A connector is executable access to something you own, so every record here states what it exposes as plainly as what it enables. All six are Anthropic's reference implementations, which their own README says are for demonstrating MCP rather than production deployment.
These records describe what providers officially document. benchr has not run these capabilities itself, so nothing here is a test result. Each record shows the vendor page it was read from and the date.
Ledger updated: September 3, 2026
Filesystem
Node.js
AnthropicMITDocs checked September 3, 2026
Read, write, move and search files under directories you explicitly allow.
Why this comes up: It is the difference between pasting a file into a chat and letting an agent work through a project. Everything else in an agent workflow tends to start here.
docker run -i --rm --mount type=bind,src=/path/to/dir,dst=/projects/dir mcp/filesystem /projects
What it unlocks
An agent reads and edits real files instead of a pasted excerpt.
Directory trees come back as JSON, so an agent can reason about structure before touching anything.
Pattern-based edits through edit_file rather than rewriting whole files.
Requirements
Node.js, and at least one allowed directory - the server refuses to start without one.
Access is scoped by command-line arguments, or dynamically through the MCP Roots protocol, which the README recommends because it updates without a restart.
What it puts at risk
This is write access to your filesystem. Scope it to a project directory, never a home directory.
write_file overwrites existing files, and edit_file is documented as not idempotent - run it twice and you may not get the same result.
All tools set openWorldHint: false, meaning they touch the local filesystem only.
A reference implementation by the maintainer's own description, not a hardened production server.
Retrieve a web page and convert it to markdown, with length and offset controls for paging through long documents.
Why this comes up: The cheapest way to give a model a specific page you already trust, without a browser or a search bill.
uvx mcp-server-fetch
python -m mcp_server_fetch
What it unlocks
A named URL becomes readable context, converted to markdown rather than raw HTML.
max_length and start_index let a long page be read in pieces instead of blowing the context window.
A raw mode when the markdown conversion loses something that matters.
Requirements
Python with uvx or pip.
Nothing else - this is the smallest useful server in the set.
What it puts at risk
The README states plainly that this server can access local and internal IP addresses and may represent a security risk. On a machine with internal services, that is the whole warning.
robots.txt is respected for model-initiated requests and ignored for user-initiated ones, and --ignore-robots-txt disables the check entirely. Know which mode you are in before pointing it at someone else's site.
A reference implementation, not a hardened production server.
A local knowledge graph of entities, relations and observations that survives between conversations.
Why this comes up: Every session otherwise starts from zero. This is the difference between a tool you re-explain each morning and one that already knows the project.
A structured thinking tool: numbered thoughts that can be revised or branched rather than written once and defended.
Why this comes up: Long problems fail when an early wrong turn never gets revisited. This makes revision an explicit move instead of something a model has to be talked into.
Numbered reasoning steps with an explicit totalThoughts, so a plan has a shape before it has an answer.
isRevision and revisesThought let a later step correct an earlier one on the record.
branchFromThought and branchId allow two approaches to be carried in parallel.
Requirements
Node.js.
Nothing else - it stores no data and touches nothing outside the conversation.
What it puts at risk
The lowest-risk server in this set: no filesystem, no network, no persistence.
The README does not state when the tool is counterproductive; on simple lookups the structure is overhead rather than help, and benchr has not tested where that line sits.
A reference implementation, not a hardened production server.
benchr