MCP (the Model Context Protocol) lets AI assistants like Claude and ChatGPT reach out to external tools and data. Most research-related MCP servers are generic plumbing — web search, file access, scratchpad memory — useful but stateless about what you've actually read. We lead with Pith (our server) because it's the persistent, cited reading-memory in this list, then cover the notable real servers you'd pair it with, honestly noting what each does and doesn't give you.

pithlab.app

1. 1

   Pith

   Visit site

   Pith is the reading-memory MCP server: it exposes your own cited wiki — built automatically from the articles you bookmark — so Claude or ChatGPT can query what you've read, with source citations rather than copy-pasted context. Unlike generic tool servers, it's a persistent, EU-hosted (Frankfurt) knowledge layer that remembers across sessions and ties every answer back to a source. Best when you want your assistant grounded in your reading rather than the open web.

   Good for: Querying your own cited reading memory from inside an assistant

   ⚠ It only serves knowledge you've saved into Pith, and the product is young and DACH-focused — it's not a general-purpose web or filesystem server.

2. 2

   Notion MCP

   Visit site

   The official Notion MCP server gives AI assistants secure, OAuth-authenticated access to your Notion workspace, with around 18 tools focused on search and content retrieval. If your notes and docs already live in Notion, it lets Claude or ChatGPT read and act on them in place. It's hosted and actively maintained by Notion.

   Good for: Letting assistants read and search an existing Notion workspace

   ⚠ It only reaches what's in Notion, and it surfaces pages as-is without any citation or provenance layer over the content.

3. 3

   Brave Search MCP

   Visit site

   Brave's official MCP server wires the Brave Search API into your assistant for web, news, image, video, and local search, plus AI summarisation. It's a privacy-respecting way to give a model fresh, grounded web results instead of relying on stale training data. Open-source and easy to drop into Claude Desktop with an API key.

   Good for: Giving an assistant live, privacy-respecting web search

   ⚠ It's a stateless search pipe — it fetches results but doesn't remember or organise anything you read across sessions.

4. 4

   Filesystem MCP

   Visit site

   The official Filesystem reference server lets an assistant read, write, and search files in directories you explicitly allow, with path-based access controls and dynamic Roots. It's the simplest way to point a model at a local folder of papers, notes, or drafts. Maintained as part of the core MCP servers repo.

   Good for: Working with a local folder of documents and notes

   ⚠ It's raw file access with no understanding of content, citations, or structure — organising and grounding are entirely on you.

5. 5

   Memory (knowledge graph) MCP

   Visit site

   The official Memory server gives assistants a persistent knowledge graph of entities, relations, and observations, stored locally as JSON, so facts survive across sessions. It's a lightweight way to let a model remember preferences, project details, and decisions over time. A good general-purpose scratchpad memory.

   Good for: Persisting facts and preferences across assistant sessions

   ⚠ Memories are whatever the model chooses to write down — there's no source grounding, so entries aren't cited or verifiable.

6. 6

   Exa MCP

   Visit site

   Exa's MCP server connects assistants to Exa's neural web search, deep search with query expansion, code search, and company research. It's tuned for retrieving high-quality, relevant pages — academic papers, GitHub, Wikipedia, company data — rather than raw keyword hits. Free and open-source, with full features behind an Exa API key.

   Good for: High-quality neural web and research retrieval

   ⚠ Like other search servers it's stateless and usage-metered — it finds sources but keeps no persistent, cited memory of them.

7. 7

   arXiv MCP

   Visit site

   The arXiv MCP server lets an assistant search, fetch, and analyse papers directly from the arXiv preprint repository, with one-click install options for Claude Desktop. It's a focused way to pull recent scientific literature into a research conversation. Community-built and open-source.

   Good for: Pulling and analysing arXiv preprints inside an assistant

   ⚠ It's scoped to arXiv only (no journals or non-preprint sources) and is community-maintained rather than officially supported.

Last reviewed: 7 June 2026 · CC BY 4.0 · cite freely with attribution to Pith.