satchel

A tiny local archive of the actual things you read.

Most "read later" tools save a URL. The link rots, the site adds a paywall, or you're offline — and the thing you saved is gone. satchel fetches the page once, extracts the actual article text (not the nav, not the related-links box, not the footer), and stores it in one local SQLite file with real full-text search. No server, no account, no network needed to search what you've already saved.

$ satchel add https://example.com/some-article
saved #4: Why Local-First Software Is Worth the Extra Effort

$ satchel search "FTS5"
#4    Why Local-First Software Is Worth the Extra Effort  (https://example.com/some-article)

$ satchel read 4
Why Local-First Software Is Worth the Extra Effort
by Jordan Rivers

Most reading tools save a link and call it done...

Save with one click instead of a terminal: satchel serve runs a local listener and prints a bookmarklet — click it on any page and that page is saved through the exact same pipeline add uses. See "Capture" below.

How

• fetch.py — plain urllib GET. Stdlib, no dependency for the easy part. Always rejects non-http(s) schemes; a URL that didn't come from the person running satchel (see "Capture") is additionally checked against loopback/private/link-local address space before and after every redirect hop, not just the first one.
• capture.py — fetch → extract → store, in one place. Both add and the capture listener call this; there is exactly one add pipeline.
• extract.py — wraps trafilatura to pull title/author/main-text out of real HTML, correctly skipping navigation, ads, and related-links boilerplate. Boilerplate-stripping is exactly the kind of thing that looks fine on a hand-rolled test page and breaks on the next real site's markup — not worth reimplementing.
• db.py — one SQLite file, an FTS5 virtual table kept in sync with the real table via triggers (external-content FTS5: the searchable index isn't a second copy of the truth that can drift from the first).
• serve.py — the local capture listener (below).

Install

pip install satchel-reader        # the command it installs is `satchel`

Or from a checkout, for development:

pip install -e .

Use

satchel add <url>          # fetch, extract, save
satchel list                # everything saved
satchel search <query>      # full-text search
satchel read <id>           # print an article's full text
satchel serve                # one-click capture -- see below

All commands take --db path/to/file.db. The default, if you don't pass one, is ~/.local/share/satchel/satchel.db (respecting XDG_DATA_HOME) — one stable location regardless of which directory you happen to run the command from, not ./satchel.db in the current directory. Run satchel --help to see the exact resolved path on your machine.

Saving the same article twice — via a tracking link, a shortener, or a plain http:// URL the site itself upgrades to https:// — is one duplicate, not two. add normalizes the URL it's given, follows redirects, and normalizes the actual address the server served before checking for a duplicate: it defers to what the server says, rather than guessing at a scheme policy.

Capture

$ satchel serve
satchel capture listening on http://127.0.0.1:8765  (db: ~/.local/share/satchel/satchel.db)

Drag this to your bookmarks bar, then click it on any page to save it:

  javascript:fetch('http://127.0.0.1:8765/add?token=...&url='+encodeURIComponent(location.href))...

^C to stop.

Drag the printed link to your bookmarks bar. Click it on any page while serve is running, and that page goes through the same fetch/extract/save pipeline as satchel add — no terminal required for the thing you do most often.

This changes the threat model, and it's handled, not ignored. While serve is running, any tab open in your browser can send it a request — a browser's same-origin policy stops a page from reading a cross-origin response it wasn't granted, but not from sending the request in the first place. Without a shared secret, that would be an open invitation for any open tab to make satchel fetch an arbitrary URL. So: the listener only binds to 127.0.0.1, a fresh token is generated every time you run serve and only ever appears in the bookmarklet you just dragged, and every captured URL is checked against loopback, private (RFC1918), link-local (this is what closes off 169.254.169.254-style cloud metadata endpoints), reserved, and multicast address space — before the first request, and again on every redirect hop, since checking only the first hop would let a URL redirect somewhere private after passing the initial check.

What that guard does not claim: it validates a hostname's DNS answer at the moment it's checked, it doesn't pin the connection to that exact resolved address. A DNS-rebinding attacker with a fast-expiring record could in principle still slip a private address past the check and have the actual connection re-resolve to it. Closing that fully means connecting to a pinned IP rather than by hostname — real added complexity for a personal tool's local listener. This is a deliberate, documented partial mitigation, not a claim that it's unbreakable.

Direct satchel add <url> from your own terminal is not restricted this way — typing your own local dev server's URL to save a draft you're writing is a legitimate thing to do, and you are not a threat to yourself.

Test

python tests/test_satchel.py

Extraction is tested against a fixture HTML file (tests/fixtures/), not the network — a real bug was caught this way during development: an ambiguous byline ("By Jordan Rivers · September 2026" in one text node) made trafilatura fold part of the date into the author field. Fixed by making the fixture look like well-structured real markup (byline and date as separate elements) rather than loosening the assertion.

The capture listener is tested the same way it's actually used: a real http.server.HTTPServer runs in a background thread and gets real HTTP requests, including SSRF attempts against loopback and link-local addresses — checked with the correct token, since the interesting question is whether the guard holds once someone's past the door, not whether the door itself works.

What's deliberately not here yet

No tagging, no folders, no read/unread state — a flat list plus full-text search covers the actual workflow (encounter → capture → search → read); add these when a flat list genuinely stops being enough, not before. No multi-device sync — one local file is the whole pitch, and sync is a separate, harder problem this project isn't trying to solve. No AI summarization, no embeddings, no recommendation engine: this is an archive of what you actually read, not a platform.

MIT licensed.