> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orchestrahq.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Edit your paper with Claude Code

> Your own Claude Code reads a paper in your research workspace, proposes edits that land as suggestions for your co-authors to accept, compiles it and checks its references, all through Orchestra's MCP server.

In a research workspace, Claude Code on your own machine can work on a paper
the way a co-author does: read every file, propose changes, compile, read the
errors and try again. It runs on your own Claude Code subscription or key.
Orchestra gives it the paper and keeps the record.

Nothing it writes goes into the paper directly. Each edit lands as a
suggestion in the paper's review mode, word by word, credited to Claude Code
and to you, for someone who can edit the paper to accept or reject.

## Set it up

1. Install the Orchestra CLI and sign in to your research workspace, if you
   have not: see [Install the CLI](https://docs.orchestrahq.ai/install-cli.md). If it is installed already,
   update it so it has the paper tools (see [Updating](https://docs.orchestrahq.ai/install-cli.md#updating)).

2. Register Orchestra with your agents:

   ```bash theme={null}
   orchestra enable
   ```

   This adds Orchestra's MCP server to Claude Code for every project (see
   [Orchestra in your agents](https://docs.orchestrahq.ai/mcp-server.md)). In a research workspace the paper
   tools are on by default.

3. Start Claude Code anywhere and ask it about your paper, by its title:

   ```
   Read the paper "Sparse adapters" and tighten the introduction to half a page.
   Compile it and fix anything that breaks.
   ```

4. Open the paper in Orchestra. The changes wait in **Review**, marked as
   Claude Code's, with your request beside them. Accept or reject each one, or
   a whole turn at once.

If you set `mcp.tools` yourself, add `papers` to it to keep these tools (see
[Turn on more tools](https://docs.orchestrahq.ai/mcp-server.md#turn-on-more-tools)).

## What Claude Code can do

| Tool | What it does |
| - | - |
| `list_papers` | The papers you can see, each with your role on it |
| `read_paper` | A paper's files, each with its path, its revision and its text |
| `propose_paper_edit` | A change to one file, as a range of its text or the whole file as it should be, with a short reason. It lands as a suggestion |
| `compile_paper` | Compiles the paper on Orchestra's compile service and returns each error and warning with its file and line. By default it compiles with your own waiting suggestions in place, so Claude Code can check a change builds before anyone accepts it. That compile is a draft: it is never shown as the paper's own |
| `read_compile_log` | The end of the compile log, for errors the line list does not place |
| `check_references` | Checks every cited reference against Crossref and arXiv |
| `check_claims` | Checks that each cited sentence is supported by the cited paper, where your lab has its PDF on the project |
| `search_lab_evidence` | Your lab's runs, notebook results, sessions, decisions and uploaded papers that bear on what Claude Code is writing, each with a link |

## Who can do what

The paper's roles apply to Claude Code exactly as they apply to you:

* **Can edit** and **can review**: Claude Code reads, compiles, proposes
  changes and checks claims.
* **Can view**: Claude Code reads and compiles, and is told it cannot propose
  changes or check claims.
* A paper you cannot see does not exist for Claude Code either.

Only someone who can edit the paper accepts a suggestion.

## When the paper changed meanwhile

Claude Code proposes each change against the revision it read. If someone
changed the same words since, the change is refused, and Claude Code reads the
file again and proposes on the text as it is now. A whole-file change is
refused whenever the file changed after Claude Code read it, so it never undoes
someone else's edit.

## Where the work shows

Your Claude Code session is captured as usual. Each suggestion it makes is
linked to that session, so:

* the paper's **Sessions** list shows the session, with who ran it, what it
  changed and how many of its suggestions were accepted or rejected;
* **Who wrote this** names the session behind every passage it wrote that was
  accepted, and opens it.

Anyone who opens the session sees it only if they can see the session itself.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.