Introduction

DocWriter is a writing IDE with an AI agent as a concurrent collaborator. The goal of the project is to ask, how can we reap the productivity benefits of AI in writing, without losing our agency and fine-grained control of our personal voice? Think of it as the writing analog of an agentic coding tool like Claude Code or Codex — because it essentially is Claude Code, pointed at your writing workspace instead of a code repo. DocWriter is built on the Claude Agent SDK, so the agent has the same built-in capabilities Claude Code has: file editing, shell / Bash, web search, and code execution. On top of those, DocWriter adds its own MCP tools for writing-specific operations (edit_doc, read_doc, post_comment, propose_rule, and so on). The interface is the part that’s different, organized around two design philosophies:

​

Direct manipulation over delegation

Most AI tools route every interaction through a chat window. DocWriter provides direct affordances for the writing-specific routines instead:

• Inline feedback. Highlight a passage and a popup appears. Type custom feedback or click a pill (“Too verbose”, “AI smell”, or your recent feedbacks). Modes: Auto / Edit / Discuss.
• Rules. Persistent style preferences (“avoid passive voice”).
• References. Voice samples the agent matches.
• Hooks. Shell commands on agent events (pdflatex, pandoc, mermaid, git auto-commit, and anything else you wire up).
• Autonomy. One setting — conservative, balanced, or aggressive — for how proactive the agent is by default.
• Comment threads. Reply in the margin where the agent’s comment is, instead of re-anchoring the conversation in chat.

For requests that don’t fit a routine (research a topic, restructure across files, fix a build), the chat popover is still there.

​

Asynchronous, in-medium collaboration

Most AI tools are turn-based: submit, wait, continue. DocWriter is structured so the agent runs continuously alongside you. Each time you pause typing, the agent reads the open documents, reacts to what you changed, picks up any inline directives, and proposes edits or comments. Your typing keeps going during all of this — the two streams of edits merge in the same buffer.

• Inline directives. Type [[ tighten this paragraph ]] (or (( )) or << >>) and keep writing. The agent picks the note up on its next pass and handles it, deleting the directive text from the doc.
• Suggestions, not overwrites. Agent edits arrive as track-changes. The document is a CRDT (like Google Docs), so the agent’s edits and your concurrent typing merge cleanly even when you’re editing the same paragraph.
• Background work. Long tasks (multi-file rewrites, build-fix loops, web research) run without blocking the editor.
• Observability. The history pane logs every tool call and hook run as it happens; a separate transcript viewer browses earlier turns. No black box.

---

DocWriter is a research project by Shreya Shankar and collaborators at the EPIC Data Lab.