astro/daylight.py
How light it is outside right now at a latitude and longitude, as one number from 0 (night) to 1 (the sun overhead): the sine of the sun's altitude, computed with ephem (no network). A body can grow by it.
70 tools in 25 packs. Each is one file that runs in a pinned container via fragletc; each pack is one Agent Skill. Agents start at llms.txt.
Written in any language with a fraglet-enabled container: the 90+ 100hellos languages or the purpose-built ofthemachine images (python3, headless-browser, latex, meme, home-automation, 3d-printing…). Two host dependencies: fragletc and Docker.
Claude Code, from a clone: make build && claude plugin marketplace add "$PWD/catalog", then claude plugin install <pack>@ofthemachine-tools.
curl -fsSLO https://tools.ofthemachine.com/<pack>/<tool>How light it is outside right now at a latitude and longitude, as one number from 0 (night) to 1 (the sun overhead): the sine of the sun's altitude, computed with ephem (no network). A body can grow by it.
The Moon's phase on a given UTC date, computed with ephem (no network): phase name, percent illuminated, and the dates of the next new and full moon. One line, ready for a dateline or an almanac box. The date is a parameter, never the clock, so the same inputs always give the same answer.
Sunrise, sunset, and day length for a latitude, longitude, and UTC date, computed with ephem (no network). Times are printed in the given UTC offset so a caller can pass the local zone; the day-length delta against the day before says whether days are getting longer or shorter. Companion to net/ip-location.py, whose lat,lon this takes.
Compile a catalog of level-0 tools -- a tarball of pack directories, each an index.md plus fraglet files -- into its projections: one Agent Skill per pack (SKILL.md beside byte-identical copies of the tools), a Claude Code plugin per pack with a marketplace.json, llms.txt and llms-full.txt for agents arriving over HTTP, manifest.json with every tool's contract and procedure_hash, and an Open Knowledge Format bundle. Everything is derived from the tools' own headers and the pack index files; nothing is authored twice.
Render a compiled tools catalog (archived) into a static website: the catalog itself served verbatim -- llms.txt, manifest.json, every pack's SKILL.md and every tool file at the same paths -- plus an HTML layer over it: an explorer of every tool with search, tag and pack filters, grid and table views and a dark/light theme; one page per pack; one page per tool with its contract, provenance, a curl one-liner and its full source. Relative links throughout, so the result works from a file:// URL as well as a web root.
Validate a compiled tools catalog (archived): every pack's SKILL.md against the agentskills.io specification using skills-ref, the spec's own reference validator; every tool file present, executable, and hashing to the procedure_hash manifest.json records for it; every llms.txt link resolving; marketplace.json plugins matching the pack directories; and, if an okf/ bundle is present, every concept typed and every relative link resolving. Prints one line per finding and a summary; exit 0 only when clean.
Fetch the catalog tools a project depends on, verify each against its pinned procedure_hash, and pack them into vendor.tar.gz (executable, at their catalog paths, reproducible bytes). The lock file names where the tools are published ("source https://tools.ofthemachine.com") and lists one "<pack>/<stem>.<ext> sha256:<hex>" per line (# comments allowed); the hashes are what manifest.json and every receipt call procedure_hash. A file whose hash does not match is refused and nothing is written, so base_url may fetch from anywhere -- a mirror, or a local `make serve` -- without weakening the pin or changing the recorded source.
Generate standalone SVG vector charts (bar, line, scatter, pie) from CSV or JSON data.
How fast a Cloudflare site's origin answers, path by path: requests, p50/p95/p99 and size over the last 24 hours, against the last hour and the 24 hours before, flagging a real slowdown. An address making most of a window's requests is left out of it and named, so a flood cannot skew the numbers. Long polls are labelled and never flagged: their time is waiting. Read-only.
A Cloudflare zone's traffic over the last 24 minutes, 24 hours and 24 days, drawn and diagnosed: traffic.svg stacks the three as small multiples (requests from clients by status class, the daily Worker quota and today's projection on the days), worker.svg is the Worker's CPU time per hour against the free plan's 10 ms limit, and report.md is a Markdown report (a headline, the golden signals against the 24 hours before, the top talkers by kind of client and path, and diagnostics that name a poller, the side a 5xx came from, CPU headroom and the quota's pace). Clients are counted by kind (a browser by name, else a program). Read-only; cloudflare/usage is the quota alone, as gauges. It names the top client addresses, with country and client.
A zone's Worker requests today (the UTC day Cloudflare's daily quota counts) against that quota: how many so far, where the day ends if the last hour's rate holds, and the share of the quota that is, by Worker script; then the last hour's requests to the zone by path and by kind of client (a browser by name, else a program), never by address. The gauges are files (used.txt, projected.txt, share.txt, the projected share as a fraction) and stdout is a short report. Read-only: the token needs only Analytics read and Zone read.
Extract compact hierarchical polyglot symbol outlines (classes, interfaces, functions, methods, structs) from source code across 140+ languages using Universal Ctags, omitting implementation bodies for up to 95% token savings.
Slice, transform, filter, and extract JSON data using jq query expressions on JSON text or JSON files.
Run SQL queries on CSV, TSV, or tabular data, outputting results as Markdown, JSON, or CSV.
Render Graphviz DOT source into an SVG, PNG, PDF, or JSON diagram. Choose a layout engine with engine=dot for hierarchies, neato or fdp for spring layouts, circo for cyclic structures, twopi for radial trees, or sfdp for large graphs. Source comes from the dot parameter, a dot_source file, or stdin, so a model can write the graph and this tool renders it deterministically.
Extract readable text out of PDF, DOCX, PPTX, XLSX, ODT, RTF, HTML, and plain-text documents into structured Markdown carrying page, slide, and sheet markers. The format is detected from file content, so no extension is needed; pages selects a 1-indexed subset such as "1-5" or "2,4,7-9" (PDF pages, PPTX slides, XLSX sheets).
Compile a LaTeX document to PDF (full TeX Live, offline, deterministic -- no runtime package fetching). Mounts input TeX file via fragletc file-shaped param (:file) to /input/tex_source.
Render Markdown to one PDF with headless Chromium: inline text, a single .md (or .html) file, or a tar/tar.gz/zip bundle of pages plus the images and fonts they reference by relative path. Several pages become one document with page breaks (ordered by manifest.txt/order.txt when present, else by path) and a table of contents. style=reader applies clean reader typography; style=none renders the page exactly as authored, adding only the page box -- for pages that bring their own <style>, like newspapers, posters, and one-sheets.
Render PDF pages to high-resolution PNG images for visual inspection, slide viewing, or multimodal agent review.
Whether a GitHub event's sender is an approver -- one of the people allowed to say "go" to an automated action -- printed as one line naming who approved what, or a failure giving the reason. The approvers file lists one GitHub username per line (# comments; a leading @ or "- " is ignored, so a plain YAML list also works); matching ignores case. label=<name> also requires the event to be that label being applied, so only that label counts as approval. Its receipt is the approval record: which event, which list, which answer.
The fields of a GitHub issue form as JSON: {"issue": {number, title, author, url, labels}, "fields": {<label_slug>: value}}. GitHub renders a form submission as "### Label" headings with values beneath; each heading becomes a snake_case key, "_No response_" becomes null, and checkbox lists become arrays of the checked options. Accepts a whole GitHub event payload (issues/issue_comment, e.g. $GITHUB_EVENT_PATH) or a bare issue body.
List the lights on a Philips Hue bridge via its local v1 API -- one "id\tname\ton\tbri%\treachable\thex" line per light. hex is the bulb's color as #rrggbb at full brightness (brightness stays in its own column), or "-" when the bulb has no color. hs mode is the inverse of hue-set's hex encoding, so a hex this prints can be sent back.
Set a Philips Hue light (or all lights) via the bridge's local v1 API -- on/off/toggle, brightness percent, white temperature (kelvin or mired), color by hue degrees + saturation percent or by hex (e.g. hex=#ff8000). Light by numeric id or by name (exact then substring match, like home-automation/roku-launch.py); id=all drives group 0. Pairs with home-automation/hue-lights.py for discovery.
List the apps installed on a Roku device via its ECP API -- one "id\tname\tversion" line per app, so a caller can find an app's numeric id for home-automation/roku-launch.py.
Launch an app on a Roku device via its ECP API, optionally deep-linking straight into a piece of content when the app supports it (e.g. a Netflix title id) -- pairs with home-automation/roku-apps.py for the app_id.
Search Roku's on-device universal search for a title and play the Nth result -- Home to a known state, opens Search, types the query via Lit_ keys, arrows Right into the results grid, opens the chosen result, and presses Select to clear provider-choice/profile screens. NOTE: Roku sunset the /search/browse ECP endpoint in OS 12.0, so this deliberately drives the on-screen keyboard instead; and keypress/Enter is a documented no-op on the OS 15 search screen (focus stays parked on "A", so Select-after-Enter types a stray "a") -- do NOT re-add either. Live-verified 2026-10-03 on Roku Ultra OS 15.3.4 (Home, Down x6 to Search, type, Right x6 to result 1, Select opens result, Select picks provider -> playback); result_index beyond 1 extrapolates one more Right per position and is untested. Timing is fixed-delay and can be thrown off by slow app intro animations -- if playback doesn't start, a manual follow-up home-automation/roku-press.py -p key=Select is the fix, same as a human would do.
Send one or more remote-control key presses to a Roku device over its ECP API (no auth, LAN only) -- the same buttons as the physical remote: navigation, Select, Home, Back, playback, volume, channel. Pass a comma-separated list (e.g. "Down,Down,Down,Select") to send a whole sequence in one call instead of one fraglet run per key.
Ask a Roku what it is doing right now via ECP queries -- the foreground app (query/active-app) plus player state/position/duration when the app reports them (query/media-player). Read-only: touches nothing on screen, safe to run mid-playback.
Type free text into a Roku device's currently focused on-screen keyboard field (e.g. a search box) over its ECP API -- sends one keypress/Lit_<char> per character in a single call, instead of one fraglet run per letter. Pairs with home-automation/roku-press.py to navigate to the field first.
Draw a TV-news lower third as a PNG: a studio backdrop, a red kicker box (BREAKING, LIVE, DEVELOPING), the headline bar, an optional ticker strip along the bottom, and the show's name as the corner bug. A still for a satirical or real broadcast; meme/caption puts text on a meme template instead.
Compare two images pixel by pixel and report a machine-readable result (RESULT: IDENTICAL or RESULT: DIFFERENT) with the changed-pixel percentage, maximum channel delta, and bounding box of the change, alongside a PNG that tints every changed region red over the after image. Set tolerance to ignore per-channel noise below a 0-255 threshold, such as antialiasing jitter.
Make a QR code (PNG or SVG) for a link or any text, computed locally with segno and no network: what it encodes never leaves the container. Text that is a secret (an invite link, a Wi-Fi password) goes on stdin, not in text=: params are written into every receipt in plain text, stdin only as a hash. Prints the code's version and size, and with show=true the code itself, drawn in the terminal.
Resize, crop, convert, and compress images (PNG, JPEG, WebP, GIF, TIFF, BMP), reporting the byte savings on stdout. Set max_dimension to fit the longest edge inside a box while preserving aspect ratio, or give width and/or height directly; crop takes "x,y,w,h" in pixels and is applied before any resize. Shrinks oversized screenshots before they consume an agent's context or get embedded in a PDF.
One Cohere chat completion (api.cohere.com/v2/chat) from a prompt file or inline prompt, with an optional system prompt file, written to response.md. The host is fixed -- there is no base-URL parameter -- so the API key can only ever be sent to Cohere. seed makes sampling best-effort reproducible; stdout is a one-line summary (model, seed, tokens, finish reason), never the completion.
One chat completion from a local, OpenAI-compatible model server (LM Studio, llama.cpp, Ollama) from a prompt file and an optional system prompt file, written to response.md; no key. With think=false a reasoning model is asked to answer at once (an empty think block is prefilled), and any think block left in the answer is removed. stdout is a one-line summary, never the completion.
The two files an LLM step reads, for llm/cohere's prompt and system: prompt.md (the task, its subject, reference text, constraints) and system.md (the voice or house style, examples). Reference text and examples are length-capped and fenced as material, never instructions, so an earlier step's output (headlines, a transcript, a page) cannot rewrite the task. For poems, poem/prompt knows the forms.
The Nth Fibonacci number (0-indexed: F(0)=0, F(1)=1). Pure computation, no network.
Check whether a number is prime, and show its smallest prime factor if not. Pure computation, no network.
Render a caption onto a meme-cli template. lines is a single "|"-delimited string, one caption per box in order, whatever the box count -- a 1-box template is just "one caption", a 2-box is "top|bottom", an 8-box is "a|b|c|d|e|f|g|h". Run templates.sh show <id> first to check the real box count/order.
Dump the whole meme catalog in one run: meme-cli list --json. Each template's alias is its id, its field count is the length of text_boxes, and its description is name. The image has no shell, so this is its own fraglet rather than a flag on templates.sh.
Every meme template, one per line: its id, how many captions it takes ("2 boxes") and its name, in columns at least two spaces apart (meme-cli list). For keywords and box geometry, meme/catalog's JSON.
Browse meme-cli's template library: templates.sh search <term>, templates.sh show <template> (boxes and their order), templates.sh list (popular templates), or templates.sh llms (meme-cli's guidance for LLMs). Arguments pass straight through to meme-cli; there are no -p parameters.
Where an IP address is, from ip-api.com (no API key; its free tier is plain HTTP, so treat the answer as a hint, not a secret). Prints one line "City, Region, Country (CC) lat,lon tz" -- the City is the shape weather/current.py, weather/forecast.py and web/search.py take as their location. Pass ip= to look up a specific address; omitted, the caller's own public address is located.
The host's public IPv4/IPv6 address as the internet sees it, from ifconfig.me (no API key) -- what `curl ifconfig.me` prints, as a fraglet so a workflow can record where it ran. Prints the bare address and nothing else.
The latest headlines from a news feed (RSS or Atom), newest first. Clean markdown list by default, or TSV when format=tsv.
Query package registries (pypi, npm, crates, dockerhub) for latest versions, release dates, licenses, and dependencies.
The prompts for a poem about any subject: system.md (the poet's brief: form, voice, length, the answer's shape "Title: ..." then the poem) and prompt.md (the subject and what it must keep), for llm/cohere's system and prompt. Every field is length-capped and fenced as material, never instructions, so whoever fills the subject cannot rewrite the brief. Common forms (haiku, limerick, sonnet, ...) get their rules spelled out.
Attest a fragletc run receipt (fraglet-receipt/2) the way OKF's Attested Computation attester does -- deterministic, no re-execution -- and report RESULT: ATTESTED or RESULT: REJECTED with every reason. Checks that the computation that ran is the sanctioned script (procedure_hash equals sha256 of the computation file, or the expected hash), that the receipt is internally consistent (memo_key recomputes from its own procedure_hash, params, and inputs under fragletc's documented formula, so an edited params field fails), that any claimed params match, that the run succeeded, and that an artifact claimed to be an output of the run hashes to what the receipt recorded.
Typeset an archive of fragletc receipts (fraglet-receipt/2) as a ledger: one row per run with its tool, memo key, params, inputs and outputs by content hash, and exit code, in run order. Where a run's input hash is an earlier run's output hash the row says so, which makes the archive a provenance graph a reader can follow by hand. If the archive also carries blobs/<sha256 hex> files, every hash present in it is marked, so the reader knows which bytes travel with the ledger and which are hash-only. Emits an HTML fragment (default) or Markdown to stdout, for embedding in a page such as a Daily Fraglet edition.
Convert a sheet from an Excel (XLSX) or ODS spreadsheet into clean CSV, TSV, or JSON.
Inspect spreadsheet workbooks (XLSX, ODS), listing sheet names, row and column dimensions, and header columns.
Split text into token-budgeted chunks using OpenAI tiktoken encodings, preferring paragraph boundaries and hard-splitting only paragraphs that exceed the budget on their own. Each chunk is introduced by a "[chunk i/n tokens=k]" marker line, and overlap carries that many tokens of tail context into the next chunk so meaning is not severed mid-thought. Companion to counter, which measures what this splits.
Calculate BPE token counts and token statistics for text or files using OpenAI tiktoken tokenizers (cl100k_base or o200k_base).
A random piece of advice (Advice Slip API, no API key).
A random cat fact (catfact.ninja, no API key).
Fetch a random humorous joke or one-liner (icanhazdadjoke.com, no API key).
Notable events that happened on a month and day in history, from Wikipedia's "On this day" feed (Wikimedia API, no key). One "YEAR<TAB>text" line per event, most recent first, count-limited, so a briefing can pick the one that rhymes with today.
Real current conditions for any city, via wttr.in's compact one-line format (%l: +%t %C) -- genuinely current, not a canned response. Outputs a single short phrase ready for captions or notifications, avoiding unrenderable emoji glyphs.
Three-day weather forecast for a city, region, or landmark (Open-Meteo geocoding + forecast, no API key). Companion to weather/current.py, which only covers right now and takes the same location parameter. Pass day=0/1/2 to print just today, tomorrow, or day after; omitted prints all three days.
A weather frog (Wetterfrosch): how nice one day will be somewhere, as a rung on its ladder from 0 (stay in) to 1 (glorious), plus a one-line croak saying why and whether to take an umbrella, and the chance of rain (rain.txt, 0 to 1) a body can drink from. Open-Meteo, no API key. The rung is a number a body can climb by; the croak is what it says.
Look up whether archive.ph already has a snapshot of a URL -- GET /newest/<url> and report its Location header on a redirect. Lookup only: never submits a new snapshot, so it can't be used to force-archive something that isn't already there. Prints nothing (not an error) when no snapshot exists -- the caller falls back to the original link.
Fetch a URL and extract its <title> and Open Graph tags (og:title, og:description, og:image) -- stdlib html.parser only, no external service.
Extract clean, readable Markdown article text from any web page using Trafilatura, stripping ads, navigations, sidebars, and boilerplate.
Render a webpage to a PDF (headless Chromium via Playwright) -- the print-to-PDF cousin of web/screenshot.py. Pass headers=<json object> to send extra request headers, as in web/screenshot.py.
Capture a full-page PNG screenshot of a webpage (headless Chromium via Playwright). Pass headers=<json object> to send extra request headers. settle_ms (default 2000) is an extra wait after page load, for pages whose real content only appears after their own async JS runs.
Search the web through DuckDuckGo and print one "title<TAB>url<TAB>snippet" result per line, with no API key required. Set site to constrain results to a single domain, timelimit to d, w, m, or y for recency, and region for locale-weighted results. Pairs with page-to-markdown to turn a query into full article text.
The lead summaries of the Wikipedia articles that best match one or more topics (Wikipedia's own search, then each article's page summary), written to summary.md as each article's title, link and a few paragraphs. Topics are separated by semicolons (cat behavior; object permanence) and share the byte budget. An article whose title shares no word with its topic is not taken (a search for a sentence finds noise), so a topic may find nothing; a query nothing matches still succeeds, and summary.md says no article was found, so a workflow that cites it goes on with no evidence rather than stopping. A rate limit or a server error exits 75 (temporary: fragletc --retry). Unlike web/page-to-markdown it needs no URL, only words.
Check whether text is a palindrome, ignoring case, spaces, and punctuation. Pure computation, no network.
ROT13 a string (its own inverse -- run it twice to get the original back). Pure computation, no network.
Filter possible 5-letter Wordle answers given one or more guess+feedback rounds, using 100hellos/java's real wordalytica fluent WordSet API and word list (not a reimplementation). clues is "guess:result|guess:result|...", each result a 5-char G/Y/X string matching that guess letter-for-letter (G=green/correct spot, Y=yellow/wrong spot, X=gray/absent), e.g. "crane:GYXXX|mount:XXGXY". Report the raw colors you see -- this tool does the Wordle-logic-to-predicate translation, not you.
| tool | pack | tags | reach | image | description |
|---|---|---|---|---|---|
| astro/daylight.py | astro | world | ○ hermetic | python3 | How light it is outside right now at a latitude and longitude, as one number from 0 (night) to 1 (the sun overhead): the sine of the sun's altitude, computed with ephem (no network). A body can grow by it. |
| astro/moon-phase.py | astro | world | ○ hermetic | python3 | The Moon's phase on a given UTC date, computed with ephem (no network): phase name, percent illuminated, and the dates of the next new and full moon. One line, ready for a dateline or an almanac box. The date is a parameter, never the clock, so the same inputs always give the same answer. |
| astro/sun-times.py | astro | world | ○ hermetic | python3 | Sunrise, sunset, and day length for a latitude, longitude, and UTC date, computed with ephem (no network). Times are printed in the given UTC offset so a caller can pass the local zone; the day-length delta against the day before says whether days are getting longer or shorter. Companion to net/ip-location.py, whose lat,lon this takes. |
| catalog/compile.py | catalog | engineering, internet | ○ hermetic | python3 | Compile a catalog of level-0 tools -- a tarball of pack directories, each an index.md plus fraglet files -- into its projections: one Agent Skill per pack (SKILL.md beside byte-identical copies of the tools), a Claude Code plugin per pack with a marketplace.json, llms.txt and llms-full.txt for agents arriving over HTTP, manifest.json with every tool's contract and procedure_hash, and an Open Knowledge Format bundle. Everything is derived from the tools' own headers and the pack index files; nothing is authored twice. |
| catalog/site.py | catalog | engineering, internet | ○ hermetic | python3 | Render a compiled tools catalog (archived) into a static website: the catalog itself served verbatim -- llms.txt, manifest.json, every pack's SKILL.md and every tool file at the same paths -- plus an HTML layer over it: an explorer of every tool with search, tag and pack filters, grid and table views and a dark/light theme; one page per pack; one page per tool with its contract, provenance, a curl one-liner and its full source. Relative links throughout, so the result works from a file:// URL as well as a web root. |
| catalog/validate.py | catalog | engineering, internet | ○ hermetic | python3 | Validate a compiled tools catalog (archived): every pack's SKILL.md against the agentskills.io specification using skills-ref, the spec's own reference validator; every tool file present, executable, and hashing to the procedure_hash manifest.json records for it; every llms.txt link resolving; marketplace.json plugins matching the pack directories; and, if an okf/ bundle is present, every concept typed and every relative link resolving. Prints one line per finding and a summary; exit 0 only when clean. |
| catalog/vendor.py | catalog | engineering, internet | ● network | python3 | Fetch the catalog tools a project depends on, verify each against its pinned procedure_hash, and pack them into vendor.tar.gz (executable, at their catalog paths, reproducible bytes). The lock file names where the tools are published ("source https://tools.ofthemachine.com") and lists one "<pack>/<stem>.<ext> sha256:<hex>" per line (# comments allowed); the hashes are what manifest.json and every receipt call procedure_hash. A file whose hash does not match is refused and nothing is written, so base_url may fetch from anywhere -- a mirror, or a local `make serve` -- without weakening the pin or changing the recorded source. |
| chart/plot.py | chart | media, computation | ○ hermetic | python3 | Generate standalone SVG vector charts (bar, line, scatter, pie) from CSV or JSON data. |
| cloudflare/latency.py | cloudflare | engineering, internet | ● network | python3 | How fast a Cloudflare site's origin answers, path by path: requests, p50/p95/p99 and size over the last 24 hours, against the last hour and the 24 hours before, flagging a real slowdown. An address making most of a window's requests is left out of it and named, so a flood cannot skew the numbers. Long polls are labelled and never flagged: their time is waiting. Read-only. |
| cloudflare/traffic.py | cloudflare | engineering, internet | ● network | python3 | A Cloudflare zone's traffic over the last 24 minutes, 24 hours and 24 days, drawn and diagnosed: traffic.svg stacks the three as small multiples (requests from clients by status class, the daily Worker quota and today's projection on the days), worker.svg is the Worker's CPU time per hour against the free plan's 10 ms limit, and report.md is a Markdown report (a headline, the golden signals against the 24 hours before, the top talkers by kind of client and path, and diagnostics that name a poller, the side a 5xx came from, CPU headroom and the quota's pace). Clients are counted by kind (a browser by name, else a program). Read-only; cloudflare/usage is the quota alone, as gauges. It names the top client addresses, with country and client. |
| cloudflare/usage.py | cloudflare | engineering, internet | ● network | python3 | A zone's Worker requests today (the UTC day Cloudflare's daily quota counts) against that quota: how many so far, where the day ends if the last hour's rate holds, and the share of the quota that is, by Worker script; then the last hour's requests to the zone by path and by kind of client (a browser by name, else a program), never by address. The gauges are files (used.txt, projected.txt, share.txt, the projected share as a fraction) and stdout is a short report. Read-only: the token needs only Analytics read and Zone read. |
| code/symbol-outline.py | code | engineering | ○ hermetic | python3 | Extract compact hierarchical polyglot symbol outlines (classes, interfaces, functions, methods, structs) from source code across 140+ languages using Universal Ctags, omitting implementation bodies for up to 95% token savings. |
| data/jq.py | data | engineering, computation | ○ hermetic | python3 | Slice, transform, filter, and extract JSON data using jq query expressions on JSON text or JSON files. |
| data/sql.py | data | engineering, computation | ○ hermetic | python3 | Run SQL queries on CSV, TSV, or tabular data, outputting results as Markdown, JSON, or CSV. |
| diagram/graphviz.py | diagram | media | ○ hermetic | python3 | Render Graphviz DOT source into an SVG, PNG, PDF, or JSON diagram. Choose a layout engine with engine=dot for hierarchies, neato or fdp for spring layouts, circo for cyclic structures, twopi for radial trees, or sfdp for large graphs. Source comes from the dot parameter, a dot_source file, or stdin, so a model can write the graph and this tool renders it deterministically. |
| doc/extract-text.py | doc | documents | ○ hermetic | python3 | Extract readable text out of PDF, DOCX, PPTX, XLSX, ODT, RTF, HTML, and plain-text documents into structured Markdown carrying page, slide, and sheet markers. The format is detected from file content, so no extension is needed; pages selects a 1-indexed subset such as "1-5" or "2,4,7-9" (PDF pages, PPTX slides, XLSX sheets). |
| doc/latex-to-pdf.sh | doc | documents | ○ hermetic | latex | Compile a LaTeX document to PDF (full TeX Live, offline, deterministic -- no runtime package fetching). Mounts input TeX file via fragletc file-shaped param (:file) to /input/tex_source. |
| doc/markdown-to-pdf.py | doc | documents | ○ hermetic | headless-browser | Render Markdown to one PDF with headless Chromium: inline text, a single .md (or .html) file, or a tar/tar.gz/zip bundle of pages plus the images and fonts they reference by relative path. Several pages become one document with page breaks (ordered by manifest.txt/order.txt when present, else by path) and a table of contents. style=reader applies clean reader typography; style=none renders the page exactly as authored, adding only the page box -- for pages that bring their own <style>, like newspapers, posters, and one-sheets. |
| doc/pdf-to-images.py | doc | documents | ○ hermetic | python3 | Render PDF pages to high-resolution PNG images for visual inspection, slide viewing, or multimodal agent review. |
| github/check-approver.py | github | engineering | ○ hermetic | python3 | Whether a GitHub event's sender is an approver -- one of the people allowed to say "go" to an automated action -- printed as one line naming who approved what, or a failure giving the reason. The approvers file lists one GitHub username per line (# comments; a leading @ or "- " is ignored, so a plain YAML list also works); matching ignores case. label=<name> also requires the event to be that label being applied, so only that label counts as approval. Its receipt is the approval record: which event, which list, which answer. |
| github/read-issue-form.py | github | engineering | ○ hermetic | python3 | The fields of a GitHub issue form as JSON: {"issue": {number, title, author, url, labels}, "fields": {<label_slug>: value}}. GitHub renders a form submission as "### Label" headings with values beneath; each heading becomes a snake_case key, "_No response_" becomes null, and checkbox lists become arrays of the checked options. Accepts a whole GitHub event payload (issues/issue_comment, e.g. $GITHUB_EVENT_PATH) or a bare issue body. |
| home-automation/hue-lights.py | home-automation | internet | ● network | python3 | List the lights on a Philips Hue bridge via its local v1 API -- one "id\tname\ton\tbri%\treachable\thex" line per light. hex is the bulb's color as #rrggbb at full brightness (brightness stays in its own column), or "-" when the bulb has no color. hs mode is the inverse of hue-set's hex encoding, so a hex this prints can be sent back. |
| home-automation/hue-set.py | home-automation | internet | ● network | python3 | Set a Philips Hue light (or all lights) via the bridge's local v1 API -- on/off/toggle, brightness percent, white temperature (kelvin or mired), color by hue degrees + saturation percent or by hex (e.g. hex=#ff8000). Light by numeric id or by name (exact then substring match, like home-automation/roku-launch.py); id=all drives group 0. Pairs with home-automation/hue-lights.py for discovery. |
| home-automation/roku-apps.py | home-automation | internet | ● network | python3 | List the apps installed on a Roku device via its ECP API -- one "id\tname\tversion" line per app, so a caller can find an app's numeric id for home-automation/roku-launch.py. |
| home-automation/roku-launch.py | home-automation | internet | ● network | python3 | Launch an app on a Roku device via its ECP API, optionally deep-linking straight into a piece of content when the app supports it (e.g. a Netflix title id) -- pairs with home-automation/roku-apps.py for the app_id. |
| home-automation/roku-play.py | home-automation | internet | ● network | python3 | Search Roku's on-device universal search for a title and play the Nth result -- Home to a known state, opens Search, types the query via Lit_ keys, arrows Right into the results grid, opens the chosen result, and presses Select to clear provider-choice/profile screens. NOTE: Roku sunset the /search/browse ECP endpoint in OS 12.0, so this deliberately drives the on-screen keyboard instead; and keypress/Enter is a documented no-op on the OS 15 search screen (focus stays parked on "A", so Select-after-Enter types a stray "a") -- do NOT re-add either. Live-verified 2026-10-03 on Roku Ultra OS 15.3.4 (Home, Down x6 to Search, type, Right x6 to result 1, Select opens result, Select picks provider -> playback); result_index beyond 1 extrapolates one more Right per position and is untested. Timing is fixed-delay and can be thrown off by slow app intro animations -- if playback doesn't start, a manual follow-up home-automation/roku-press.py -p key=Select is the fix, same as a human would do. |
| home-automation/roku-press.py | home-automation | internet | ● network | python3 | Send one or more remote-control key presses to a Roku device over its ECP API (no auth, LAN only) -- the same buttons as the physical remote: navigation, Select, Home, Back, playback, volume, channel. Pass a comma-separated list (e.g. "Down,Down,Down,Select") to send a whole sequence in one call instead of one fraglet run per key. |
| home-automation/roku-status.py | home-automation | internet | ● network | python3 | Ask a Roku what it is doing right now via ECP queries -- the foreground app (query/active-app) plus player state/position/duration when the app reports them (query/media-player). Read-only: touches nothing on screen, safe to run mid-playback. |
| home-automation/roku-type.py | home-automation | internet | ● network | python3 | Type free text into a Roku device's currently focused on-screen keyboard field (e.g. a search box) over its ECP API -- sends one keypress/Lit_<char> per character in a single call, instead of one fraglet run per letter. Pairs with home-automation/roku-press.py to navigate to the field first. |
| image/chyron.py | image | media | ○ hermetic | python3 | Draw a TV-news lower third as a PNG: a studio backdrop, a red kicker box (BREAKING, LIVE, DEVELOPING), the headline bar, an optional ticker strip along the bottom, and the show's name as the corner bug. A still for a satirical or real broadcast; meme/caption puts text on a meme template instead. |
| image/diff.py | image | media | ○ hermetic | python3 | Compare two images pixel by pixel and report a machine-readable result (RESULT: IDENTICAL or RESULT: DIFFERENT) with the changed-pixel percentage, maximum channel delta, and bounding box of the change, alongside a PNG that tints every changed region red over the after image. Set tolerance to ignore per-channel noise below a 0-255 threshold, such as antialiasing jitter. |
| image/qr.py | image | media | ○ hermetic | python3 | Make a QR code (PNG or SVG) for a link or any text, computed locally with segno and no network: what it encodes never leaves the container. Text that is a secret (an invite link, a Wi-Fi password) goes on stdin, not in text=: params are written into every receipt in plain text, stdin only as a hash. Prints the code's version and size, and with show=true the code itself, drawn in the terminal. |
| image/transform.py | image | media | ○ hermetic | python3 | Resize, crop, convert, and compress images (PNG, JPEG, WebP, GIF, TIFF, BMP), reporting the byte savings on stdout. Set max_dimension to fit the longest edge inside a box while preserving aspect ratio, or give width and/or height directly; crop takes "x,y,w,h" in pixels and is applied before any resize. Shrinks oversized screenshots before they consume an agent's context or get embedded in a PDF. |
| llm/cohere.py | llm | internet | ● network | python3 | One Cohere chat completion (api.cohere.com/v2/chat) from a prompt file or inline prompt, with an optional system prompt file, written to response.md. The host is fixed -- there is no base-URL parameter -- so the API key can only ever be sent to Cohere. seed makes sampling best-effort reproducible; stdout is a one-line summary (model, seed, tokens, finish reason), never the completion. |
| llm/local.py | llm | internet | ● network | python3 | One chat completion from a local, OpenAI-compatible model server (LM Studio, llama.cpp, Ollama) from a prompt file and an optional system prompt file, written to response.md; no key. With think=false a reasoning model is asked to answer at once (an empty think block is prefilled), and any think block left in the answer is removed. stdout is a one-line summary, never the completion. |
| llm/prompt.py | llm | internet | ○ hermetic | python3 | The two files an LLM step reads, for llm/cohere's prompt and system: prompt.md (the task, its subject, reference text, constraints) and system.md (the voice or house style, examples). Reference text and examples are length-capped and fenced as material, never instructions, so an earlier step's output (headlines, a transcript, a page) cannot rewrite the task. For poems, poem/prompt knows the forms. |
| math/fibonacci.py | math | computation | ○ hermetic | python3 | The Nth Fibonacci number (0-indexed: F(0)=0, F(1)=1). Pure computation, no network. |
| math/is-prime.py | math | computation | ○ hermetic | python3 | Check whether a number is prime, and show its smallest prime factor if not. Pure computation, no network. |
| meme/caption.sh | meme | media | ○ hermetic | meme | Render a caption onto a meme-cli template. lines is a single "|"-delimited string, one caption per box in order, whatever the box count -- a 1-box template is just "one caption", a 2-box is "top|bottom", an 8-box is "a|b|c|d|e|f|g|h". Run templates.sh show <id> first to check the real box count/order. |
| meme/catalog.sh | meme | media | ○ hermetic | meme | Dump the whole meme catalog in one run: meme-cli list --json. Each template's alias is its id, its field count is the length of text_boxes, and its description is name. The image has no shell, so this is its own fraglet rather than a flag on templates.sh. |
| meme/list.sh | meme | media | ○ hermetic | meme | Every meme template, one per line: its id, how many captions it takes ("2 boxes") and its name, in columns at least two spaces apart (meme-cli list). For keywords and box geometry, meme/catalog's JSON. |
| meme/templates.sh | meme | media | ○ hermetic | meme | Browse meme-cli's template library: templates.sh search <term>, templates.sh show <template> (boxes and their order), templates.sh list (popular templates), or templates.sh llms (meme-cli's guidance for LLMs). Arguments pass straight through to meme-cli; there are no -p parameters. |
| net/ip-location.py | net | internet | ● network | python3 | Where an IP address is, from ip-api.com (no API key; its free tier is plain HTTP, so treat the answer as a hint, not a secret). Prints one line "City, Region, Country (CC) lat,lon tz" -- the City is the shape weather/current.py, weather/forecast.py and web/search.py take as their location. Pass ip= to look up a specific address; omitted, the caller's own public address is located. |
| net/public-ip.py | net | internet | ● network | python3 | The host's public IPv4/IPv6 address as the internet sees it, from ifconfig.me (no API key) -- what `curl ifconfig.me` prints, as a fraglet so a workflow can record where it ran. Prints the bare address and nothing else. |
| news/headlines.py | news | world | ● network | python3 | The latest headlines from a news feed (RSS or Atom), newest first. Clean markdown list by default, or TSV when format=tsv. |
| package/metadata.py | package | engineering, internet | ● network | python3 | Query package registries (pypi, npm, crates, dockerhub) for latest versions, release dates, licenses, and dependencies. |
| poem/prompt.py | poem | computation | ○ hermetic | python3 | The prompts for a poem about any subject: system.md (the poet's brief: form, voice, length, the answer's shape "Title: ..." then the poem) and prompt.md (the subject and what it must keep), for llm/cohere's system and prompt. Every field is length-capped and fenced as material, never instructions, so whoever fills the subject cannot rewrite the brief. Common forms (haiku, limerick, sonnet, ...) get their rules spelled out. |
| receipt/attest.py | receipt | engineering, computation | ○ hermetic | python3 | Attest a fragletc run receipt (fraglet-receipt/2) the way OKF's Attested Computation attester does -- deterministic, no re-execution -- and report RESULT: ATTESTED or RESULT: REJECTED with every reason. Checks that the computation that ran is the sanctioned script (procedure_hash equals sha256 of the computation file, or the expected hash), that the receipt is internally consistent (memo_key recomputes from its own procedure_hash, params, and inputs under fragletc's documented formula, so an edited params field fails), that any claimed params match, that the run succeeded, and that an artifact claimed to be an output of the run hashes to what the receipt recorded. |
| receipt/ledger.py | receipt | engineering, computation | ○ hermetic | python3 | Typeset an archive of fragletc receipts (fraglet-receipt/2) as a ledger: one row per run with its tool, memo key, params, inputs and outputs by content hash, and exit code, in run order. Where a run's input hash is an earlier run's output hash the row says so, which makes the archive a provenance graph a reader can follow by hand. If the archive also carries blobs/<sha256 hex> files, every hash present in it is marked, so the reader knows which bytes travel with the ledger and which are hash-only. Emits an HTML fragment (default) or Markdown to stdout, for embedding in a page such as a Daily Fraglet edition. |
| sheet/convert.py | sheet | documents, engineering | ○ hermetic | python3 | Convert a sheet from an Excel (XLSX) or ODS spreadsheet into clean CSV, TSV, or JSON. |
| sheet/inspect.py | sheet | documents, engineering | ○ hermetic | python3 | Inspect spreadsheet workbooks (XLSX, ODS), listing sheet names, row and column dimensions, and header columns. |
| token/chunk.py | token | engineering | ● network | python3 | Split text into token-budgeted chunks using OpenAI tiktoken encodings, preferring paragraph boundaries and hard-splitting only paragraphs that exceed the budget on their own. Each chunk is introduced by a "[chunk i/n tokens=k]" marker line, and overlap carries that many tokens of tail context into the next chunk so meaning is not severed mid-thought. Companion to counter, which measures what this splits. |
| token/count.py | token | engineering | ● network | python3 | Calculate BPE token counts and token statistics for text or files using OpenAI tiktoken tokenizers (cl100k_base or o200k_base). |
| trivia/advice.py | trivia | world | ● network | python3 | A random piece of advice (Advice Slip API, no API key). |
| trivia/cat-fact.py | trivia | world | ● network | python3 | A random cat fact (catfact.ninja, no API key). |
| trivia/joke.py | trivia | world | ● network | python3 | Fetch a random humorous joke or one-liner (icanhazdadjoke.com, no API key). |
| trivia/on-this-day.py | trivia | world | ● network | python3 | Notable events that happened on a month and day in history, from Wikipedia's "On this day" feed (Wikimedia API, no key). One "YEAR<TAB>text" line per event, most recent first, count-limited, so a briefing can pick the one that rhymes with today. |
| weather/current.py | weather | world | ● network | python3 | Real current conditions for any city, via wttr.in's compact one-line format (%l: +%t %C) -- genuinely current, not a canned response. Outputs a single short phrase ready for captions or notifications, avoiding unrenderable emoji glyphs. |
| weather/forecast.py | weather | world | ● network | python3 | Three-day weather forecast for a city, region, or landmark (Open-Meteo geocoding + forecast, no API key). Companion to weather/current.py, which only covers right now and takes the same location parameter. Pass day=0/1/2 to print just today, tomorrow, or day after; omitted prints all three days. |
| weather/frog.py | weather | world | ● network | python3 | A weather frog (Wetterfrosch): how nice one day will be somewhere, as a rung on its ladder from 0 (stay in) to 1 (glorious), plus a one-line croak saying why and whether to take an umbrella, and the chance of rain (rain.txt, 0 to 1) a body can drink from. Open-Meteo, no API key. The rung is a number a body can climb by; the croak is what it says. |
| web/archive-snapshot.py | web | internet, documents | ● network | python3 | Look up whether archive.ph already has a snapshot of a URL -- GET /newest/<url> and report its Location header on a redirect. Lookup only: never submits a new snapshot, so it can't be used to force-archive something that isn't already there. Prints nothing (not an error) when no snapshot exists -- the caller falls back to the original link. |
| web/page-metadata.py | web | internet, documents | ● network | python3 | Fetch a URL and extract its <title> and Open Graph tags (og:title, og:description, og:image) -- stdlib html.parser only, no external service. |
| web/page-to-markdown.py | web | internet, documents | ● network | python3 | Extract clean, readable Markdown article text from any web page using Trafilatura, stripping ads, navigations, sidebars, and boilerplate. |
| web/page-to-pdf.py | web | internet, documents | ● network | headless-browser | Render a webpage to a PDF (headless Chromium via Playwright) -- the print-to-PDF cousin of web/screenshot.py. Pass headers=<json object> to send extra request headers, as in web/screenshot.py. |
| web/screenshot.py | web | internet, documents | ● network | headless-browser | Capture a full-page PNG screenshot of a webpage (headless Chromium via Playwright). Pass headers=<json object> to send extra request headers. settle_ms (default 2000) is an extra wait after page load, for pages whose real content only appears after their own async JS runs. |
| web/search.py | web | internet, documents | ● network | python3 | Search the web through DuckDuckGo and print one "title<TAB>url<TAB>snippet" result per line, with no API key required. Set site to constrain results to a single domain, timelimit to d, w, m, or y for recency, and region for locale-weighted results. Pairs with page-to-markdown to turn a query into full article text. |
| web/wiki-summary.py | web | internet, documents | ● network | python3 | The lead summaries of the Wikipedia articles that best match one or more topics (Wikipedia's own search, then each article's page summary), written to summary.md as each article's title, link and a few paragraphs. Topics are separated by semicolons (cat behavior; object permanence) and share the byte budget. An article whose title shares no word with its topic is not taken (a search for a sentence finds noise), so a topic may find nothing; a query nothing matches still succeeds, and summary.md says no article was found, so a workflow that cites it goes on with no evidence rather than stopping. A rate limit or a server error exits 75 (temporary: fragletc --retry). Unlike web/page-to-markdown it needs no URL, only words. |
| words/is-palindrome.py | words | computation | ○ hermetic | python3 | Check whether text is a palindrome, ignoring case, spaces, and punctuation. Pure computation, no network. |
| words/rot13.py | words | computation | ○ hermetic | python3 | ROT13 a string (its own inverse -- run it twice to get the original back). Pure computation, no network. |
| words/solve-wordle.java | words | computation | ○ hermetic | java | Filter possible 5-letter Wordle answers given one or more guess+feedback rounds, using 100hellos/java's real wordalytica fluent WordSet API and word list (not a reimplementation). clues is "guess:result|guess:result|...", each result a 5-char G/Y/X string matching that guess letter-for-letter (G=green/correct spot, Y=yellow/wrong spot, X=gray/absent), e.g. "crane:GYXXX|mount:XXGXY". Report the raw colors you see -- this tool does the Wordle-logic-to-predicate translation, not you. |