MCP server de Diffbot

Diffbot MCP Server

Web knowledge fetching tools for your agent.

One click adds the hosted server and signs you in with a Diffbot account, no token to paste. For the local uvx diffbot-mcp server instead: Cursor · VS Code.

What is Diffbot? Diffbot is an applied AI lab building models that structure all of human knowledge. Our commercial product suite make it possible to fetch and structure knowledge at the scale of the entire public web.

This is the MCP counterpart to diffbot-skills without requiring bash access.

Tools

Web

🧬 extract

Web fetch tool alternative. Conventional fetch tools return either 500k+ tokens of HTML or a markdown approximation of the page's content. Extract classifies the page it is fetching and extracts content into a meaningful and consistent JSON schema. Responses are optimized for minimal token usage while maintaining data fidelity. Learn more..

🔎 search_web

Web search tool that ranks accuracy, not popularity. Built on Diffbot's independent web index. Ranking model heavily favors primary sources over secondary sources (e.g. press release > news piece on said press release). Returns markdown formatted content chunks for source citation, with num_results and max_tokens controls. Learn more.

Also includes a page cache reader. Query url:<URL> and it returns Diffbot's already-parsed copy of that page from the index in a fraction of the time for live fetch. Use it for full content from reference prose (docs, articles, papers) and reach for extract when the page's current state matters.

🕷️ crawl

Find all links on a web page (the seed) and extract them all. A classic web crawl without the need for scraping rules. Crawls run as background jobs and can be polled for status. Crawl automatically classifies and structures pages in a standard ontology for easy querying. Use crawl to archive documentation, take snapshots of web projects, or site mapping. Learn More.

🧠 resolve_entities

Recognize and resolve entities in text. Uses Diffbot NLP to recognize organizations, people, places, addresses, and other proper nouns in text. Entities are resolved to entity IDs in the Diffbot Knowledge Graph, and the response includes a ready-made dql_filter (id:or(...)) to fetch their full records with dql. Use to keep AI hallucinations in check and explore deep research opportunities. Learn more.

Knowledge Graph

The Diffbot Knowledge Graph contains over 10 billion entities: Organizations, people, news, places, and more, all linked together by their relationship to each other. It is a representation of things, facts, and connections mentioned by the trillions of websites that make up the public web.

📰 search_news

News and articles by mentioned entity, topic, publisher, language, date range, or sentiment. Newest first by default, relevance-ranked inside a date window.

🏢 search_organizations

Companies by industry, headquarters, headcount, revenue, funding, ownership, or leadership, plus similar_to lookalike lists for competitor mapping.

👤 search_people

People by title, employer, employer industry, skills, education, location, or nationality. Covers people with a public online professional presence only.

🗺️ search_places

Cities, counties, states, countries, and points of interest by name, population, prominence, containing place, or proximity.

💸 search_deals

Funding rounds, investments, and acquisitions by date, size, series, investor, acquirer, or industry.

🎯 dql

Raw querying of the Diffbot Knowledge Graph. The general-purpose layer under the five tools above, for any entity type or query shape they do not cover: products, patents, brands, job posts, cross-entity queries, and facet aggregations. Returns compact csv rows with per-type default columns, or full json records on request.

🔬 dql_probe

Parallel hit counts for several candidate queries. The way to check that a query is neither too broad nor too narrow before pulling rows, and to compare the {} co-constrained form of a query against the loose one.

🧭 dql_ontology

Helper utility for accurate dql ontology navigation. Holds the keys to Diffbot's entire ontology and taxonomy reference: entity types, fields, taxonomies, and enums. A guessed field name is silently ignored by the graph, so this is how the exact spelling gets confirmed. Humans don't need to touch this, but feel free to satisfy your curiosity.

Example Prompts

# extract
→ Install Diffbot agent skills following these instructions https://www.diffbot.com/docs/interfaces/agent-skills

# crawl
→ Crawl Philips' websites for product data sheet PDFs

# search_organizations
→ AI companies based in San Francisco started by previous YC founders

# search_deals
→ recent M&As in pharma

Authentication

A free Diffbot token is required for tool use. Get one here.

MCP clients that support OAuth (Claude, Cursor, VS Code) can connect to https://mcp.diffbot.com/mcp without a token and sign in with a Diffbot account instead.

Otherwise the server looks for a token in this order:

1. Authorization: Bearer <token> header or ?token=<token> query parameter (HTTP transport)
2. DIFFBOT_TOKEN or DIFFBOT_API_TOKEN environment variable
3. A DIFFBOT_API_TOKEN=<token> line in ~/.diffbot/credentials, the same file the db CLI and diffbot-skills read, so one credential serves all three

To enable OAuth on a self-hosted HTTP server, set:

Variable	Example
DIFFBOT_OAUTH_ISSUER	https://app.diffbot.com
MCP_BASE_URL	Public URL of this server, e.g. https://mcp.diffbot.com
DIFFBOT_OAUTH_CLIENT_ID / DIFFBOT_OAUTH_CLIENT_SECRET	Confidential client allowed to call the dashboard's introspection endpoint

OAuth is off when DIFFBOT_OAUTH_ISSUER is unset. With docker compose, MCP_BASE_URL and DIFFBOT_OAUTH_CLIENT_ID are set in docker-compose.yml; DIFFBOT_OAUTH_ISSUER and DIFFBOT_OAUTH_CLIENT_SECRET go in a .env next to it, which the production deploy writes from the DIFFBOT_OAUTH_CLIENT_SECRET GitHub Actions secret. The server runs without it.

Get Started

Two ways to connect. The remote server is hosted by Diffbot and deploys from this repo. The local server runs on your machine from PyPI with uv.

Remote server

https://mcp.diffbot.com/mcp?token=<YOUR_DIFFBOT_TOKEN>

Local server

uvx diffbot-mcp

uvx installs the package into its own environment and runs it. The process reads a token from DIFFBOT_TOKEN, DIFFBOT_API_TOKEN, or ~/.diffbot/credentials.

Setup with MCP Clients

For a local process, this config works with Claude Desktop and most other MCP clients. Follow the official installation guide for where the file lives.

{
    "mcpServers": {
        "diffbot-mcp": {
            "command": "uvx",
            "args": ["diffbot-mcp"],
            "env": {
                "DIFFBOT_TOKEN": "<YOUR_DIFFBOT_TOKEN>"
            }
        }
    }
}

To use the remote server instead, point the client at https://mcp.diffbot.com/mcp?token=<YOUR_DIFFBOT_TOKEN>.

For a Docker install, build the image with docker build -t diffbot-mcp . and use this configuration.

{
    "mcpServers": {
        "diffbot-mcp": {
            "command": "docker",
            "args": [
                "run",
                "--rm",
                "-i",
                "-e", "DIFFBOT_TOKEN=<YOUR_DIFFBOT_TOKEN>",
                "diffbot-mcp"
            ]
        }
    }
}

The repo also packs as a Claude Desktop extension, so the local server installs without editing a config file. Download diffbot-mcp-<version>.mcpb from the latest release, then double-click it, drag it onto the Claude Desktop window, or use Settings > Extensions > Install Extension. The GitHub release carries the file once the v* tag workflow has run.

Claude Desktop asks for a Diffbot API token on install. Leave it blank to use ~/.diffbot/credentials instead. Python and the server's dependencies are resolved by uv, which Claude Desktop manages itself, so there is nothing else to install.

To build the file from a checkout:

npm install --global @anthropic-ai/mcpb
python3 scripts/sync_version.py        # stamp manifest.json from pyproject.toml
mcpb pack . diffbot-mcp.mcpb

Requires uv on PATH. The extension starts the server with uvx.

gemini extensions install https://github.com/diffbot/diffbot-mcp

Gemini CLI asks for a Diffbot API token on install. Leave it blank to use ~/.diffbot/credentials.

For remote setup, simply enter the remote server URL and include "http" as the transport type.

{
  "servers": {
    "diffbot-mcp": {
      "type": "http",
      "url": "https://mcp.diffbot.com/mcp/?token=<YOUR_DIFFBOT_TOKEN>"
    }
  }
}

For a local server, this config prompts for the token instead of storing it in the file.

{
	"servers": {
        "diffbot-mcp": {
            "command": "uvx",
            "args": ["diffbot-mcp"],
            "env": {
                "DIFFBOT_TOKEN": "${input:diffbot-api-key}"
            }
        }
	},
	"inputs": [
		{
			"password": true,
			"id": "diffbot-api-key",
			"type": "promptString",
			"description": "Diffbot API Key"
		}
	]
}

Development

python3 -m venv env && env/bin/pip install -e '.[dev]'
env/bin/python -m pytest tests/test_builder.py     # generated-DQL assertions, no network
env/bin/python tests/smoke.py                      # one live call per tool, needs a token
env/bin/python tests/eval_routing.py               # real questions through `claude -p`, asserts tool routing; costs model tokens
MCP_TRANSPORT=http env/bin/python diffbot_mcp_server.py & ./test.sh   # HTTP transport smoke

The code lives in diffbot_mcp/: api.py (token resolution and the shared request helper), dql.py (query execution, default columns, the cached ontology), dql_builder.py (pure DQL assembly), and one module per tool under tools/. diffbot_mcp_server.py is the entry point.

Releases are tag-driven: bump the version in pyproject.toml, run python3 scripts/sync_version.py, then push a vX.Y.Z tag — see RELEASING.md.

Help & Support

This project is built and maintained by @jeromechoo. For support, file an issue and he'll get back to you as soon as he can. For faster answers, write to sales@diffbot.com 😜.