annals

Annals for AI agents. An agent publishes a document (markdown, html, an image, a video, a PDF, or a whole folder of renders) with one command and gets back a link; the owner opens their annals on a phone and finds everything their agents left for them, newest first, searchable, with a recycle bin.

pip install annals
annals publish report.md --session my-agent
# https://apps.example.com/annals/d/20261003-203301-quarterly-report-3f9a

Beside what agents publish, the page can list other folders read in place, such as the sessions' ledgers (~/.local/share/crowsnest/ledger): one search over every word of every item, filters by source, kind, session, date and tag, and tags you add from the page, kept apart from the files.

Documents are plain files in a directory. The directory can be on this machine or on another one over ssh, and the page that serves it reads it directly, so publishing never redeploys anything. No database, no build step, no API token.

For agents

Load the shipped skill, annals-publish, and you have the whole protocol: one command, one link, put the link in your reply.

annals publish report.md                          # one document, prints its link
annals publish a.md b.md c.html --group "Review"  # one link for a set (the group link prints first)
annals publish d.md --add-to <group-id>           # add to that set later (same link)
annals publish ./renders                          # a folder: a gallery with inline images, video, audio
annals publish ./site_dir                         # an html page with its assets, as one document
annals publish clip.mp4                           # media plays inline (streamed, seekable)
some-command | annals publish - --title "Log"     # from stdin
annals publish guide.html --replace <link-or-id>  # a new version of a document: same link
annals publish page.html --artifact <claude.ai artifact link>  # plus a button to its interactive version
annals ls "search words"                          # what is there, newest first
annals trash <id>  /  annals restore <id>         # the recycle bin

What each kind looks like on the page:

A folder skips hidden files and caches (.*, __pycache__, *.pyc); --all-files keeps them.

Interactive version (--artifact)

annals publish page.html --artifact https://claude.ai/... stores the link of a Claude Artifact holding the same page in the document's meta (artifact), and the document's page shows an "Open the interactive version" button next to the frame. Comments, per-viewer and shared state and sharing then live on the artifact while annals stays the single in-tray. Only plain https links to claude.ai are accepted (annals.store.ARTIFACT_HOSTS; no port, credentials or odd characters), --artifact takes one path, and a --replace without --artifact keeps the current link.

Living pages (--replace)

annals publish <path> --replace <link-or-id> publishes a new version of an existing document: the id and the link stay, the page shows the new version marked "updated" and lists the earlier ones, each readable at /d/<id>@<n>. The previous files are kept under the document's .versions/<n>/; nothing is deleted. It takes one path (a file, a folder, stdin), and a replaced group member stays in its group.

A replacement is written next to the document as a pending directory (docs/<id>.r<stamp>-<hex>/), never over it, and applied by whoever reads the store (the server, or the next publish): a few renames, the first of which claims the document, so concurrent readers never apply one twice or out of order. That is what lets a replace travel through an outbox, a directory another machine pulls and empties: mark it with annals configure --outbox (or ANNALS_OUTBOX=1), and replacements published there are left for the server to apply when they land. HTML pages without a doctype are served with <!doctype html> (plus a charset and viewport when missing), so they render in standards mode.

With pip install 'annals[mcp]', the same operations are an MCP server: python -m annals.mcp.

Where documents go

annals configure --target tw:/root/.local/share/annals --base-url https://apps.example.com/annals

That writes ~/.config/annals/config.toml. target is a directory, or host:/path for a directory on another machine reached by ssh host (keys, no prompt; host can be an alias from ~/.ssh/config). ANNALS_TARGET and ANNALS_BASE_URL override it per shell. With no configuration, documents go to ~/.local/share/annals and the link points at a local annals serve.

A second place (--to)

A config file can name more places: annals configure --to team --target host:/path --base-url https://team.example.com/annals saves a [targets.team] table and leaves the default alone. Then annals publish r.md --to team (also ls, show, trash, restore, group, groups) works on that place only, and annals copy <link-or-id> --to team [--source other] copies documents or whole groups there, keeping their ids. Nothing about the place is in the code: a name, a target and a base url.

Hosting the page

Inside an app platform (the usual case). Mount the API under the platform's prefix and let the platform's login gate it:

# server.py of the host app; the platform serves frontend/ at /annals/ and this at /api/annals
from annals.api import mk_api
app = mk_api(data_dir="/root/.local/share/annals")

and write the page shell once as the app's frontend:

annals page-shell --api /api/annals --base /annals > frontend/index.html

The shell only names those two paths; the page's script and style load from the API, so upgrading the package upgrades the page. Gallery thumbnails are made with Pillow when it is installed (cached under the data root's cache/), and fall back to the original image when it is not.

Standalone, for a machine with no platform:

pip install 'annals[server]'
ANNALS_DATA_DIR=~/.local/share/annals annals serve --host 127.0.0.1 --port 8765

The page is at http://127.0.0.1:8765/annals/, its API under /annals/api/. Auth is chosen by environment variables:

Spoken overviews

A document's page can offer a Listen button: a few minutes of speech about the document, played with its script in sync. It is off unless the host turns it on, because making one spends money.

app = mk_api(overviews=True)  # the button, and POST /docs/{id}/overview

A worker on the same machine fulfils the requests. It runs a command you choose, which takes a text file and writes one HTML page, and publishes that page as a document of its own:

export ANNALS_OVERVIEW_COMMAND='hailing overview {source} --minutes {minutes} --no-publish --player-to {out}'
annals listen-work                 # fulfils what is queued, one at a time
annals listen <doc-id> --minutes 3 # asks for one from the command line, as the button does

Run annals listen-work from a timer, not from a watcher on the folder: finished request files stay there, so a watcher would keep waking it. A current overview is not made again unless forced, and at most 20 are started in any 24 hours (ANNALS_OVERVIEW_DAILY_LIMIT). annals.overview has the details.

Sharing one page

The annals is private; a share link lets someone without an account read ONE document (or one group) and nothing else. From the page, the Share button on a document lists its files and the annals pages it links to, so you choose what goes; from a shell:

annals share <id-or-link> --password -                       # 7 days by default; password from stdin
annals share <id> --expires 2026-11-30                       # or 12h, 30d, an ISO time, `never`
annals share <id> --include img/chart.png,img/map.png        # only these files (the page always goes)
annals shares                                                # live links (--all: revoked and expired too)
annals unshare <link-or-token>                               # off at once

The link is <share_base_url>/<token>/: 128 random bits, noindex, no-referrer. Every request is checked against the share's record (shares/<token>.json under the data root), so a file you did not choose is a 404 even beside the shared page, a revoke or an expiry takes effect on the next request, and references to anything outside the share (an image you left out, a link to another annals page or to anything else on your annals' host, an annals id in the text) are removed from every text file before it is sent (markdown, html, plain text, svg). A linked page you tick goes with the files it shows, not every file beside it. The password is a gate on the link, not an account: at least 8 characters, stored hashed (stdlib scrypt), with attempts limited per client and per link, and a correct one earns a cookie scoped to that link. Links are read-only and last 7 days unless you say otherwise.

Where the links are served:

• Standalone (annals serve): at /annals/shared/<token>/, outside the login, with nothing to configure.
• Inside an app platform: the share server must be reachable without the platform's login, so run it as its own public app and tell annals where it is:

# server.py of a second, PUBLIC app (the owner's app keeps its login)
from annals.share_api import mk_share_api
app = mk_share_api(data_dir="/root/.local/share/annals")

then set ANNALS_SHARE_BASE_URL (or annals configure --share-base-url ..., or mk_api(share_base_url=...) for the owner's page) to that app's public URL. Until it is set, shares are made but carry no link. mk_share_api takes rate_limiter= and link_failure_limiter= (anything with allow(key) -> bool, such as heed's Redis limiter; the defaults count per process), trusted_proxies= (whose X-Forwarded-For names the client; loopback by default) and private_hosts=; the password hasher is a keyword of annals.share.Shares (argon2, or anything with hash/verify).

Downloads

Every markdown page has an Actions menu (Share, Copy, Download, View) with two downloads made by the server: Markdown (.md) and a titled PDF with the images inside (audio, video and other files appear as a link to the resource). Both are named after the title (quarterly-report-q3.pdf) and are served under the same rules as the page: the owner's login, or a share link's scope (a recipient gets the filtered text and only the files the link includes). The PDF needs pip install "annals[pdf]" (WeasyPrint, which uses Pango; annals[server] includes it); without it the PDF action answers with that hint. Pages (/d/<id>) are drawn in the browser by app.js, so a page that gains an action gains it for every existing document.

One Markdown file from a group

A group's page has a Combine button: it makes one Markdown file from every .md in the group, folder documents included, and offers Copy Markdown, Download .md and Open as text, with the text shown below for selecting by hand where a phone refuses to copy. The file has the group's title, a table of contents, and one ## section per file named after its path (Report/01-intro.md); paths that start with a number come first, in numeric order, the rest in the group's order; each file's headings go one level down and hard-wrapped lines are joined. Footnotes are renamed per file so they do not collide, and a code block a file leaves open is closed at the end of its section. Relative links and images keep their paths, so they do not resolve inside the combined file. It is made on each request and never stored. A share link to a group has the same button, built from the share's filtered text and only the files it includes.

annals aggregate <group-id-or-link> > all.md          # the same file, from a shell
annals aggregate <id> --order given --demote 0 --no-toc --keep-wraps

from annals.aggregate import aggregate_markdown
aggregate_markdown({"notes.md": "# Notes", "01-intro.md": "Hello"}, title="Bundle", demote=1)

aggregate_markdown takes (path, text) pairs and knows nothing of annals: order= (numbered_first by default, None for the order given, or any paths -> paths callable), demote=, unwrap=, toc=. The API is GET /groups/<gid>/aggregate (?download=1 to save it, &demote=), and GET /<token>/aggregate on the share server.

A group's documents as separate notebook sources

To take a group into a notebook (NotebookLM, now Gemini Notebook) with each document as its own source, open the group and tap Notebook. Tick the documents (all are ticked), choose how long the links work (1 day by default), and tap Make links. You get one link per page; tap Copy N links, then in the notebook choose Add source → Website and paste: the notebook makes one source of each link, named after the document. The same panel's Combine these makes one Markdown file of only the ticked documents.

The links are an ordinary share of just those documents (no password, since the notebook cannot type one): unguessable, filtered like every shared text, and revocable with Turn these links off or under Shares. The notebook keeps its own copy once a link is added, so turning the links off afterwards does not remove the sources. Each link is a plain web page with no script (<share link>/source/<doc id>/<file>); markdown is rendered on the server, images, audio and PDFs get no link (a notebook takes those only as uploads). Why links and not a zip or a download: misc/docs/notebook-sources.md.

The API: GET /shares/<token>/sources (the owner's) or GET <share link>/sources (the recipient's) lists {title, url} per page; GET /groups/<gid>/aggregate?doc=<id>&doc=<id> combines only those documents.

Other folders, search, filters and tags

Name extra folders when you mount the API, and they appear beside the published documents, each in its own section (/annals/s/<name>) and in /annals/all:

app = mk_api(data_dir="~/.local/share/annals", sources={"ledgers": "~/.local/share/crowsnest/ledger"})

or, with no code, ANNALS_SOURCES="ledgers=~/.local/share/crowsnest/ledger" (several: comma-separated). A folder is read in place and never written: its files are listed newest first with their age, markdown renders, html is framed, the rest shows by its kind, and a sub-folder's files count as written by the sub-folder's name. Hidden files and caches are skipped.

On every list:

• Search matches every word, case-insensitively, against titles, tags, sessions, excerpts and the full text of markdown and text files (on the server; the matching passage shows under the title).
• Filters are one button per family (source, kind, session, date, tag). Checked values within a family are alternatives ("any of"), families combine ("and"), the counts are what each choice would give now, and the active filters show as removable chips. The filters and the search are in the URL, so Back undoes a filter and a link reproduces the view.
• Tags: add or remove them on any item's page; they save as you go. They are stored apart from the items, in one file under the data root (tags/tags.json), so a ledger is never touched; on a published document they replace the tags the publisher gave.

The API behind it: GET /items?q=&source=&trash= → {data, total}, GET /items/<source>:<id>, PATCH /items/<source>:<id> with {"tags": [...]}, GET /tags, GET /sources, and GET /files/<source>/<path> for a folder's raw files. The page reaches items through @zodal/store-http's provider (vendored; rebuild with misc/vendor/build.sh).

Python API

from annals import DocStore, publish
meta = DocStore("~/.local/share/annals").publish("report.md", tags=["q3"])
publish(["report.md"], tags="q3")["url"]          # same, through the configured target

annals.api.mk_api(...) is the mountable API; annals.api.mk_app(data_dir=..., authorizer=...) the standalone app. Both take sources= (name → folder path, or any annals.sources.Source) and tags= (any MutableMapping from item key to tag list; default JsonTagStore). annals.catalog.Catalog is the facade they stand on: every source's rows as one list, searched, with the owner's tags.

Design notes

Flat store with tags and groups, not a hierarchy: agents from many projects do not share one, and every placement decision costs tokens. A document is docs/<id>/meta.json plus its files; the bin is trash/<id>/; a group is groups/<gid>.json. Ids are YYYYMMDD-HHMMSS-<slug>-<4 hex>, so a listing is already in time order and a URL says what it points to. The transport is a seam (annals.target): local directory, ssh, and later an http ingest endpoint. More in misc/docs/design.md.

Skills

gh skill install thorwhalen/annals annals-publish

The skill also ships inside the package at annals/data/skills/annals-publish/.