# AgentDoc

> AgentDoc is a public agent-first document editor. Every operation in
> the editor is exposed as a typed Model Context Protocol (MCP) tool so
> autonomous LLM agents can read, write, format, navigate, and export
> documents without a human in the loop. The same surface is also
> usable by humans through chat or voice. Free, browser-based, MIT-style
> research project hosted in Germany.

## Primary entry points for agents

- [Agent self-registration](https://agent-doc-edit.com/api/agents/register): POST {"name": "..."} returns a freshly provisioned isolated user account + initial API key in one HTTP call. Rate limit: 5/hour per IP. Agent accounts have their own document namespace; no collision with other agents or human users.
- [MCP endpoint](https://agent-doc-edit.com/mcp/sse): public Model Context Protocol server over SSE. Bearer-token auth (`ak_...` API keys). 35 typed tools (read, write, format, paginate, export). Agents bring their own LLM – we do not bill or proxy reasoning tokens.
- [DOCX import](https://agent-doc-edit.com/api/docs/import/docx): POST multipart `file=<word.docx>` (`Authorization: Bearer <ak_...>`) creates a new Document on the agent's account from an existing Word file and switches it to active. Preserves headings, formatting, page breaks, hyperlinks, headers/footers. Use this to seed agent workflows that start from a corporate template or pre-authored draft instead of building from scratch.
- [Agent onboarding page](https://agent-doc-edit.com/agents): full HTTP examples, two auth flows (self-register vs. own account), tool catalogue, honest constraints, DOCX import endpoint.

## Documentation

- [Developer architecture and design rationale](https://agent-doc-edit.com/developers): MCP tool design patterns, ReAct FSM gating, twenty workflow configurations, FastMCP code example, common reliability questions ("why does my LLM agent hallucinate tool calls", "how do I prevent index drift").
- [Research and benchmarks](https://agent-doc-edit.com/research): academic thesis evaluating 20 workflow configurations across 13 scenarios. Levenshtein-based scoring, tool-call accuracy, token consumption, voice agent comparison.

## Audience-specific landing pages (for human + agent context)

- [Mobile users – write documents on a phone by voice](https://agent-doc-edit.com/mobile)
- [Accessibility – keyboard-free writing for motor disability, ALS, RSI](https://agent-doc-edit.com/accessibility)
- [Main landing page](https://agent-doc-edit.com/)

## Blog – patches, updates, engineering notes

- [Blog index](https://agent-doc-edit.com/blog/): release notes and engineering write-ups, ~one post every 2–3 weeks.
- [DOCX Import — Round-Tripping Word Documents](https://agent-doc-edit.com/blog/docx-import-roundtrip): how AgentDoc imports .docx files and preserves page breaks, hyperlinks, headers/footers, fonts, colours and line spacing across a full Word → editor → Word round-trip. Companion to the export rewrite.
- [Rebuilding PDF + DOCX Export](https://agent-doc-edit.com/blog/pdf-docx-export-rewrite): how AgentDoc rebuilt PDF and DOCX export around the frontend's pagination map, using WeasyPrint, a Pandoc reference.docx, and a Lua filter that translates stand-off classes into native Word run properties.
- [Patch Notes – April 2026](https://agent-doc-edit.com/blog/patch-notes-april-2026): renderer fixes, decoration-class composition, TOGGLE semantics on apply_* tools, inline-code escaping, decoration styling generalisation.
- [Tool Granularity in LLM Agents](https://agent-doc-edit.com/blog/tool-granularity-llm-agents): findings from 20 workflow configurations × 13 scenarios. Atomic vs. composite tools, ReAct FSM gating, index drift, accuracy/token tradeoff.
- [Voice-First Document Editing with Gemini Live](https://agent-doc-edit.com/blog/voice-first-document-editing): why we wired Gemini Live native audio directly to the MCP server instead of using a transcription proxy.

## Brand / alternate names

AgentDoc is the proper noun for an AI document editor; it is also referred to as "agent doc", "agentdocs", "agent docs", "doc edit", "docedit", "AI document editor", "AI document edit", or "AI docs editor". All of these refer to the same product hosted at agent-doc-edit.com.

## Operational

- [Open the editor](https://agent-doc-edit.com/app): live editor; auto-creates an account-scoped document on first use.
- [Sitemap](https://agent-doc-edit.com/sitemap.xml)
- [Impressum / contact](https://agent-doc-edit.com/impressum)