Skip to content

MCP Connectors ​

Screens for managing Model Context Protocol servers: the connectors your product's assistants can call, what each tool is allowed to do, and who signed in to what.

They follow the MCP specification dated 2025-11-25 and the MCP Registry's server.json format. If your product lets an AI model call tools on somebody's behalf, these are the screens an administrator asks for first: which servers are installed and healthy, which tools can run without asking, what changed since somebody approved it, and a record of every call.


The catalogue ​

PagePathWhat it is for
MCP connectors/account/mcp-connectors.htmlInstalled servers: transport, sign-in state, health and latency, tools exposed and blocked, scope, owner; bulk on and off; the tool budget; changed-tool alert
MCP connector/account/mcp-connector.htmlOne server in tabs: overview with a latency sparkline, tool policies, resources and prompts, OAuth and service key, the call log with CSV export, settings from server.json
MCP directory/account/mcp-directory.htmlA searchable catalogue by category, with Verified and Official publishers and a reads-only or can-change-data chip on every listing
Add an MCP connector/account/mcp-new.htmlA wizard: choose, endpoint, probe, sign in, tool policies, access, then a test call and client config to copy

All four sit under Account in the sidebar and the account tabs, with a sub-rail (.av-subnav) for Installed, Directory and Add a connector. Every one is registered in PAGES in tests/lib/server.mjs, so the browser suites hold them to the same gates as the rest of the demo.


What the screens model ​

Transports ​

A connector reaches its server one of three ways, and the list says which:

  • Remote HTTP: the Streamable HTTP transport, a single endpoint such as https://api.githubcopilot.com/mcp/.
  • Local stdio: a process the workspace starts, from a package in server.json's packages.
  • SSE: the older HTTP with Server-Sent Events transport, deprecated since the 2025-03-26 protocol. The badge is .av-badge--warning-soft and the row says to move to Streamable HTTP.

Sign-in states ​

Connected, Needs sign-in, Expired and Error are the values of the status chips above the table, and each row carries one in data-av-status. Health is separate from sign-in: a dot (.av-dot), the last check and the latency, because a server can be signed in and slow, or healthy and unauthorised.

OAuth, as the specification describes it ​

The probe step in the wizard and the Authentication tab show each part of MCP authorisation so an administrator can see what was negotiated:

  • the server's 401 with a WWW-Authenticate header naming its protected resource metadata (RFC 9728) and the scopes it wants;
  • the authorisation server that metadata points to, and its own metadata;
  • PKCE with S256, which the wizard requires rather than falling back to plain;
  • the resource indicator (RFC 8707), so a token is bound to one server;
  • step-up: a scope left out at sign-in is asked for when a tool gets 403 insufficient_scope.

Tool annotations are hints ​

readOnlyHint, destructiveHint, idempotentHint and openWorldHint come from the server, and anyone running a server can set them. The Tools tab groups tools by them ("as declared by the server") and says so in words, but the policy beside each tool is what decides. The pages start destructive tools at Ask whatever the default is; enforce the same rule where your app stores the policies.

Changed tools ​

When a server's tool list changes (notifications/tools/list_changed), a tool whose description or input schema differs from the approved version goes back to Ask and is marked Changed. The review dialog shows each change with .av-diff, descriptions included, because a description is an instruction to the model.


Components these pages added ​

None of these is specific to MCP. Each is documented here because these pages are where they first appear.

Policy list ​

A list of things that each carry one of three permissions. The row is a size container: from 34rem the control sits beside the text, below that it drops underneath it.

html
<ul class="av-policy">
  <li class="av-policy__item" data-av-state="changed">
    <div class="av-policy__row">
      <div class="av-policy__main">
        <p class="av-policy__name">
          <code>create_issue</code>
          <span class="av-badge av-badge--warning-soft">Changed</span>
        </p>
        <p class="av-policy__text">Open a new issue in a repository.</p>
        <div class="av-policy__hints">
          <span class="av-badge av-badge--neutral-soft">Open world</span>
          <button class="av-btn av-btn--sm av-btn--quiet" type="button" data-av-collapse
                  aria-expanded="false" aria-controls="schema-create-issue">Input schema</button>
        </div>
      </div>
      <div class="av-segmented av-policy__control" role="radiogroup" aria-label="Policy for create_issue">
        <label><input type="radio" name="pol-create_issue" value="allow">Always allow</label>
        <label><input type="radio" name="pol-create_issue" value="ask" checked>Ask</label>
        <label><input type="radio" name="pol-create_issue" value="block">Block</label>
      </div>
    </div>
    <pre class="av-code av-disclosure__panel av-policy__detail" id="schema-create-issue"
         data-av-state="collapsed"><code>{ "type": "object" }</code></pre>
  </li>
</ul>

data-av-state="changed" draws a warning edge on the row; the badge carries the meaning for a screen reader. A permission that somebody else fixed is a sentence, not a control: replace the .av-segmented with <p class="av-policy__locked">, a lock icon and "Set by admin". A radio group nobody can change still announces itself as something to operate.

Group control: [data-av-cascade] ​

One radio group in a header that sets every radio group below it. Put data-av-cascade on the element that holds both, data-av-cascade-parent on the header's group, and data-av-cascade-status on the line that explains a split.

html
<section class="av-card" data-av-cascade data-av-cascade-noun="write tool">
  <div class="av-card__header av-policy-head">
    <div class="av-policy-head__row">
      <div>
        <h2 class="av-card__title">Write</h2>
        <p class="av-policy-head__status" data-av-cascade-status role="status" hidden></p>
      </div>
      <div class="av-segmented av-policy__control" data-av-cascade-parent
           role="radiogroup" aria-label="Policy for every write tool">
        <label><input type="radio" name="grp-write" value="allow">Always allow</label>
        <label><input type="radio" name="grp-write" value="ask">Ask</label>
        <label><input type="radio" name="grp-write" value="block">Block</label>
      </div>
    </div>
  </div>
  <div class="av-card__body">
    <ul class="av-policy"><!-- rows as above --></ul>
  </div>
</section>
  • Choosing a value in the parent checks the radio with the same value in each child group and fires change on it. A child group without that value, or with disabled radios, is left alone.
  • When the children disagree, the parent shows no selection, the wrapper takes data-av-state="mixed", and the status line reads, for example, "Mixed: 2 Always allow, 1 Ask". A header that kept showing the last value it set would be a summary that contradicts the rows.
  • data-av-cascade-noun names what a child is, for the announcement ("Ask set for 3 write tools").
  • The module emits av:cascade:change with { el, value, changed }.

Children are found by name each time, so a row you insert later is included without calling init() again for the cascade.

Diff ​

Two versions of a text, line by line. The + and - are text in .av-diff__mark, and each changed line starts with a visually hidden "Added:" or "Removed:", so the colour is never the only cue.

html
<div class="av-diff" role="group" aria-label="Changes to the search_code description">
  <div class="av-diff__line av-diff__line--removed">
    <span class="av-diff__mark">-</span><span class="av-visually-hidden">Removed: </span>Search code in public repositories.
  </div>
  <div class="av-diff__line av-diff__line--added">
    <span class="av-diff__mark">+</span><span class="av-visually-hidden">Added: </span>Search code across every repository the account can read.
  </div>
</div>

The block scrolls inside itself, so a long line never widens the page.

List filter: [data-av-list-filter] ​

Search and one row of chips over a list that is not a table, such as the directory's card grid. [data-av-datagrid] does the same for table rows; this has no sort, pages or export.

html
<div data-av-list-filter data-av-list-filter-noun="connector">
  <label class="av-toolbar__search">
    <input type="search" name="q" data-av-list-filter-search aria-label="Search the directory">
  </label>
  <span class="av-toolbar__count" data-av-list-filter-count aria-live="polite"></span>
  <div class="av-demo-row" role="group" aria-label="Filter by category">
    <button class="av-filter-chip" type="button" data-av-list-filter-chip="all" aria-pressed="true">All</button>
    <button class="av-filter-chip" type="button" data-av-list-filter-chip="dev" aria-pressed="false">Developer tools</button>
  </div>
  <div class="av-integration-grid">
    <article class="av-integration" data-av-list-filter-item data-av-list-filter-tags="dev">…</article>
  </div>
  <div class="av-empty" data-av-list-filter-empty hidden>…</div>
</div>

An item shows when its text contains the search and its data-av-list-filter-tags include the pressed chip's value (all matches everything). The count is written from the result rather than typed into the markup, and a chip press is announced. The module emits av:list-filter:change with { el, shown, query, tag }.

Bulk on and off ​

The connectors table turns several connectors on or off at once with data-av-grid-bulk-switch, documented with the rest of the grid in Tables & grids. The required connector's switch carries aria-disabled="true", so the bulk action skips it and says so.


Building on these pages ​

  • The approval prompt on the connector page (#approvalPreview) is the dialog your chat shows before a tool set to Ask runs. Keep the arguments exact and complete: "Allow GitHub to create an issue?" without them is a question nobody can answer.
  • Redact before you store. The call log shows [redacted] for values the server marks isSecret and for anything shaped like a token. Do that on the server; the page only displays it.
  • Secrets stay masked. Service keys and the gateway token use [data-av-secret], described in Account, Access & Storefront. The data-av-secret attribute is where a static page keeps the value; in your app it becomes a request made on Reveal.
  • The tool budget is a .av-usage row with .av-usage__row--over. Count what you actually send to the model, after policies, because a blocked tool is not in the prompt.

The cascade, the list filter and the bulk switch are in the runtime (src/js/adminvora.js), so they work with the dist/ bundle. The masked secret is demo page behaviour from src/demo/demo.js, like the rest of the account pages' secrets.

Documentation for @creativeigniter/adminvora