Repository navigation
feat(extensions): choose agents per connector, account apps shown once, a cleaner Installed and Discover - #927
Conversation
… and show each agent's step A cold Codex can spend most of its 30-second limit listing plugins, and the account apps were read only after it, so they were never read and Gmail showed "Not connected" for Codex. They are now read beside the listing. The service page says "Checking…" for each agent until the inventories are read, instead of claiming one is not connected, and each agent's row carries its own action: Connect in ChatGPT, Connect on claude.ai, or Sign in. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rives The service page waited for every agent's inventory before trusting any, so one agent that never reports kept Claude and the header on "Checking…". An agent now counts as read once its own servers arrive or the whole read ends, the header shows only the next step, and agents still checking are left out of Works with. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tGPT in place A Scient Agent signed in to Claude or another account showed a bare Sign in, as if it had no sign-in at all. Its row now says what it is signed in with and that the service needs its ChatGPT sign-in, and Add ChatGPT (or Sign in with ChatGPT) opens that one sign-in right in the row, run by the agent as on the Providers page. A ChatGPT sign-in on a version without ChatGPT apps reads Update needed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
claude.ai moved connectors from Settings to Customize, so the old settings/connectors address only says they have moved. Connect on claude.ai and Add more on claude.ai now open claude.ai/customize/connectors. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Featured becomes Most used: Gmail, Google Drive, Google Calendar, GitHub, Slack, Notion, PubMed and Semantic Scholar, what nearly everyone needs whatever their field. The field picks it held (Wolfram, Hugging Face, NASA Earthdata, BioRender) go to their fields. Teamwork splits into Email, calendar & chat and Projects & tasks; Writing & references becomes Files & documents, with citation tools under Papers & literature; Data becomes Data & databases. The fields follow, led by where most users work: papers, biology & medicine, code, data, physics & maths. Agent store categories map to the new ones, and a home section counts the account services it shows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Added since opening, from the owner's dev-app review:
Checked: type check, lint, server tests (338), Extensions web tests (120), seam check. Screenshots are in reviews/account-apps-20261114 (discover-most-used.png, discover-fields.png). 🤖 Generated with Claude Code |
Thread transfer impact✅ Thread transfer remains within every enforced ceiling.
Other scenariosAll 64 scenario measurements match All 64 scenario measurements
Baseline: Scenario and decoded snapshot size10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed. |
…in its own section From Claude hid the claude.ai Gmail, Calendar and Drive connectors, so it looked as if Claude had nothing installed. They show again in each agent's section and open the service's one page. A Scient Agent signed in to ChatGPT on a version without ChatGPT apps now says the apps come with the next update. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cted services fills in as agents report Gmail, Calendar and Drive rows in an agent's section showed "Claude only" though every agent reaching the service has it; they now carry the service's own Works with. Connected services waited for every agent's inventory, so one that never reports left it showing only connectors; it now fills in as each agent reports. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…onnectors Installed showed Connected services, Scient plugins and a section per agent, each a grid of cards with its own Refresh and Show all. It is now one table, Yours: each thing you have is one row with its name, its type, the logos of the agents that use it, and its action. Gmail, Calendar and Drive are one row naming every agent that reaches them; Scient's connectors and plugins and each agent's own plugins, apps and servers follow. One Refresh reads every agent again, Show all opens the rest past 24 rows, Works with still narrows it to one agent, and Add apps on ChatGPT and Add connectors on claude.ai sit under it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s own table A ChatGPT app Scient Agent and Codex both reach through their ChatGPT accounts showed twice; it is one row naming both. An agent's plugin named like one of those rows, or like another of its own plugins, adds no row. Codex's ChatGPT apps match Gmail, Calendar and Drive by name too. Scient's built-in plugins move out of Yours into their own table at the bottom, Built into Scient. The Works with column says All agents for what reaches every agent and shows at most three logos, and columns keep one width across rows, so the table never runs wider than the page. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…plugin pages show each agent Only Gmail, Calendar and Drive had the page that shows each agent with its own step. Any ChatGPT app the agents report is now such a service: Codex and Scient Agent reach it through ChatGPT, Claude through a claude.ai connector of the same name. A Codex plugin's page shows, under each of its apps, every agent with its state and action (Connect in ChatGPT, Connect on claude.ai, Add ChatGPT), instead of a lone Connect next to Remove; that button now reads Remove from Codex. A ChatGPT app's row in Yours opens its page. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…w each agent's step Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…idth Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The app dialog's "Connect Figma" opened the same ChatGPT page as the agent rows' "Connect in ChatGPT", so it read like a second, different step. It now has the same words. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
claude.ai gives each connector a page (claude.ai/customize/connectors/id/ figma), so Claude's Connect for Gmail, Figma or any other service now opens the connector itself, not the list of every connector. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…give way As on claude.ai's Connectors page: the table's rows divide by lines, and the row under the pointer turns into a rounded tile while the lines above and below it disappear. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Discover's rows added a facts line (Official · Account needed · Registered OAuth client needed) under the tagline, which wrapped and made some rows taller than the rest. Rows now show only the one-line tagline; the facts stay on each item's page. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The two big top-bar buttons become one + Add menu: create a connector with an agent, or add from a file, folder or link. Discover lists each agent store in its own section, which names the agents once, so rows say Add and Remove; each section has a single Show more that asks every store in it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Discover's + and an item's Add now ask which of your agents get it when more than one takes connectors (all ticked); with one agent it adds at once. A connector keeps an optional agents list: delivery skips other agents (agent-deselected) and the proxy re-checks the session's agent on each request. The connector page gains an Agents editor (all or selected) beside Available in. Most used adds Playwright (pinned @playwright/mcp, run locally, added in one step), Vercel and Cloudflare's API server, and the home page shows all of Most used. Card rows drop the Works with line and the store links; the item's page carries both. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Claude route linked to claude.ai/customize/connectors/id/<name> for every ChatGPT app, a guess that only held for some (Outlook led nowhere). claude.ai links now come from a checked-in list of each connector's own Add to Claude address in Anthropic's public directory (993 connectors, refreshed by scripts/generate-claude-connectors.mjs). An app claude.ai lacks gets no Claude row; Microsoft's apps map to Microsoft 365, and the row names claude.ai's connector. ChatGPT links follow Codex's exact rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… were; Telegram, Exa, Context7 in Most used A category page shows the picks Discover already has while the full list loads, starts loading when its link is hovered or focused, keeps store results for ten minutes, and shows loading notes only after 700 ms. An item page's back link returns to the list it was opened from (Discover, a category, a store, All MCP servers or Installed), by name. Most used adds Telegram (@ghostpaw/telegram 0.7.9, four tools, asks for a @Botfather token), Exa and Context7. The local add dialog says the chosen agents get it when agents were chosen. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Installed: a plugin two agents share is one row, plugin-owned servers and Codex's app gateway are not listed again, every account app is one App row with all its agents, bundles sit with Yours, the agent filter honours a connector's chosen agents and lists only agents the person has. Type says Connector, App, Plugin, MCP server or Built in; descriptions carry no ids. Pages: an account app shows its agents first, with no second status or Works with row; Works with values drop the repeated words; a connector page orders Agents before Projects and says No account needed once; item pages drop internal ids. Removing a connector returns to the list. The agents note no longer appears because an agent is off or keeps no inventory; an account app page stops reading once the read has ended. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…tests at the right level - providerStoreAudience replaces three inline copies; the catalog list is built once per render; the unreachable reconnect action and skill group are deleted; skill rows show descriptions without Markdown marks. - Tests: the add forms pass the chosen agents; the account app page shows Claude only where claude.ai has the connector; the checked-in link list and the installed-services rule are unit-tested; leaked mocks and the remembered list page are reset; three page tests that repeated providerInventory unit tests are deleted. - Docs follow the Type names, the Projects setting and the Yours table. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… its connectors; reads fail soft Agent choice: - Agents' own connector tools (list, check, control, authoring with an installed connector's keys) refuse an agent that was not chosen. - The state file is written as format 3 when a connector has an agent choice, so an older build refuses it instead of dropping the choice; files without choices stay format 2. - GitHub through Source Control keeps or applies the choice; ids that are not configured agents are dropped, and a list left empty is refused. Delivery: for Scient Agent only a connected server of its own takes a connector's name. Its listed servers come from config and plugin files; a Claude plugin's Playwright is filtered by the agent itself, which left the agent with neither that nor Scient's Playwright connector. Reads: Scient Agent's plugins and apps are read together, and a failed Claude plugin read keeps its own; a failed or timed-out apps read keeps the last apps; a failed Claude account probe retries after 15 s; a store entry that could not be read is no longer reported as gone. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
|
This PR changes the test suite by +23 test cases (browser +3, unit +3, server +17). Keep tests that would fail for a real reason, and test logic below the UI where it can be. See Testing. |
…es-fixes-20261114
…ce; rows keep their controls - scient_connector_setup refuses an agent that was not chosen, before any account card or change; scient_connector_list with a query no longer names a connector given to other agents only, nor offers a second copy. - Installed: one row per installation (owner, plugin id, workspace); only a Claude plugin Scient Agent also loads collapses. A plugin's server that still needs sign-in keeps its row, the only place that offers it. - Seam manifest classifies the connector-list generator and its knip entry. Telegram is labelled Community. User docs describe the App page. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
An agent that was not chosen could still copy a connector's code with scient_connector_copy, and scient_connector_add by catalog name returned its server name, status and missing setup. Copy now finds nothing by that name for such an agent; add says only that it is added for other agents. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… add or check Every connector tool an agent has now goes through one check. Adding the address of a connector given to other agents only no longer names it, and checking one returns before its setup or sign-in state is read; the runtime's own test refuses the agent before it reports a needed sign-in. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…t say Setup echoes the name the agent gave; folder check and save look the connector up as that agent, so one given to other agents is not found; the runtime's own refusal names no connector. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…es-fixes-20261114
…tores Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…row per thing CI: - The row hover hid its neighbours' lines with a sibling :has(), which main now bans in Scient styles. Each row carries its line on top and the next row drops its own when the one before it is hovered. - storeCategories still expected the removed writing category. Installed, with every project loaded: - A project that reports a machine-wide plugin again, in the same state, is not a second installation; one turned off for that project is. - A plugin that carries an app already listed adds no row: the App row stands for it, and its page stays reachable from the agent's store. - The note about missing inventories shows only when a read failed, with the reason; an agent that keeps no per-project list is not a gap. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Settings → Extensions, reworked with the owner over a day of dev-app testing, then reviewed end to end (three Codex code reviews, a test review, a design review on screenshots, and live tries with each agent).
What people get
Choose which agents get a connector.
Apps reached through an agent's own account, shown once.
scripts/generate-claude-connectors.mjs); ChatGPT links follow Codex's exact rule. An app claude.ai lacks gets no Claude row; Microsoft's apps map to Microsoft 365.Discover. Most used first (now with Playwright, Telegram, Exa, Context7, Vercel, Cloudflare), categories by kind of work, one section and one Show more per agent store, one + Add menu, categories open filled, item pages go back to the list they came from.
Installed. One Yours table with one row per installation, Type (Connector, App, Plugin, MCP server, Built in) and Works with; built-ins follow. No raw ids, no repeated status. A Claude plugin Scient Agent also loads is one row; a plugin's server is listed only while it still needs a sign-in.
Fixes found by trying it
playwright__browser_navigateworks.scient_connector_list,add,control,check,copy,save,setup) now refuse or hide it, and no refusal names anything the agent did not type. Tests go through the tool paths.Not in this PR
Caching the shared store listing, keeping the last Installed result while it refreshes, and merging the two store-list implementations. Each is larger than this PR should carry. The confirm dialog's button stays "Confirm" (shared T3 component).
Checks
Platforms
Behaviour is the same on macOS, Windows and Linux. The local Playwright entry runs
npx @playwright/mcp@0.0.85, which needs Node on the machine as any local package connector does; it was run on macOS only. ChatGPT links were not opened signed in.🤖 Generated with Claude Code