myrssfeed

Read RSS/Atom feeds, and print only the items that match criteria you describe in plain English. Claude makes the call on each item; you get the headline, the link, and a one-line reason.

$ myrssfeed --prompt "Export controls on AI chips, and EU tech regulation" \
         --feed https://feeds.arstechnica.com/arstechnica/technology-lab \
         --feed https://www.theregister.com/headlines.atom
US tightens accelerator export rules for third-country resellers
  https://www.theregister.com/2026/10/06/export_rules/
  The Register · 2026-10-06T09:12:00+00:00
  why: Directly about chip export controls.

EU Commission opens consultation on cloud switching rules
  https://arstechnica.com/2026/10/06/eu-cloud/
  Ars Technica · 2026-10-06T07:40:00+00:00
  why: EU tech regulation affecting cloud providers.

Everything else in those feeds is dropped silently.

Why

Feed readers filter on keywords, which is both too narrow ("accelerator" misses "GPU") and too broad (every mention of "chips" matches). A prompt describes what you actually want: "new CVEs in Python packages I'd have to patch, not general security commentary."

Install

myrssfeed is not on PyPI yet, so install it from a checkout:

git clone https://github.com/fabianhogger/myrssfeed
cd myrssfeed
pip install .

For development, install it editable with the test and lint tools:

pip install -e ".[dev]"

Once the project is published (see Releasing), the usual one-liner will work:

pip install myrssfeed     # not available until the first release is published

Requires Python 3.9+ and an Anthropic API key:

export ANTHROPIC_API_KEY=sk-ant-...

Check that it landed:

myrssfeed --version

If the myrssfeed command is not found, your Python scripts directory is not on PATH; python -m myrssfeed works identically.

Usage

One-off run

myrssfeed --prompt "Rust async ecosystem news" --feed https://blog.rust-lang.org/feed.xml

With a config file

myrssfeed.toml:

system_prompt = """
Keep items about:
  - security advisories affecting Python or Node packages
  - outages at major cloud providers

Skip funding rounds, opinion pieces, and conference announcements.
"""

feeds = [
  "https://feeds.feedburner.com/TheHackersNews",
  "https://status.cloud.google.com/en/feed.atom",
]

interval = "30m"
state_path = "~/.local/state/myrssfeed/state.json"
output_format = "markdown"

myrssfeed --config myrssfeed.toml

JSON config files work the same way; the format is chosen by file extension.

Scheduling

Two options, depending on whether you want a resident process.

Built-in loop — runs until you stop it, and shuts down cleanly on Ctrl-C/SIGTERM so the state file is never left half-written:

myrssfeed --config myrssfeed.toml --interval 45m

cron or systemd — have myrssfeed run one pass and exit. The state file means each run only reports items it has not reported before:

*/30 * * * * ANTHROPIC_API_KEY=sk-ant-... /usr/local/bin/myrssfeed --config ~/myrssfeed.toml --once >> ~/myrssfeed.log 2>&1

A systemd timer unit is in examples/.

Checking a setup before spending anything

--dry-run fetches the feeds and lists what would be judged, without calling the API, and without marking anything as seen:

myrssfeed --config myrssfeed.toml --dry-run -v

Output formats

Logs go to stderr and results to stdout, so myrssfeed ... | jq works as expected.

How it works

1. Each feed is fetched with If-None-Match/If-Modified-Since, so an unchanged feed costs one small request and no tokens.
2. Entries already handled on an earlier run are dropped using the state file.
3. The rest are sent to Claude in batches, with your prompt as the system prompt and a JSON schema constraining the reply to one verdict per item.
4. Matching items are printed; every item that got a verdict is recorded as seen.

Items in a batch that fails (a transport error, a declined request) are not recorded, so the next run retries them rather than dropping them silently.

Cost

Only titles, summaries and dates are sent — never full articles — and summaries are truncated. The system prompt is cached across the batches of a run.

The knobs that matter:

• --batch-size (default 10): larger batches spread the prompt over more items.
• --effort (default low): relevance is a classification task and rarely needs more. Raise it if the criteria are subtle.
• --no-reasons: drops the per-item justification, and the output tokens with it.
• --max-items N: caps how many entries per feed are considered at all.

Configuration reference

Every key below can be set in a config file, as an environment variable MYRSSFEED_<KEY> (uppercase), or as a command-line flag. Precedence is file < environment < flags.

--api-key also exists, but prefer the environment variable: flags are visible in ps output and end up in shell history.

Exit status

Library use

The CLI is a thin wrapper; the same thing works from Python:

from myrssfeed import Config, Runner

config = Config(
    system_prompt="Release notes for Python web frameworks.",
    feeds=["https://blog.djangoproject.com/rss/"],
    state_path="~/.local/state/myrssfeed/state.json",
)

result = Runner(config).run_once()
for match in result.matches:
    print(match.item.title, match.item.link, "-", match.verdict.reason)

Runner accepts a reader, a relevance_filter and a state if you want to substitute your own — which is how the test suite avoids the network entirely.

Writing a good prompt

The prompt is the whole filter, so be concrete about both halves:

Keep:
  - Postgres release announcements and migration-breaking changes
  - Write-ups of real production incidents with root-cause detail

Skip:
  - Vendor benchmarks and marketing posts
  - "Top 10 database tips" listicles
  - Anything about MySQL unless it compares directly to Postgres

The model is told to prefer precision over recall, so tangential items are dropped. If you are missing things you wanted, say so in the prompt explicitly ("err on the side of including borderline items") or raise --effort.

Development

pip install -e ".[dev]"
pytest                 # no network or API calls
ruff check . && ruff format --check .
mypy

The test suite never reaches the network or the Anthropic API: feeds are parsed from in-memory XML and the classifier is a scripted fake, so it runs offline and without an API key.

Releasing

Publishing uses trusted publishing, so no API token is stored in this repository. One-time setup, done once per index at https://pypi.org/manage/account/publishing/ and the same page on https://test.pypi.org:

Then, per release:

# 1. Rehearse on TestPyPI, and check the result installs
gh workflow run release.yml
pip install --index-url https://test.pypi.org/simple/ \
            --extra-index-url https://pypi.org/simple/ myrssfeed

# 2. Bump __version__ in src/myrssfeed/__init__.py, move the CHANGELOG entry out
#    of [Unreleased] into a version heading, then tag and push
git commit -am "Release 0.1.0"
git tag -a v0.1.0 -m "0.1.0"
git push origin main v0.1.0

The tag push is what publishes to the real PyPI. The workflow refuses to upload if the tag does not match __version__, because PyPI never allows a version number to be reused — a mistake means yanking the release and shipping a new number, so it is checked before the upload rather than after.

To inspect a build locally first:

pip install build twine
python -m build
twine check --strict dist/*

License

MIT.