allthings from your terminal: what's coming up, who's on stage and how to get in. It reads the same public MCP server agents use, so what you see here is what they see.
curl -fsSL https://allthings.dev/install.bash | bashThat installs the latest stable release, allthings 2, in ~/.allthings/bin.
It is also on npm as allthings, for Bun:
bunx allthingsallthings # the next evening, and what to run from there
allthings events # what's coming up
allthings events --past --limit 5 # the five most recent evenings
allthings event next # the next evening: when, where, the talks
allthings event latest # the last one, with its video when there is one
allthings speakers react # who has been on stage about react
allthings rsvp next # open the next evening's page to say you're in
allthings open latest # open the last evening's page on allthings.dev
allthings guide # conferences worth the trip, starred as Michelin stars restaurants
allthings about # where to find everyone between evenings
allthings coach slides.md talk.vtt # coaching for your talk, before you give it
allthings propose # propose a talk for an eveningevent, rsvp and open take an evening's slug, from events, or next and latest, so every line above runs as typed. With nothing on the calendar, allthings shows the last evening and where the next one is announced first, and event next exits 3.
Add --json to any of them except rsvp and open, which print just the link, for the public contract, unformatted, for scripts and agents. On its own, allthings --json gives { next, latest }, each an evening or null. Exit codes are part of that contract: 0 ok, 1 service error, 2 usage error, 3 no such evening.
ALLTHINGS_MCP_URL points it at another server, such as a local Worker.
allthings coach slides.md rehearsal.vtt # a 3-minute open-floor slot
allthings coach outline.md --slot 20 --goal teach-a-technique
allthings coach rehearsal.txt --goal first-talk --jsoncoach sends your talk to coach_talk, the MCP server's coaching, and prints what to change first, then each finding (must, should, consider) with what it measured and what to do, what already works, and what it couldn't check without more. Every check is explained at allthings.dev/coach.
A file's extension says what it is: .md is an outline, or slides split by lines of ---, and its first heading is the title unless --title gives one; .txt, .vtt and .srt are a rehearsal's transcript, and with timestamps (.vtt, .srt) it also checks your pace. Send at most one of each. --slot is the slot in minutes: 3, an open floor's, unless given, or 20 with --talk, and a slot over 5 minutes is a talk in the lineup. --goal says what you want from the talk, as often as you like: first-talk, teach-a-technique, find-contributors, show-a-launch or get-hired. --abstract adds the abstract.
The files are read only to send them; coach_talk keeps none of what it is sent. A file it can't read, one too long for coach_talk, or an unknown goal exits 2 before anything is sent.
allthings propose # asks for each part, shows it, and sends it on your yes
allthings propose --title "Effect in production" --abstract "What a year of Effect in a payments service taught us about errors." \
--name "Ada Lovelace" --email ada@example.com --format demo --link https://github.com/adaAn open door and a high bar: anyone may propose, and Erik and Andre read every proposal and reply by email. In a terminal it asks for whatever the options leave out (everything, with none), then shows the proposal and asks before sending; --yes skips that. Elsewhere, as in a script or an agent, --title, --abstract (at least 40 characters), --name and --email are required; --format (talk, lightning or demo), --about, --link (up to four) and --notes are not. With --json it prints the server's answer. It exits 2 with what to fix when the server finds a problem, and 1 when it can't take it (too many at once, or a stage that doesn't keep them). It sends through the MCP server's propose_talk, naming itself in its User-Agent, so the organizers see it came from the CLI. Nothing you send shows on allthings.dev unless you give the talk.
allthings mcp # a line to add allthings to Claude Code, Codex, Grok Build, VS Code or CursorIt prints the server's URL, its tools and a line for each agent, for example claude mcp add --transport http allthings https://allthings.dev/mcp. It reads nothing, so it works offline. With ALLTHINGS_MCP_URL set, the lines name that server instead. allthings mcp --json gives { name, url, transport, tools }. An agent with a shell can also just run this CLI with --json.
At an evening that passes the aux, guests set the room's lights from their terminal, with the code on the card from the check-in table:
allthings aux login <code> # save the code from your card
allthings aux status # who has the aux, your place in line, your cooldown
allthings aux scenes # tonight's scenes
allthings aux scene ocean # set the room to one of them
allthings aux color mint # set a color, everywhere or on one --zone
allthings aux request # get in line for the aux
allthings aux pass # hand it to the next person
allthings aux tracks # tonight's tracks, when music is on
allthings aux queue <slug> # queue one
allthings aux skip # skip the track playing, if you hold the auxlogin saves the code in ~/.allthings/aux-code ($ALLTHINGS_HOME/aux-code when the installer's ALLTHINGS_HOME is set), readable by you alone. No command prints the code, even when the hub sends it back. ALLTHINGS_AUX_CODE wins over the saved code.
The hub is an MCP server at https://aux.allthings.dev/mcp, and ALLTHINGS_AUX_URL points at another, such as one on localhost. Every call sends the code as Authorization: Bearer <code>, so agents can use the same endpoint and code. Its tools are aux_status, set_scene, set_color, request_aux and pass_aux, plus list_tracks, queue_track and skip_track when music is on. The music commands check tools/list first and say "Music isn't on tonight" when they're missing.
When the hub refuses, the CLI says why from the tool error's code: going too fast (rate_limited, with the seconds to wait from retry_after_seconds, or from Retry-After on an HTTP 429), guest control paused by the host (frozen), a scene or color tonight doesn't have (unknown_scene, unknown_color), something only the aux holder can do (not_holder), or a code the hub doesn't accept (unauthorized, or an HTTP 401). Exit codes: 1 when the hub refused or couldn't be reached, 2 for an unknown scene or color, a missing code and other usage errors.
status reads holder, queue_position, cooldown_seconds, frozen, scenes and colors (or their camelCase names) from aux_status, and prints the hub's own text when its result has another shape. --json prints what the hub returned.
This CLI used to be atw. The installer still puts an atw beside allthings, and the package still has an atw bin. Both run the same CLI and print a note on stderr, so scripts reading stdout keep working. The old name goes away in a later release. ATW_MCP_URL and ATW_VERSION are still read when the new names are unset.
bun install
bun run check && bun run typecheck && bun test
bun run src/index.ts eventssrc/config.ts holds the site it talks to, siteOrigin: https://allthings.dev, whose MCP server is at /mcp.
Releases are GitHub releases, cut by pushing an annotated tag that matches version in package.json (.github/workflows/cli-release.yaml). A hyphenated tag (2.0.0-alpha.2) is a prerelease, and the installer skips it unless asked.
The same workflow then publishes the package to npm with bun publish, after the binaries. bun run build (which packing runs) bundles src/index.ts and src/atw.ts for Bun into dist/, and the package ships only that, with zod inside. A prerelease gets the alpha dist-tag, a release latest. npm gives a package's first version latest whatever its tag, and latest can't be removed, so until there is a stable release latest follows the newest prerelease. The step then checks the registry's dist-tags. It publishes with --tolerate-republish, so rerunning it after a publish that went through repairs and checks the dist-tags instead of failing on the existing version. It authenticates with npm trusted publishing: the job trades GitHub's OIDC token at npm's exchange for a short-lived token that can publish allthings, so no long-lived npm token is stored anywhere. npm accepts the trade only from this workflow, set as the trusted publisher in the package's settings.
Publishing to npm waits for a person. The npm job runs in the npm-release environment, which needs Erik's approval in GitHub (admins can't bypass it) and accepts only version tags, and npm's trusted publisher names that environment. The job also stops unless the tagged commit is on main. So having write access to the repository isn't enough to publish the CLI.