Newskit

English | Français

Automated topic monitoring for Claude Code. You name a subject and give a few sources; at regular intervals Newskit collects them, writes one note per item, extracts the concepts and how they relate, and keeps a knowledge graph up to date, within a spending cap. Two things set it apart from a feed reader: every answer drawn from the graph cites its note and its line, and the graph extends itself to new themes and new sources.

For developers who already use Claude Code: every model call goes through claude -p, under your own Claude subscription. Notes, feed and newsletter are written in French; an English translation can be added (traductions: [en]).

governance.news: a newsletter and a knowledge graph on AI governance in business, built on Newskit.

What you get

All of it lives in your media folder, as plain files:

• Notes: one Markdown file per collected item, under sources/, with its title, author, date and address.
• The graph: graphe/<name>/graphe.json, its map (carte.md, the main themes and their concepts) and the report of each pass (ROUTINE.md).
• Answers with their source: newskit query searches the graph without any model call and gives, for each answer, the note and the line that support it.
• A feed and a newsletter: short posts on the latest items (fil.json) and a newsletter of the graph, sent through your own delivery module (diffusion: in config.yaml).

$ newskit query --graph ai-governance sources "general-purpose AI model"
  Sources attestant « General-purpose AI models »
    sources/ai_compliance/rss/ec_digital/2026-08-02_navigating-ai-act.md L211

Quick start

You need Python 3.12 or later, git, macOS or Linux, and a logged-in Claude Code (claude).

The words used below: the media is your project folder; it holds one or more graphs; each graph follows subjects; a pass collects the graph's sources and writes one note per item.

1. Install the engine and test it in an empty folder:

   pip install newskit
   mkdir my-media && cd my-media
   newskit test
   

   The test runs one full pass on a sample graph (medieval beekeeping), with model answers replayed from a recording: no paid call, about one minute. It must end with « témoin vert » (test passed).

2. Open Claude Code in that folder (claude) and install the plugin:

   /plugin marketplace add lfoletti/newskit
   /plugin install newskit@newskit
   

3. Ask Claude to follow a subject: “start a monitoring graph on urban gardening”. The newskit skill asks its questions, writes config.yaml and sources.yaml, suggests themes drawn from the sources you give, creates the graph and runs a first pass in front of you.

4. Read the result: graphe/<name>/ROUTINE.md for the report, newskit query --graph <name> search "…" for the graph.

5. Schedule the next passes. newskit new-graph has written the graph's task under taches/ and printed the two sudo lines that install it in launchd (macOS). On Linux, add a cron line at each time of passages that runs newskit test then newskit run --graph <name> --scheduled; on a day outside the interval, the scheduled run stops by itself.

Sources

Sources are declared in sources.yaml, next to config.yaml, one block per source, each tied to a graph. Supported: RSS feeds, sitemaps and listing pages (for sites without a feed), Substack, Medium, LinkedIn newsletters, arXiv, journals by ISSN, OpenAlex, GitHub and Hugging Face searches, an e-mail inbox, and local folders of Markdown or PDF files.

version: 1
sources:
  - cle: city_farmer
    plateforme: Flux RSS
    nom: City Farmer News
    dossier: potagers/rss
    abonnement: gratuit
    graphe: potagers-urbains
    flux: https://cityfarmer.info/feed/
  - cle: soil_papers
    plateforme: OpenAlex
    nom: OpenAlex, urban soil
    dossier: potagers/openalex
    abonnement: gratuit
    graphe: potagers-urbains
    requetes: ["urban allotment soil contamination"]

The skill writes this file for you from the addresses you give; a source the engine cannot read is refused with its reason.

How the graph grows

The graph does not stay within the subjects you declared at the start: it expands.

• New themes. After each pass, the engine looks for clusters of concepts that none of the declared subjects covers. A candidate that comes back three passes in a row joins the subjects; the sources and searches follow it.
• Auto-completion. A question the graph could not answer becomes a search for the next pass. A web scout searches the declared subjects; a question agent probes the graph the way a practitioner would; what a language model considers essential and the graph lacks becomes a shopping list of sources to find.
• The source list. The engine assesses a candidate domain before it is declared, and keeps track of every source a user drops in (a link sent from the browser). A source that keeps failing is audited and, when a route exists, repaired on its own (see accesspath below).
• Clean concepts. The same concept found under several labels (a translation, an acronym, a variant) is merged after a model confirms it is exactly the same concept; refusals are kept and never asked again.

The media folder

media_sample/ is a complete example to copy. The media is a folder that holds a config.yaml. The engine attaches to it: the NEWSKIT_DONNEES variable first, otherwise the current folder. The code stays in the installed package; the data (notes, graph, cost ledger) live in the media folder.

version: 1
media:
  nom: potagers               # name: lowercase, digits and hyphens
  langue: fr                  # language of the prompts and of the publication
  traductions: [en]           # translated feed and newsletter; none by default
  passages: {tous_les: 1, heures: ["05:30"]}   # every N days, at these times
  # service_maj: kg-maj       # the server service started over ssh after a pass; none by default
  # cerveau_du_site: media-site    # the Thalamus brain of the site graph; "<media>/<graph>" by default
  # cerveau_lacunes: media-lacunes # the Thalamus brain of reported gaps; "lacunes-signalees" by default
modeles:                      # one model per role; a missing role keeps its default
  extraction: claude-opus-5-5
  redaction: claude-sonnet-5-5
sources: sources.yaml         # the source list, kept apart
# registre_des_graphes: graphes.json   # labels and declared subjects, kept by the engine (default)
graphes:
  potagers-urbains:
    libelle: Potagers urbains   # label
    domaine: "urban gardening: allotments, soil, seed exchange"   # the domain, in English, for extraction

newskit media shows what was read; an unknown field is refused and named. The roles: extraction (reading a note and drawing its concepts), analyse (judging a whole graph), redaction (summaries and newsletter), lecture (deciding whether two labels are the same concept), traduction (translation), questions (training questions for the local encoder).

Cost

Every model call goes through your subscription and counts against its usage limits; the cost ledger records each one at its API-equivalent price, so you can see where the usage goes (newskit costs --by graph). Concept extraction is the main item: measured on governance.news, about $0.25 per note with the default model (Opus), so about $10 per pass at the default cap of 40 notes. The cap is set per graph (--cap of newskit new-graph).

Commands

Safeguards

• The test. Before the real pass, the engine runs newskit test on the sample graph with recorded answers. The real pass starts only if a test passed the same day with the same engine version. An engine update therefore cannot break a pass unnoticed: it stops before spending anything.
• The caps. Each step has its own cap on notes and on spending; every model call is written to the cost ledger, under a step and a graph.
• The usage limit. After three usage limits in a row, the engine sets a marker that lasts three hours. The plugin's hook then refuses any model call started from Claude Code.

Works with

Newskit relies on:

• graphify, installed with the package. Reads the structure of the documents, validates each extraction against the schema, builds the graph and splits it into communities.
• trafilatura, installed with the package. Main text extraction from web pages.
• pytest, installed with the package: newskit test runs on it.

The following tools extend its features when they are installed:

• accesspath (sap). Website access audit. When a source keeps failing, Newskit asks sap which route still reaches it (RSS feed, API, sitemap, plain HTML, browser) and repairs the source.
• cortexdb. AI memory and knowledge graph in one SQLite file. A graph can be loaded into a cortexdb brain, each relation tied to the passage that supports it, and queried from any MCP client.
• thalamus. A governance layer over cortexdb for an organisation's shared knowledge. Questions left unanswered by its clients return to the engine as searches.

Limits of version 0.1

• The prompts to the models are written in French; the only translation is English.
• macOS and Linux. Scheduled passes run through launchd on macOS; on Linux, through a cron line.
• newskit new-graph writes the graph's scheduled task under taches/, ready to install in /Library/LaunchDaemons.

Contributing

See CONTRIBUTING.md. Report a vulnerability privately: SECURITY.md.

Licence

MIT.