the document
const client = createSuperDocClient();
await client.connect();
const doc = await client.open({ doc: './contract.docx' });

// 2. Load the toolkit (tools + system prompt + dispatch, core preset)
const { tools, systemPrompt, dispatch } = await createAgentToolkit({
  provider: 'openai',
  preset: 'core',
});

// 3. Build the conversation
const openai = new OpenAI();
const messages: OpenAI.Chat.ChatCompletionMessageParam[] = [
  { role: 'system', content: systemPrompt },
  { role: 'user', content: 'Find the termination clause and rewrite it to allow 30-day notice.' },
];

// 4. Agent loop
while (true) {
  const response = await openai.chat.completions.create({
    model: 'gpt-5.4',
    messages,
    tools,
  });

  const choice = response.choices[0];
  messages.push(choice.message);

  if (choice.finish_reason === 'stop' || !choice.message.tool_calls?.length) {
    console.log(choice.message.content);
    break;
  }

  for (const toolCall of choice.message.tool_calls) {
    if (toolCall.type !== 'function') continue;

    try {
      const result = await dispatch(doc,
        toolCall.function.name,
        JSON.parse(toolCall.function.arguments),
      );
      messages.push({
        role: 'tool',
        tool_call_id: toolCall.id,
        content: JSON.stringify(result),
      });
    } catch (err: any) {
      messages.push({
        role: 'tool',
        tool_call_id: toolCall.id,
        content: JSON.stringify({ error: err.message }),
      });
    }
  }
}

// 5. Save and clean up
await doc.save({ inPlace: true });
await doc.close();
await client.dispose();

import Anthropic from '@anthropic-ai/sdk';
import { createSuperDocClient, createAgentToolkit } from '@superdoc-dev/sdk';

// 1. Open the document
const client = createSuperDocClient();
await client.connect();
const doc = await client.open({ doc: './contract.docx' });

// 2. Load the toolkit (tools + system prompt + dispatch, core preset)
const { tools, systemPrompt, dispatch } = await createAgentToolkit({
  provider: 'anthropic',
  preset: 'core',
});

// 3. Build the conversation
const anthropic = new Anthropic();
const messages: Anthropic.MessageParam[] = [
  { role: 'user', content: 'Find the termination clause and rewrite it to allow 30-day notice.' },
];

// 4. Agent loop
while (true) {
  const response = await anthropic.messages.create({
    model: 'claude-sonnet-5',
    max_tokens: 4096,
    system: systemPrompt,
    messages,
    tools,
  });

  messages.push({ role: 'assistant', content: response.content });

  if (response.stop_reason === 'end_turn' || !response.content.some((b) => b.type === 'tool_use')) {
    const textBlock = response.content.find((b) => b.type === 'text');
    console.log(textBlock?.text);
    break;
  }

  const toolResults: Anthropic.ToolResultBlockParam[] = [];
  for (const block of response.content) {
    if (block.type !== 'tool_use') continue;

    try {
      const result = await dispatch(doc,
        block.name,
        block.input as Record<string, unknown>,
      );
      toolResults.push({
        type: 'tool_result',
        tool_use_id: block.id,
        content: JSON.stringify(result),
      });
    } catch (err: any) {
      toolResults.push({
        type: 'tool_result',
        tool_use_id: block.id,
        content: JSON.stringify({ error: err.message }),
        is_error: true,
      });
    }
  }

  messages.push({ role: 'user', content: toolResults });
}

// 5. Save and clean up
await doc.save({ inPlace: true });
await doc.close();
await client.dispose();

import { generateText, jsonSchema, stepCountIs } from 'ai';
import { openai } from '@ai-sdk/openai';
import { createSuperDocClient, createAgentToolkit } from '@superdoc-dev/sdk';

// 1. Open the document
const client = createSuperDocClient();
await client.connect();
const doc = await client.open({ doc: './contract.docx' });

// 2. Load tools and system prompt
const { tools: sdkTools, systemPrompt, dispatch } = await createAgentToolkit({
  provider: 'vercel',
  preset: 'core',
});

// 3. Convert SDK tools into Vercel AI tool objects
// The vercel dialect is flat: { name, description, inputSchema }.
const tools: Record<string, any> = {};
for (const t of sdkTools as any[]) {
  tools[t.name] = {
    description: t.description,
    inputSchema: jsonSchema<Record<string, unknown>>(t.inputSchema),
    execute: async (args: Record<string, unknown>) => {
      try {
        return await dispatch(doc, t.name, args);
      } catch (err: any) {
        return { error: err.message };
      }
    },
  };
}

// 4. Run the agent (loop handled by generateText)
const { text } = await generateText({
  model: openai.chat('gpt-5.4'),
  system: systemPrompt,
  messages: [
    { role: 'user', content: 'Find the termination clause and rewrite it to allow 30-day notice.' },
  ],
  tools,
  stopWhen: stepCountIs(10),
});

console.log(text);

// 5. Save and clean up
await doc.save({ inPlace: true });
await doc.close();
await client.dispose();

Other providers

AWS Bedrock

Use the toolkit with provider: 'anthropic' and convert to Bedrock’s toolSpec shape:

Node.js
Python

import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';
import { createSuperDocClient, createAgentToolkit } from '@superdoc-dev/sdk';

const client = createSuperDocClient();
await client.connect();
const doc = await client.open({ doc: './contract.docx' });

// Get tools in Anthropic format, convert to Bedrock toolSpec shape
const { tools, systemPrompt, dispatch } = await createAgentToolkit({
  provider: 'anthropic',
  preset: 'core',
});
const toolConfig = {
  tools: tools.map((t: any) => ({
    toolSpec: {
      name: t.name,
      description: t.description,
      inputSchema: { json: t.input_schema },
    },
  })),
};

const bedrock = new BedrockRuntimeClient({ region: 'us-east-1' });
const messages = [
  { role: 'user', content: [{ text: 'Review this contract.' }] },
];

while (true) {
  const res = await bedrock.send(new ConverseCommand({
    // Bedrock uses regional inference-profile IDs, not Anthropic model
    // names — copy the exact ID from your Bedrock console's model catalog.
    modelId: 'us.anthropic.claude-sonnet-4-6',
    messages,
    system: [{ text: systemPrompt }],
    toolConfig,
  }));

  const output = res.output?.message;
  if (!output) break;
  messages.push(output);

  const toolUses = output.content?.filter((b) => b.toolUse) ?? [];
  if (!toolUses.length) break;

  const results = [];
  for (const block of toolUses) {
    const { name, input, toolUseId } = block.toolUse;
    const result = await dispatch(doc, name, input ?? {});
    const json = typeof result === 'object' && result !== null ? result : { result };
    results.push({ toolResult: { toolUseId, content: [{ json }] } });
  }
  messages.push({ role: 'user', content: results });
}

await doc.save();
await doc.close();
await client.dispose();

import boto3
from superdoc import SuperDocClient, create_agent_toolkit

client = SuperDocClient()
client.connect()
doc = client.open({"doc": "./contract.docx"})

# Get tools in Anthropic format, convert to Bedrock toolSpec shape
kit = create_agent_toolkit({"provider": "anthropic", "preset": "core"})
tool_config = {
    "tools": [
        {
            "toolSpec": {
                "name": t["name"],
                "description": t["description"],
                "inputSchema": {"json": t.get("input_schema", {})},
            }
        }
        for t in kit["tools"]
    ]
}

bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
messages = [{"role": "user", "content": [{"text": "Review this contract."}]}]

while True:
    response = bedrock.converse(
        # Bedrock uses regional inference-profile IDs, not Anthropic model
        # names — copy the exact ID from your Bedrock console's model catalog.
        modelId="us.anthropic.claude-sonnet-4-6",
        messages=messages,
        system=[{"text": kit["system_prompt"]}],
        toolConfig=tool_config,
    )

    output = response["output"]["message"]
    messages.append(output)

    tool_uses = [b for b in output.get("content", []) if "toolUse" in b]
    if not tool_uses:
        break

    tool_results = []
    for block in tool_uses:
        tu = block["toolUse"]
        result = kit["dispatch"](doc, tu["name"], tu.get("input", {}))
        json_result = result if isinstance(result, dict) else {"result": result}
        tool_results.append(
            {"toolResult": {"toolUseId": tu["toolUseId"], "content": [{"json": json_result}]}}
        )
    messages.append({"role": "user", "content": tool_results})

doc.save({})
doc.close({})
client.dispose()

Auth: AWS credentials via aws configure, env vars, or IAM role. No API key needed.

Streaming generated text into a visible editor

Sometimes you don’t need a full agent loop. You just want the model to write into the document while the user watches. Stream the output through a small backend proxy and append each delta to the editor:

for await (const chunk of streamFromServer(prompt, signal)) {
  buffer += chunk;
  if (chunk.includes('\n')) flush();
  else if (!pendingFlush) pendingFlush = setTimeout(flush, 150);
}

function flush() {
  editor.doc.insert({ value: buffer, type: 'text' });
  buffer = '';
}

editor.doc.insert is the public Document API. With no target, content appends at the end. Newlines from the model become real paragraph breaks. A few things to get right:

Keep the model key on the server. A small Node proxy that forwards Server-Sent Events keeps the key out of client bundles.
Buffer deltas. Inserting on every token causes one document mutation per token, which floods the layout engine and undo stack. Flush on a timer (~150ms) or whenever a newline arrives.
Abort on unmount and Stop. Tie an AbortController to the fetch and call it from your cleanup. The server should also abort upstream when the client disconnects so neither side burns tokens.

Full working example: examples/ai/streaming.

LLM tools: tool catalog and SDK functions
Best practices: prompting, workflow tips, and tested prompt examples
Debugging: troubleshoot tool call failures
Collaboration: add real-time sync between agent and frontend
SDKs: typed Node.js and Python wrappers

Getting Started

Editor

Document Engine

AI

Solutions

Resources

Prerequisites

Step 1: Install

Step 2: Open a document

Step 3: Load the toolkit

Step 4: Run the agent loop

Step 5: Save and clean up

Full example

Other providers

AWS Bedrock

Streaming generated text into a visible editor

​Prerequisites

​Step 1: Install

​Step 2: Open a document

​Step 3: Load the toolkit

​Step 4: Run the agent loop

​Step 5: Save and clean up

​Full example

​Other providers

​AWS Bedrock

​Streaming generated text into a visible editor

​Related

Prerequisites

Step 1: Install

Step 2: Open a document

Step 3: Load the toolkit

Step 4: Run the agent loop

Step 5: Save and clean up

Full example

Other providers

AWS Bedrock

Streaming generated text into a visible editor

Related