Readwise MCP: Use Readwise in Claude (2026 Guide)

Readwise sits in an odd spot in the knowledge-worker stack. It is one of the very few apps where almost every user runs it for years, builds up a library of thousands of highlights, and still mostly forgets what is in there. The promise of a personal database of everything you have read is real. The day-to-day reality is a backlog of highlights you saw once at import and never opened again.

For most of the app’s history, the only way to do something useful with those highlights was the Daily Review email. That is fine for surfacing four random highlights a day. It is not how you actually want to use your library when you sit down to write a brief or work through a problem. You want to ask a question and get back the relevant highlights you already saved, in your own words, ranked by what matters.

The Readwise MCP server is the thing that closes that gap. It exposes your Readwise library, including Reader documents and highlights, to Claude, ChatGPT, Cursor, or any MCP client through a single hosted connector. This guide walks through the setup that takes about five minutes, the actions you can take once it is wired in, and where it stops being enough.

What Readwise MCP actually exposes

The Model Context Protocol is an open standard released by Anthropic in late 2024 that defines how AI clients talk to external tools. An MCP server exposes a set of tools, the AI client picks one mid-conversation, and the result flows back into the model’s reasoning. The official Readwise MCP server wraps the Readwise API in that interface.

What gets exposed is every highlight in your Readwise library and every document in your Reader inbox, archive, shortlist, and feed folders. The server combines full-text matching with semantic search, so a question like “what have I read about pricing power” returns highlights that argue the concept without needing the literal phrase. The Readwise docs note that the index updates automatically as you add content, and subsequent syncs only process new or changed items so they run quickly in the background.

The action layer is the part most setup guides skip. Read-only access would already be useful. Readwise MCP goes further. You can move documents between inbox, shortlist, and archive. You can add or remove tags on a document. You can create new highlights from a passage you point Claude at. That makes Reader usable as a working surface from inside a chat, not just a place to dump things and forget about them.

The hosted setup in Claude Desktop (5 minutes)

The hosted server is the default in 2026. No local install, no API key in a config file, no npx command to debug. Setup is three clicks and an OAuth grant.

Step 1. Open the Claude Desktop connectors panel. Go to Settings, then Customize, then Connectors. Click “Add custom connector.” The Anthropic walkthrough for this panel is documented in the official custom connectors guide.

Step 2. Paste the Readwise hosted URL. The server lives at https://mcp2.readwise.io/mcp. Paste that into the URL field. Give the connector any name you like (“Readwise” is fine) and confirm.

Step 3. OAuth into Readwise. Claude opens a browser window. Sign in to your Readwise account if you are not already. Click “Authorize.” The connector flips to a connected state in Claude’s panel. You can close the browser tab.

Step 4. Restart Claude Desktop. Cmd-Q on macOS, full close on Windows. A window reload does not pick up new connectors. The Obsidian and Readwise community forums are full of “why is my MCP server not showing up” threads, and roughly half of them are people who reloaded instead of restarted.

Step 5. Test it. Open a new chat and ask “what have I highlighted about second-brain systems?” or “what is in my Reader inbox right now?” Claude calls the Readwise tools, returns the matches, and you are running.

The legacy local install (still works)

The original @readwise/readwise-mcp npm package still works if you want a self-hosted setup. The Readwise team has marked it deprecated in favor of the hosted server, but the install path is identical to any other Node-based MCP server.

Step 1. Generate a Readwise access token. Sign in at readwise.io, go to readwise.io/access_token, and copy the token. Treat it like a password. Do not commit it to a repo.

Step 2. Edit claude_desktop_config.json. Open Claude Desktop, Settings, Developer, Edit Config. The file lives at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows.

Step 3. Add the Readwise block.

{
  "mcpServers": {
    "readwise": {
      "command": "npx",
      "args": ["-y", "@readwise/readwise-mcp"],
      "env": {
        "ACCESS_TOKEN": "your-access-token-here"
      }
    }
  }
}

The -y flag is not optional. Without it npx waits for an interactive install prompt that Claude cannot answer, and the server hangs at startup. Node must be installed for this path to work.

Step 4. Restart Claude Desktop. Same rule. Full quit, not reload.

The local install gives you the same tool surface as the hosted server. The main reason to pick it is if you have a self-hosting policy that rules out hosted connectors, or if you want to fork the source and add your own tools. For most readers, the hosted route is faster and easier to maintain.

Hosted vs local, which to pick

If you are picking from scratch in 2026, use the hosted server. The OAuth path is cleaner, the maintenance is zero, and the same connector works on Claude mobile and ChatGPT Developer Mode without re-doing the install. The local install path is a fallback, not a feature.

What to actually ask Claude once it is connected

A new MCP server feels underwhelming for the first ten minutes and obvious after a week. The prompts that pay off are not “summarize my Readwise library.” They are the specific, small asks that you used to either skip or do by hand.

Surface old highlights during a current task. “What have I highlighted about cold-email open rates?” Claude searches the library, returns the four or five highlights that match, and you remember the framework you forgot you had. This single prompt is most people’s first “oh, this is the point” moment.

Triage your Reader inbox. “What is in my Reader inbox from this week, sorted by source?” The hosted server can pull the inbox and offer to archive items in the same chat. The Readwise blog post about the new MCP server uses this as the headline use case, and it holds up. Reader inboxes accumulate. The MCP turns a 30-minute triage into a five-minute one.

Pull quotes for a brief. “I am writing about pricing strategy. Give me the five strongest highlights from my library on price anchoring.” Claude returns highlights with their source and author, so you can drop the quote into a draft and link it back to the original book or article. The semantic-search step matters here. Direct keyword search misses anything that uses different vocabulary for the same idea.

Cluster what you have been reading. “Show me everything I saved in Reader last month, clustered by theme.” Useful for end-of-month review and obvious for anyone who keeps a “what I am thinking about” page. The Readwise team’s own readwise-skills repo packages this exact pattern as reusable Claude Skills, layered on top of the MCP server.

Create highlights from a draft. “Save this passage as a highlight on the Naval Almanack book in Readwise.” The MCP can write back into your library, not just read from it. Useful when you are referencing something in a Claude chat and want the quote saved for later without context-switching to Readwise.

The pattern across all of these is that the MCP server makes Readwise useful in the moment of work, not after the fact. The Daily Review email is still fine. The MCP layer is the bit that turns a passive library into a tool you reach for.

The five mistakes that break Readwise MCP setups

Six months of forum threads and the same handful of issues come up again and again.

Reloading instead of restarting. Claude Desktop only reads its connector list at full startup. A window reload does nothing. Cmd-Q on macOS, fully close on Windows, then reopen. This is the most common “the connector is not showing up” thread on the Readwise forum and the answer is almost always the same.

Skipping the paid Readwise check. The MCP server works against the Readwise API. The API returns empty for free or expired accounts. If your search calls keep returning zero results no matter what you ask, the first check is whether the underlying subscription is active.

Confusing Readwise with Reader. Readwise (the highlights service) and Reader (the read-later app) are the same subscription but two surfaces. The MCP server exposes both. If you only use Reader and have zero highlights synced from Kindle, Apple Books, or RSS, expect the highlights tools to return less than the documents tools. Your library is doing the work, not the server.

Skipping -y on local-install npx commands. Without -y, npx waits for an interactive “install this package” prompt that Claude cannot answer. The server appears to hang at startup. This applies to dozens of community MCP servers, not just Readwise. The -y is not optional.

Running 20 MCP servers at once. Each connector is a subprocess with startup cost and memory overhead. Past ten, you will feel it. Past twenty, Claude Desktop’s memory use climbs fast and tool calls get slower. The right rule is to keep three to six servers connected at a time, each covering a distinct part of your stack. Readwise plus one bookmarks server plus a notes server is a sane starting baseline.

What Readwise MCP does not cover

This is the honest gap. Readwise is built around long-form reading. Articles, books, Kindle highlights, RSS, podcasts you transcribe into Reader. It is not built around social-feed bookmarking. If you save threads on X, posts on Reddit, or articles inside LinkedIn, none of that lands inside Readwise unless you remember to manually push each one through Reader.

For most knowledge workers, the social-saves stack is a separate input source. X bookmarks alone, for many users, outnumber Reader saves by a wide margin. They are also the input source most likely to disappear silently as platforms change policy. None of it is reachable through Readwise MCP, because none of it is in Readwise.

This is where a dedicated bookmarks MCP earns its slot in the stack. ContextBolt is a Chrome extension that captures bookmarks from X, Reddit, and LinkedIn, tags them automatically, and exposes them through a personal MCP endpoint. The free tier (150 bookmarks) covers casual use. Pro at $6 a month adds unlimited bookmarks, encrypted cloud sync, and the MCP endpoint that any client can call.

Wire it the same way as the Readwise server. Settings, Connectors, Add custom connector, paste your endpoint URL, restart. Claude now has access to two parallel libraries. The long-form one in Readwise and the social-feed one in ContextBolt. Asking “what have I been thinking about this quarter?” returns a real answer because both halves of the input are visible.

For more on the broader pattern of multi-server setups, see Best MCP Servers for Knowledge Workers and Personal AI Context Stack for Claude.

The one opinion worth holding

Almost every “use Readwise with Claude” guide on the internet treats this as a productivity hack. It is not. It is the smallest reasonable version of a personal knowledge stack that actually works.

The hard part of knowledge management is not capturing things. It is reaching for them at the right moment. Readwise has been excellent at the first half for years and weak at the second half by design. The MCP layer is the missing piece. Your library is finally available the way you want it to be available, at the moment you have a question, from inside the chat you would have asked the question in anyway.

The version of this that most people skip is the second-MCP step. Connecting Readwise alone gives you everything you have read in long form. That is half your input. The other half lives in social feeds and stays there unless you do something. Add ContextBolt as the second connector and Claude can answer “what have I been thinking about” with both halves of the answer instead of one. The marginal cost of the second server is small. The marginal value is large. That is the entire decision tree.