Connect Claude Code to your indexed papers, figures included

Agent Bayes now runs an MCP server. Claude Code, Cursor, Claude Desktop, or any other agent that speaks the protocol can search the papers you indexed, ask for cited answers, and read or build mindmaps, the trees of cited claims Agent Bayes builds from your papers. Every citation comes back attached to the passage it came from.

As of this release, the server also lets your agent look at the pages themselves: figures, maps, charts, plates, and the images stored on mindmap nodes.

This post covers what the connection gives you, the two ways people tend to use it, and why the vision tools matter in any field where the evidence lives in a figure.

A general agent needs evidence it cannot invent

Many researchers now draft, analyze, and code inside a general-purpose agent. Those agents are good at structure and prose. They are weak at the one thing a researcher cannot give up: a claim tied to a real passage on a real page.

Ask a general agent about your field and it answers from its training data. It produces references that look right, and nothing in the loop checks them against a paper.

The MCPModel Context Protocol, an open standard that lets an AI agent call tools and read data from an external service. server gives the agent a different place to get evidence. Every search result is a passage from your own corpus, with its page number and full reference attached. The agent can quote it, cite it, and look it up again later.

What the connected agent can do

Everything runs inside your own account, through an API key you create in the app. The tools cover the whole research loop:

• Build the corpus: create knowledge bases (collections of indexed papers) and projects (workspaces that search one or more knowledge bases), then upload PDFs from disk.
• Search: send up to 10 query variants in one call and get back cited passages.
• Ask: put a question to project chat, where the Agent Bayes research agent answers in short, cited claims.
• Map: run the research agent on a mindmap, then read it, edit it, or undo its changes.
• Look: view rendered PDF pages and the images on mindmap nodes.
• Account: check your credits and see which models your plan includes.

In a mindmap, each claim node carries the citation keys of the passages that support it, and parent nodes act as section titles over the claims beneath them.

Borrow the evidence, or delegate the mindmap

People use the server in two ways, and many mix them.

In the first, your agent is the author. It uses search to scout a topic and find the right sources, then asks project chat when it needs a synthesized conclusion rather than raw passages. It writes the draft, the review, or the analysis itself, carrying each citation along with the claim it backs.

In the second, the mindmap is the deliverable. Your agent creates a mindmap, sends the research question, waits for the run to finish, and reads the result. Follow-up questions can point the research agent at specific nodes, so each round deepens one branch instead of starting over.

The first suits people who already have a writing workflow in their agent. The second suits people who want Agent Bayes to do the synthesis and their agent to drive it.

Citations survive the handoff because the server teaches the rules

The weak point of any chain of agents is summarization. Each step compresses text, and citations are the first thing to fall off.

The server deals with this before the first real tool call. Every session starts with an operating manual, and the other tools refuse to run until the agent has read it. The manual tells the agent to keep citation keys verbatim with the claim they back, to prefer printed page numbers when a document has them, and to build bibliographies from the stored full reference without retyping it.

Answers from the research agent come back with a short reminder of those rules attached. And any citation key the agent carries can be resolved back to its passage with rag_resolve_chunks, so a claim can be checked again long after the search that produced it. The agent does not have to remember a reference. It can ask for the passage behind it.

Costs are visible before they are spent

Reading is free: search, listings, mindmap reads and edits, and viewing images cost no credits. The tools that do spend credits say so at the start of their description, in capital letters, so the agent sees it before it calls them.

Indexing costs 0.6 credits per page, reserved up front. The manual tells the agent to count the pages locally and tell you the total before it indexes a batch. Indexing calls also accept a retry key, so retrying after a timeout with the same key does not charge twice.

Some evidence is a picture

In archaeology, geography, biology, medicine, and engineering, a paper's key evidence often sits in a figure: a site plan, a section drawing, a growth curve, a microscope image. Text search finds the caption, and the caption often leaves the finding out.

Agent Bayes already writes a text description of every figure, map, chart, and photo when it indexes a PDF, and those descriptions are searchable and citable like any passage. A description is still a summary. Sometimes the agent needs to see the figure itself, to read a value off an axis, compare two plates, or check whether a map shows what a sentence claims.

Two tools now make that possible.

document_view_pages returns images of up to 4 pages of an indexed document. The agent names the pages the same way search results do: PDF page numbers from page_range, or printed page labels from actual_page_range. Each image arrives with a caption naming the PDF page, the printed page when the document has one, and the source's author and year.

mindmap_view_node_images returns the images stored on up to 4 mindmap nodes. When the agent reads a mindmap, nodes that hold an image are marked has_image, so it knows which ones to ask for.

A typical round looks like this. Search returns a passage that describes a settlement map on a given page. The agent passes that page to document_view_pages, reads the map, and checks the claim against the map before it writes the sentence.

Viewing pages costs no credits, but it does cost context. Each page image takes space in your agent's context window. That is why the manual tells the agent to find the page with search first and only then look at it. Your agent also needs a model that accepts images, which current Claude models do.

Connecting takes one command

First, create an API key in the app under Settings → API Keys. The settings guide covers creating, naming, and revoking keys. Every plan includes API access.

Then add the server to Claude Code:

claude mcp add --transport http agent-bayes https://api.agentbayes.com/mcp/ \
  --header "X-API-Key: sk-your-key"

Cursor and other clients that take a JSON config use the same URL and header:

{
  "mcpServers": {
    "agent-bayes": {
      "url": "https://api.agentbayes.com/mcp/",
      "headers": { "X-API-Key": "sk-your-key" }
    }
  }
}

The MCP page has the Codex config and the rest of the setup details.

The agent fetches the operating manual on its own when the session starts. From there, a first request can be as plain as "find what my papers say about this question, with page numbers", or "open page 14 of that paper and tell me what the map shows."

Which figure in your field would you want an agent to check before it trusts the caption?