Create an EditorBridge backed by a DocxReviewer. The agent (or MCP client) can read, comment, propose changes, etc., against a parsed DOCX file on disk. Call `reviewer.toBuffer()` afterwards to get the modified DOCX.

declare function createReviewerBridge(reviewer: DocxReviewer): EditorBridge;

Execute a tool call against an EditorBridge. Returns the result (never throws).

declare function executeToolCall(toolName: string, input: Record<string, unknown>, bridge: EditorBridge): AgentToolResult;

Friendly UI label for a tool — sourced from the registry's `displayName`, falling back to a sentence-case version of the snake_case name. Used by `<AgentTimeline>` and any other UI that lists running / completed tools.

declare function getToolDisplayName(name: string): string;

getToolDisplayName('add_comment') // → 'Adding comment'

getToolDisplayName('fetch_clause_template') // → 'Fetch clause template'

Get tool schemas in OpenAI function-calling format. Works directly with the OpenAI SDK and Anthropic's tools API. For Vercel AI SDK, LangChain, or other agent runtimes, transform this output to that runtime's required shape — see `examples/agent-chat-demo/` for a Vercel AI SDK example. The package stays runtime-agnostic.

declare function getToolSchemas(): {
    type: "function";
    function: {
        name: string;
        description: string;
        parameters: Record<string, unknown>;
    };
}[];

Headless DOCX reviewer — parse a file, read/comment/track changes against the document model, write the modified DOCX back out. No DOM, no editor instance. Pair with `createReviewerBridge()` to drive the built-in agent tools against a file on disk.

declare class DocxReviewer

```ts
const reviewer = await DocxReviewer.fromBuffer(buffer, 'AI Reviewer');
reviewer.addComment(5, 'Fix this paragraph.');
reviewer.replace(5, '$50k', '$500k');
const output = await reviewer.toBuffer();
```

Snapshot of what the user is looking at — pass this to your agent's system prompt so it knows the current selection / page without an extra `read_selection` round-trip.

interface AgentContextSnapshot

Definition of an agent tool — name, JSON-schema input, handler. Use this to build custom tools alongside the built-in `agentTools`.

interface AgentToolDefinition<TInput = Record<string, unknown>>

Result returned by a tool handler. `success: false` carries an `error` message; `success: true` may carry tool-specific `data`.

interface AgentToolResult

High-level agent surface over a live editor (or a headless reviewer). Every built-in tool calls into this contract — implement it once and the agent toolkit works against your editor.

interface EditorBridge

interface SelectionInfo

All built-in agent tools — read/write document content, comments, and tracked changes. Use `getToolSchemas()` to feed them to an LLM and `executeToolCall()` to run the handlers against an `EditorBridge`.

agentTools: AgentToolDefinition<any>[]