Ref MCP

A ModelContextProtocol server that gives your AI coding tool or agent access to documentation for APIs, services, libraries etc. It's your one-stop-shop to keep your agent up-to-date on documentation in a fast and token-efficient way.

For more see info ref.tools

Agentic search for exactly the right context

Ref's tools are designed to match how models search while using as little context as possible to reduce context rot. The goal is to find exactly the context your coding agent needs to be successful while using minimum tokens.

Depending on the complexity of the prompt, LLM coding agents like Claude Code will typically do one or more searches and then choose a few resources to read in more depth.

For a simple query about Figma's Comment REST API it will make a couple calls to get exactly what it needs:

SEARCH 'Figma API post comment endpoint documentation' (54 tokens)
READ https://www.figma.com/developers/api#post-comments-endpoint (385 tokens)

For more complex situations, the LLM will try to refine it's prompt as it reads results. For example:

SEARCH 'n8n merge node vs Code node multiple inputs best practices' (126)
READ https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-base.merge/#merge (4961)
READ https://docs.n8n.io/flow-logic/merging/#merge-data-from-multiple-node-executions (138)
SEARCH 'n8n Code node multiple inputs best practices when to use' (107)
READ https://docs.n8n.io/code/code-node/#usage (80)
SEARCH 'n8n Code node access multiple inputs from different nodes' (370)
SEARCH 'n8n Code node $input access multiple node inputs' (372)
READ https://docs.n8n.io/code/builtin/output-other-nodes/#output-of-other-nodes (2310)

Ref takes advantage of MCP sessions to track search trajectory and minimize context usage. There's a lot more ideas cooking but here's what we've implemented so far.

1. Filtering search results

For repeated similar searches in a session, Ref will never return repeated results. Traditionally, you dig farther in to search results by paging to the next result but this approach allows the agent to page AND adjust the prompt at the same time.

2. Fetching the part of the page that matters

When reading a page of documentation, Ref will use the agent's session search history to dropout less relevant sections and return the most relevant 5k tokens. This helps Ref avoid a big problem with standard fetch() web scraping which is when it hits a large documentation page you can easily end up pull in 20k+ tokens into context, most of which are irrelevant.

Why does minimizing tokens from documentation context matter?

1. More context makes models dumber

It's well documented that as of July 2025 that models get dumber as you put in more tokens. You might have heard about how models are great with long context now and that's kind of true but not the whole picture. For a quick primer on some research, checkout this video from the team at Chroma.

2. Tokens cost $$$