Catalyst evidence graph

ai.catalystprojectv0.2.0Updated Oct 4, 2026

Findings on what compounds and supplements do in the body, with evidence strength, quote and paper.

VerifiedStreamable HTTPWeb executableData & Analytics Knowledge & Memory

Overview

AI-generated overview

Lets an assistant look up curated findings on what compounds, supplements, nutrients and foods do in the human body, each with a graded evidence strength and…

What it does: Catalyst is a hosted, read-only evidence graph covering compounds, supplements, nutrients and foods and their effects in the human body. An assistant can retrieve findings for a compound or an outcome; each finding carries an evidence_strength level (strong, moderate, limited, very_limited or insufficient) computed from the underlying study, the verbatim sentence it came from, the paper, population and dose, and a permalink. It also returns relations from reference databases such as targets, pathways and reactions, labelled as hypotheses rather than reported results. No tool returns products, prices or purchase links, and nothing is ranked as a recommendation.
When to use it: Useful when a conversation touches nutrition, supplements or compound effects and you want sourced, graded evidence rather than an unsourced claim. It suits assistants that need to cite a paper, quote the sentence behind a finding, or filter by evidence strength. It is not a medical-advice or product-recommendation service.
Requirements: A remote MCP endpoint over Streamable HTTP at the provider's host; no account, key or environment variable is required. Any MCP-capable client that supports remote HTTP servers can connect, and setup instructions differ per client. Network access to the provider's host is needed.

Before you install

The graph is read-only and stores no queries, and the provider states it has no accounts. Evidence strength is not a recommendation and the content is not medical advice, so findings should not be presented as treatment guidance. Rate limits are applied per client name and a keyed hash of the network address. Content is subject to the provider's terms and per-source licences.

Installation

In SourceWeft

Open Catalyst evidence graph in the dashboard and add it to a workspace.
Enable the server for the chats that should use its tools.

Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.

Other MCP clients

Add this to your client's mcpServers config.

{
  "mcpServers": {
    "evidence-graph": {
      "type": "http",
      "url": "https://catalystproject.ai/mcp"
    }
  }
}

README

Catalyst evidence graph — MCP server, plugins and extensions

[Image]

Catalyst is a curated evidence graph of what compounds, supplements, nutrients and foods have been shown to do in the human body. This repository connects it to the assistants and editors people already use. The server itself is hosted; everything here only tells a client where it is.

Server: https://catalystproject.ai/mcp — MCP over Streamable HTTP, read-only, no account, no key. Registry name: ai.catalystproject/evidence-graph Setup for every client, the tool list and limits: https://catalystproject.ai/connect

What an assistant gets

Findings for a compound or an outcome. Every finding carries an evidence_strength — strong, moderate, limited, very_limited or insufficient — computed from the study behind it, never set by hand, plus the verbatim sentence it was drawn from, the paper, the population and dose, and a permalink to check it.
Relations from reference databases (targets, pathways, reactions), each labelled as a hypothesis rather than a reported result.

Evidence strength describes how well the evidence supports a finding. It is not a recommendation, and nothing here is medical advice. No tool returns a product, a price or a link to buy anything, and no tool ranks what to take.

Changed in 0.2.0 (ADR-0103): tool results report evidence_strength as a named level and never a letter, and get_findings takes an evidence_strength floor. The grade field and the grade argument are gone; a call that still sends grade gets an error naming the replacement.

Install

Client	How
Claude (web, desktop)	Settings → Connectors → Add custom connector → `https://catalystproject.ai/mcp`
ChatGPT	Developer mode → add a custom app/connector with `https://catalystproject.ai/mcp`
Claude Code (plugin, adds the server and a skill for presenting evidence strength)	`/plugin marketplace add colorcatalyst/catalyst-mcp` then `/plugin install catalyst@catalyst`
Claude Code (server only)	`claude mcp add --transport http catalyst https://catalystproject.ai/mcp`
VS Code / GitHub Copilot	The Catalyst Evidence Graph extension (VSIX on the releases page), the one-click link on /connect, or `.vscode/mcp.json` below
Cursor	One-click install, or `~/.cursor/mcp.json` below
Gemini CLI	`gemini extensions install https://github.com/colorcatalyst/catalyst-mcp`
Anything else	the JSON below

VS Code's own file takes a slightly different shape:

In this repository

Path	What it is
`server.json`	The entry in the official MCP registry
`.claude-plugin/marketplace.json`, `plugins/catalyst/`	A Claude Code plugin marketplace with one plugin: the server and an `evidence-strength` skill
`gemini-extension.json`, `GEMINI.md`	A Gemini CLI extension
`vscode/`	A VS Code extension that registers the server with Copilot's agent mode

Privacy

The server stores no queries and has no accounts. Rate limits count calls per minute against the client's name and a keyed hash of the network address, never the address itself. Full notice: https://catalystproject.ai/privacy.

Contact

Questions, or something an assistant got wrong from this graph: [email protected]

The code in this repository is MIT-licensed. The graph's content carries the terms on https://catalystproject.ai/terms and per-source licences on https://catalystproject.ai/attribution.

Source: README.md at commit df9e1b1

Tools

Tool metadata has not been indexed yet.

Version history

v0.2.0LatestOct 4, 2026