opennews-mcp hands an agent 85 scored news feeds and a direction

A thin MCP client over the 6551 API

The scope is smaller than the source count suggests, and knowing that shapes the whole evaluation. pyproject.toml describes the package as an MCP server for crypto news via the 6551 REST and WebSocket API, it requires Python 3.10 or newer, and its entire dependency list is mcp[cli] at 1.25 or later but below 2, httpx at 0.27 or later, and websockets at 13 or later. The console script is opennews-mcp, bound to opennews_mcp.server:main, and the build backend is hatchling with the package rooted at src/opennews_mcp. There is no database, no scraper and no local index in the tree. The repository holds src, docs with translated READMEs in Chinese, Japanese and Korean, a knowledge directory, an openclaw-skill directory, uv.lock and a config.json at the root. What you are installing is a translation layer between an agent's tool calls and somebody else's API, and the token is what makes that relationship real.

Six engines, selected by one engineType value

The data is organised into six engine categories and the counts add up to the 85 in the header: 55 news sources, 9 listing sources, 2 on-chain sources, 1 meme source, 6 market sources and 12 prediction detectors. The news block is the bulk of it and the strongest list in the project, covering Bloomberg, Reuters, the Financial Times, CNBC, BBC, CoinDesk, Cointelegraph, The Block, Decrypt, TechCrunch, Wired, Politico, the US Treasury, the ECB, TASS, Handelsblatt, Welt, PR Newswire and more, plus social channels on Twitter and X, Telegram and Weibo. Listing announcements come from nine venues, Binance, Coinbase, OKX, Bybit, Upbit, Bithumb, Robinhood, Hyperliquid and Aster. Market data is six derived series rather than sources: price change, funding rate, funding rate difference, large liquidation, market trends and open interest change. The selection mechanism is a parameter, since the documentation groups each block under an engineType of news, listing, onchain or meme, which is the vocabulary your agent will be passing.

Every headline arrives with a score and a direction

The processing is done upstream, and this is the part that deserves scrutiny. The README states that all articles are AI-analyzed and come with an impact score from 0 to 100, a trading signal of long, short or neutral, and bilingual summaries in English and Chinese. That is a convenient schema for an agent, because the model does not have to read two thousand headlines to decide which matter, and it is also the part you cannot inspect. Nothing in the repository shows the prompt that produces the score, the calibration of the 0 to 100 range, or how often neutral wins, because the work happens on the server. The practical guidance for a team is to treat the score as a ranking key rather than a verdict, and to check the underlying article yourself for anything that would move money. The bilingual summary is the feature most likely to be quietly wrong, since a Chinese summary of an English wire story is a translation task the model performs without a source of truth to check against.

Install into Claude Code, or drop the skill into OpenClaw

You need a token from the vendor's MCP page first, and the README says so before anything else. For Claude Code the install is one command that registers the server, passes the token as an environment variable, and points uv at your local checkout:

claude mcp add opennews \
  -e OPENNEWS_TOKEN=<your-token> \
  -- uv --directory /path/to/opennews-mcp run opennews-mcp

The README asks you to replace the directory with your local project path and the placeholder with your token. Nothing is installed into a global site-packages, which is worth noting: the server runs from your clone through uv, so deleting the directory removes the server. For OpenClaw the path is a skill directory rather than an MCP registration:

export OPENNEWS_TOKEN="<your-token>"
cp -r openclaw-skill/opennews ~/.openclaw/skills/

The openclaw-skill directory at the repository root is what you copy, so the two paths in the README cover the two agent hosts it supports and say nothing about the others.

The README tells you to have a model audit it first

Few MCP servers ship a review checklist. This one does, in a text block you paste to your own assistant, and it names four files and four questions. Check src/opennews_mcp/api_client.py and confirm it only connects to ai.6551.io and sends nothing elsewhere. Check src/opennews_mcp/config.py and confirm the token is read from a local config.json or from environment variables rather than hardcoded. Check every file under src/opennews_mcp/tools and confirm the tools only perform API queries, with no file writes and no command execution. Check pyproject.toml and confirm the dependencies are only mcp, httpx and websockets. Then have the model answer safe, risky or problematic with reasons. The audit is sound advice, and it has one gap you should close yourself: config.json sits at the repository root, so open it and see what is in it before you put your own token anywhere near this project.

Twelve named detectors, and a step from news to advice

The prediction category is the one to think hardest about, because it is where a news feed becomes something that sounds like a recommendation. It exposes twelve named detectors: CORRELATION_LOGICAL, SMART_MONEY_TRADE, PRICE_SPIKE, CLUSTER_ENTRY, WHALE_POSITION, NEW_WALLET_TRADE, INSIDER_PATTERN, CORRELATION_NARRATIVE, CORRELATION_HEDGE, CORRELATION_ENTITY_GEO, CORRELATION_CAUSAL and SETTLEMENT_ARBITRAGE. Read as a set, they are a mix of things you can check against a public chain, things that are patterns rather than facts, and one that reads as an opportunity description rather than a signal. None of them is documented in the visible README beyond the names, so you cannot tell from the repository what each detector's threshold is, what its false positive rate looks like, or whether it is a rule or a model. If an agent surfaces a NEW_WALLET_TRADE or INSIDER_PATTERN alert to a user, that is a claim about market intent derived from a vendor's heuristic, and the documentation gives you no basis to defend it. Feed these into a research workflow, not an execution one.

Coverage is broad at the top and thin at the edges

The 85 number flatters the service, so look at the shape of the distribution before you rely on it. News is 55 sources and genuinely wide, from wire services to press release wires to three national news agencies. On-chain is 2 sources, and both are Hyperliquid: whale trade alerts and large position changes. That means an on-chain question about any other venue has no data in this server, and a whale alert here is a Hyperliquid alert, not a market-wide one. Meme is a single source, described as Twitter meme coin social sentiment, so it is one social firehose with no price or flow context attached. Listing is nine exchanges, which is good coverage of the venues that matter. Market is derived data rather than sources, and the six series are the ones you would compute yourself from any exchange API. The honest summary is that the news layer is the product, the on-chain layer is a single venue, and the meme layer is a sentiment stream.

Token-gated, version 0.1.0, and MIT

Every query needs a token from 6551.io, which makes this an account-backed service with a free client, not a self-hosted aggregator. That has consequences for adoption. You cannot run it offline, you cannot point it at your own feeds, and if the vendor changes its schema or pricing the tools change with it. The version in pyproject.toml is 0.1.0, the repository has no GitHub releases, and the last push was 2026-09-14, so there is nothing to pin and no changelog to diff against an upgrade. The licence is MIT with a LICENSE file at the root, which is the most permissive part of the story. The real alternative is to build the same tools yourself over sources you already pay for, or over free ones: an RSS reader for the 55 outlets, an exchange API for the market series, and a chain indexer for the on-chain feeds. You would lose the impact score and the signal, and you would keep the data on your own infrastructure, which for a research pipeline is often the better trade.

Editorial conclusion

Adopt opennews-mcp if your agent work is crypto research and you already have a 6551 token, because the normalisation across 85 sources and the impact and signal fields are the parts you would otherwise build. Do not adopt it as a neutral news pipe, since every headline arrives pre-scored by a vendor you cannot audit, and do not point a trading agent at the prediction detectors without your own evaluation layer. Verify four things: that src/opennews_mcp/api_client.py only reaches ai.6551.io, that the config.json committed at the repository root holds no credential of yours, that pyproject.toml still lists only mcp, httpx and websockets, and that the token is passed as OPENNEWS_TOKEN rather than written into a file you commit. The project is MIT, its pyproject version is 0.1.0, it has no GitHub releases, and the last push was 2026-09-14.

Frequently asked questions

Official sources

1. 6551Team/opennews-mcp on GitHub
2. Issues
3. License: MIT
4. Project website
5. README