The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the NewsBlog Composer listing page.
Give it a news headline or a topic and it verifies the story is real, downloads the sources, extracts the facts with attribution, mines the keywords, writes the article, and builds the SEO, schema, banner prompt and publishing pack around it.
Two independent publishers, or it will not draft. Wire reprints are detected, so six outlets running one Reuters story are not mistaken for six sources, and a wide date spread downgrades confidence rather than passing silently.
Every fact traces to a source it fetched, not to a model's memory. Facts,
quotes and figures carry their source_url, and build_schema rejects reference
links that are aggregator redirects rather than real publisher URLs.
You choose the banner. With no style picked there is no image prompt in the
result at all, and save_and_present refuses the package. The tool asks; it does
not decide for you.
It will not hand you a post without the numbers. save_and_present refuses
unless the SEO audit and the AI-word check have both run, so the finished post
always arrives with its scores attached.
The prose is machine-written and the tool says so. Without a detector API key
score_ai_text reports Not measured rather than inventing a reassuring number
— a local style score measures cliche density and sentence rhythm, which is not a
prediction of what a detector will say. If the byline is going to claim a person
wrote it, that is your call to make with your eyes open; review_draft exists so
the draft gets edited rather than published raw.
It runs with zero API keys. Every credential is an upgrade, not a requirement. See No keys? Start here.
Download newsblog-composer-<version>.mcpb from
Releases.
Claude Desktop → Settings → Extensions → Advanced settings → Install Extension.
Toggle the extension off and on. On Windows the X button only minimises to the tray, so without this the old server keeps running.
Add this line to a Claude project's instructions, or to Customize so it applies to every chat:
When I ask for a blog post, an article, or a news write-up, always use the NewsBlog Composer tools. Start with
write_blog_post. Never search the web and draft the article yourself.
Ask for a post. It asks for your byline, company and blog domain once.
Step 4 is not optional. A host's system prompt reads "write me a blog post"
as "search the web and make an artifact", and an MCP server's instructions
field cannot outrank that — there is no API for a server to claim priority.
Without the line the client answers from web search and never mentions the
extension. Naming the server in the request works for one-offs.
Extensions are desktop only. On mobile or the web app there is nothing to call, and the client answers from web search without saying so.
That gives you the newsblog-mcp command, which speaks MCP over stdio.
Check it reaches the network before wiring it into a client — if search is blocked by a proxy or VPN, this is where you find out:
Create .vscode/mcp.json in your workspace:
Reload the window, then open Chat in Agent mode — the tools appear under the
tools picker. Use the full path to newsblog-mcp (or to the venv's Python with
"args": ["-m", "newsblog_mcp.server"]) if the command is not on your PATH.
.mcp.json in your project:
%APPDATA%\Claude\claude_desktop_config.json on Windows,
~/Library/Application Support/Claude/claude_desktop_config.json on macOS:
The .mcpb above does this for you and is the easier route.
Launch newsblog-mcp (or python -m newsblog_mcp.server) and speak MCP over
stdio. For a remote client, newsblog-mcp --http serves a streamable HTTP
endpoint — see Publishing and connecting.
Installed from PyPI, the server writes to your user data directory, so nothing is lost when you upgrade:
| Windows | %LOCALAPPDATA%\newsblog-composer-mcp |
| macOS | ~/Library/Application Support/newsblog-composer-mcp |
| Linux | ~/.local/share/newsblog-composer-mcp |
That folder holds profile.json (the identity you set at first run), output/
(every generated post) and an optional .env. Set NEWSBLOG_DATA_DIR to put
them somewhere else. Run from a cloned checkout instead and everything stays in
the project folder, beside the code.
python -m newsblog_mcp.diagnose prints the exact paths for your install.
Claude Desktop / Claude Code run it over stdio, which is what the install section above sets up.
ChatGPT cannot spawn a local process, so stdio will never reach it. Run the same server over HTTP and deploy it behind HTTPS:
Then in ChatGPT: Settings → Connectors → Advanced → Developer mode, then Add
custom connector pointing at https://your-host/mcp. Custom connectors need a
Pro, Team, Enterprise or Edu plan.
Publishing to the MCP Registry needs the package on PyPI first, then the
mcp-publisher CLI with server.json in this repo. Four things must line up or
the publish is rejected:
name must match the authenticated GitHub account: io.github.<username>/newsblog-composerdescription must be 100 characters or fewerpackages[].version must be a release that already exists on PyPImcp-name: <that same name> on its own line. The
registry proves package ownership by reading the description PyPI serves for
that exact version, so the marker has to be in the release you uploaded, not
just in the repo. It is the HTML comment at the top of this file.Three levels, cheapest first.
1. Offline, no network — proves the logic.
272 checks covering schema parity, the SEO audit, AI-word detection, publisher identity behind aggregator links, clustering, the publishing pack, the derived image concept, the trademark filter and the install paths. All should pass in about two seconds.
2. Network — proves search reaches you.
One line per provider with timings, then the verdict. What you want to see is
independent_publishers naming several real outlets and reference_candidates
holding real publisher URLs.
3. A whole post, end to end.
Builds a complete package from a real story using facts already in the file, and
prints the keywords, schema validation, AI-word check, human score and SEO score
before writing the output folder. Use it as the reference for what a good run
looks like. examples\build_mistral_example.py does the same for the
hand-written reference post.
4. In Claude Desktop, after restarting it. Say it the way you normally would — you should not have to name the tools:
Write me a blog post about ""
The server should answer by calling write_blog_post, then verify_news, and
work through to the finished package, ending with the post rendered in an
artifact and the publishing details printed in the chat.
If the client searches the web and writes its own markdown file instead, the extension did not trigger. Two things to check, in order:
Settings → Extensions shows it enabled. Note that closing the window on Windows only minimises to the tray — toggle the extension off and on to respawn the server.
Tell the client to prefer it. A host's own system prompt tends to treat
"write me a blog post" as "search the web and make an artifact", and an MCP
server's instructions cannot reliably outrank that. The fix is client-side:
put a line like this in a Claude project's instructions, or in your global
preferences, and work inside it —
When I ask for a blog post, an article, or a news write-up, always use the NewsBlog Composer tools. Start with
write_blog_post. Never search the web and draft the article yourself.
This is normal for local MCP servers; naming the server in the request works just as well for one-offs.
Extensions are desktop only. On the phone or the web app there is no extension to call, and the client will answer from web search without saying so.
The server runs on your own machine. Nothing is sent to the author of this package, and there is no telemetry of any kind.
What leaves your machine, and only while a tool is running:
GPTZERO_API_KEY or
SAPLING_API_KEY, and to an image generator only if you call generate_image.
Without those keys, no text leaves your machine for either purpose.What stays on your machine: the publishing identity you set at first run
(profile.json), every generated post (output/), and any keys you configure.
See Where it keeps your files for the exact paths.
Nothing in that folder is uploaded anywhere.
Optional API keys are read from the environment or from the extension's settings panel. They are used only to authenticate with the provider they belong to.
| Tool | What it does | Needs a key? |
|---|---|---|
capabilities | Reports which providers are live and which fallbacks are in use | no |
write_blog_post | The front door. Call it whenever someone asks for a blog post; returns the ordered plan | no |
find_stories | Turns a topic into today's actual stories, grouped and ranked by corroboration and freshness | no |
verify_news | Searches news providers, keeps matching results, counts independent publishers | no (keyless RSS) |
fetch_article_facts | Downloads sources, extracts facts, short attributed quotes and figures | never |
draft_brief | Returns the writing order: structure, sourced facts, keywords and FAQ candidates | never |
review_draft | Reads the finished draft and says where it is weak, before publishing | never |
humanize_text | Optional rewrite pass. Prefer review_draft | optional |
find_ai_words | Finds every stock AI phrase with the sentence it sits in | never |
score_ai_text | Local style score. Not measured unless a detector key is set | optional |
generate_image | Concept banner, trademark filter applied first | no (watermarked) |
seo_keywords | Mines keywords from the fetched sources; long-tail and FAQ queries from Google autocomplete | no |
seo_audit | Scores the finished body against on-page rules, returns fixes | never |
build_publishing_pack | Title, labels, permalink, alt text and a paste-ready Gemini image prompt | never |
build_schema | Renders HTML body + both JSON-LD blocks, then validates them | never |
save_and_present | Writes the package to output/, with canonical/OG/Twitter tags | never |
Resources: newsblog://house-style (structure and sourcing rules for the
drafting step) and newsblog://humanizer-rules (the rewrite rule set).
There are two entry points, depending on what you type.
A topic — "today's AI news", "electric vehicles", "Indian fintech". There is no claim to verify yet, so start by finding out what happened:
A specific headline — start at verify_news directly. Note that an old
headline correctly returns old sources; days limits how far back to look.
Output lands in a timestamped folder under output/:
| File | What it is |
|---|---|
publish-pack.md | Title, labels, custom permalink, search description, alt text, Gemini image prompt |
paste-into-blogger.html | Both JSON-LD blocks then the styled body — the file you paste |
report.md | Verification verdict, human score, AI-word status, SEO score, references |
index.html | Standalone preview with meta, canonical, OG and Twitter tags |
body.html, *.jsonld, meta.json | The pieces, separately |
Three guardrails are enforced in code, not left to the model:
verify_news returns is_legit: false unless at least two independent
publishers match the headline, or one primary/official source does.build_schema returns validation.issues listing every mismatch: an FAQ
question that differs between the HTML and the FAQPage schema, an image URL
that differs between the <img> tag and NewsArticle.image, a reference that
is not a real fetched URL. A non-empty list means don't publish.seo_audit returns must_fix for the things that actually cost rankings —
a duplicate H1, a missing keyword in the opening, images with no alt text, a
meta description of the wrong length, fewer than two external source links.Everything fetch_article_facts returns carries a source_url, so any claim in
the finished post can be traced back to the page it came from.
With an empty .env the pipeline still runs end to end. Here is what you get,
and what each key would change.
| Step | With no key | With a key |
|---|---|---|
| Search | GDELT DOC 2.0 — official, free, no signup, news-specific — then Bing/Google News RSS as backup | Tavily / Brave / Serper / Google CSE: cleaner snippets, higher limits |
| Article extraction | Full quality. trafilatura runs locally. | — no key exists |
| Keywords | Full quality. Mined from your fetched sources, plus keyless Google autocomplete. | — a paid keyword API would add search-volume data |
| Humanise | Returns the rule set and asks the calling model to rewrite. Works well in Claude; varies elsewhere. | Rewrite happens server-side, identical everywhere |
| Human score | Local heuristic, labelled is_real_detector: false | A real detector's score |
| Image | Pollinations anonymous: ~1 request/15s, and may watermark | Cloudflare/OpenAI/Stability: clean, fast |
| Schema, audit, files | Full quality. | — no key exists |
Brave is no longer the free recommendation. In February 2026 Brave removed its free tier and moved every plan to credit-based billing — a card is required, a $5 monthly credit covers roughly 1,000 requests, and you are billed past that.
Free options that still hold up, best first:
verify_news asks. Start here and only add a key
if snippet quality or freshness becomes a problem.TAVILY_API_KEY.GOOGLE_CSE_KEY and GOOGLE_CSE_ID from
programmablesearchengine.google.com.Free tiers move around; check each provider's own pricing page before committing.
generate_image runs with no key, but read this before publishing anything.
Pollinations still allows anonymous requests — about one every 15 seconds, basic models — but the free anonymous tier may watermark the image, which makes it unusable as a published banner. Three ways out, cheapest first:
POLLINATIONS_TOKEN.
Smallest change, keeps the existing provider.IMAGE_PROVIDER=cloudflare, CLOUDFLARE_ACCOUNT_ID and
CLOUDFLARE_API_TOKEN. Note the model returns 1024x1024, so crop to your
banner ratio.Whichever you use, the file lands locally. Upload it and pass the public https
URL into build_schema, or NewsArticle.image points at a path no crawler can
reach.
Add Cloudflare (or the free Pollinations token) for images. Search already works properly with no key; images are the step where the keyless output is not publishable.
TRADEMARKS in
src/newsblog_mcp/providers/imagegen.py as you hit new names, and still look
at what comes back.find_stories(days=1) is same-day; days=2 is
the default. verify_news searches a 14-day window unless you pass days.
Feeding it a three-week-old headline returns three-week-old sources, correctly.sources before trusting a medium confidence. The RSS backups are
unofficial endpoints that can change shape without warning; diagnose tells
you when one has stopped working.build_schema, or NewsArticle.image will point at a path no crawler
can reach.