Toolbox

The AI tools Every uses. Toolbox shows which AI tools and models the Every team uses for each of 11 kinds of work (coding, knowledge work, writing, research, classification, image, video, animation, text to speech, speech to text, music), and lets each person share what is in their own AI toolbox.

• Home is the team page: the most used tool and model per kind of work as plain counts ("5 of 6 use it"), the latest model launches, an Overall top 10, and filters for who is counted. Each kind has its own page with who ranks what first, second and third, and how people set their tools up.
• Sign in with Every, claim toolbox.every.to/<you>, and rank up to three picks per kind of work in about a minute. A pick is a tool (Cursor), optionally with the model behind it (Claude Opus 5.5), a context size and an effort.
• Or connect your agent (Claude, Claude Code, Cursor, Codex) over MCP with a normal sign-in. It can only suggest picks: nothing shows on your page until you confirm it on the site.
• New pages start public (Anyone with the link). Choose Private, Every team or Anyone with the link: that decides who can open your page and see your name. Every onboarded Every team member's picks count toward the team's totals either way; a private member is counted anonymously and never named. The link-preview card is served only for Anyone with the link.
• Tools and models show their real marks where the catalog has one, and a serif initial otherwise. The look is "Every dark".
• Every change to your picks is kept with its date, so a kind's page can show what the team used before a new model shipped.

Built on compound-stack-rails 0.8.0: Rails 8.1, Inertia, React 19, Tailwind 4, SQLite, and Kamal. The plans live in docs/plans/ (the v1 plan, then the Every Toolbox redesign), and the design of record is in docs/design/every-loadout/.

Local development

bin/setup            # install deps, prepare the database
bin/rails db:seed    # the catalog plus demo members
bin/dev              # boot Rails + Vite (open http://localhost:3100)
bin/rails test       # Ruby suite
npm run check        # tsc x2 + Vitest

Sign in with Every is the only production login. Locally, the sign-in page lists the seeded members under "Dev login"; they have mixed visibility, ranked picks and two open suggestions. To try real Every sign-in, set the EVERY_OAUTH_* and PUBLIC_BASE_URL variables from .env.example. Every team status needs a verified @every.to email, so a local sign-in with another address counts as "Everyone else".

Connect an agent (MCP)

The MCP server is https://toolbox.every.to/mcp (Streamable HTTP; the host is whatever PUBLIC_BASE_URL says). It uses the MCP authorization spec: OAuth 2.1 with PKCE and dynamic client registration. There are no tokens to paste. Your client opens a browser, you sign in with Every, and you approve.

Client	How
Claude (claude.ai, desktop)	Settings → Connectors → Add custom connector → https://toolbox.every.to/mcp
Claude Code	claude mcp add --transport http toolbox https://toolbox.every.to/mcp
Cursor	"Add to Cursor" on /agents, or {"mcpServers": {"toolbox": {"url": "https://toolbox.every.to/mcp"}}}
Codex	codex mcp add toolbox --url https://toolbox.every.to/mcp

Then ask: "Suggest picks for my Toolbox from what you know about how I work. Ask me before you guess, and I'll confirm them on the site."

The tools are list_categories, search_catalog, get_my_toolbox, get_team_rankings, get_recent_changes and suggest_picks. suggest_picks is the only write, and it only proposes: you confirm or dismiss each suggestion in the Rank editor, and confirming, removing, moving, visibility and revoking are web-only. get_team_rankings returns what Home and the kind pages show, as you would see it. The same tools are exposed through WebMCP while you are signed in on the site. Connected agents are listed, and can be revoked, at /agents.

Catalog

config/catalog.yml seeds categories, tools and models, including which real mark each item shows. It syncs on db:seed and on every deploy, and is idempotent. Members can add anything missing while picking; those items show to that member as pending and wait for review at /admin/catalog_items. Admins come from the ADMIN_EMAILS env var (an address counts only once Every verified it) or users.admin. Admins also set a model's release date and Vibe Check link there: a model is listed as a launch on Home only when it has both, and the link must be https on an allowed host (every.to and checks.every.to, or the hosts in VIBE_CHECK_HOSTS when that is set).

Screenshots

The marketing screenshots in docs/screenshots/ are from v1 (light theme, the members-only map). They predate the redesign and are not yet regenerated.

Deploy

See DEPLOYING.md. Toolbox deploys with Kamal to Hetzner, like happyhappy. Before the first deploy, Kieran needs:

• DNS: an A record for toolbox.every.to pointing at the Hetzner server's IPv4 address (add an AAAA record for IPv6 if the server has one). kamal-proxy issues the Let's Encrypt certificate.
• Every OAuth client: redirect URI https://toolbox.every.to/auth/every/callback, scopes basic_profile openid (see DEPLOYING.md for automatic sign-in). Put its id and secret in EVERY_OAUTH_CLIENT_ID and EVERY_OAUTH_CLIENT_SECRET.
• Env: PUBLIC_BASE_URL=https://toolbox.every.to (required in production), EVERY_OAUTH_BASE_URL=https://every.to, and ADMIN_EMAILS=kieran@every.to.

For agents

Read AGENTS.md first. The one write path for toolboxes is Toolbox::Update (app/services/toolboxes/update.rb). The web editor, MCP and WebMCP all go through it, and it records the dated history in entry_changes. Agent sources (mcp, webmcp) may run only suggest and withdraw. Every count on a page or in a tool comes from the shared read layer in app/queries/ (Audience, TeamRankings, PersonPicks, ModelLaunches, NumberOneHistory, Search), read as the viewer.