@modelcontextprotocol/inspector 2.3.0 → 2.4.0

package/README.md

@@ -30,10 +30,13 @@ inspector/
 │   ├── tui/          # TUI client (Ink + React, tsup bundle)
 │   └── launcher/     # Shared launcher — provides the `mcp-inspector` bin, dispatches to web/cli/tui
 ├── core/             # Shared code consumed via the `@inspector/core` alias (no package.json)
-│   ├── auth/         # OAuth: providers, discovery, storage, endpoint overrides, mid-session recovery (browser/node/remote backends)
+│   ├── auth/         # OAuth: providers, discovery, storage, endpoint overrides, mid-session recovery (browser/node/remote backends);
+│   │                 #   plus per-server secret storage — the keychain/file/memory SecretStore
+│   │                 #   implementations, the selection policy, and the descriptor the banner and UI report
 │   ├── client/       # Install-level client config (`client.json`): browser-safe parse/validate + Node load/save, remote backend, secrets
-│   ├── json/         # JSON + parameter/argument conversion utilities, and the nullable-union
-│   │                 #   schema collapse shared by the web and TUI form builders
+│   ├── json/         # JSON + parameter/argument conversion utilities, the nullable-union
+│   │                 #   schema collapse shared by the web and TUI form builders, and the
+│   │                 #   tool-schema portability lint all three clients report from
 │   ├── logging/      # Silent pino logger singleton
 │   ├── mcp/          # InspectorClient runtime, state stores, transports, config import,
 │   │                 #   and the RFC 6570 URI-template helpers the web form and TUI expand through
@@ -137,9 +140,11 @@ This is what lets an Inspector connection negotiating `protocolEra: "auto" | "mo
 
 Each config below is a ready-made server for exercising one feature by hand. Load one with `--config`, and unless noted, connect with **Protocol Era = Modern**.
 
-| Config                                    | Demonstrates                                       | Issue                                                                  |
-| ----------------------------------------- | -------------------------------------------------- | ---------------------------------------------------------------------- |
+| Config                                    | Demonstrates                                        | Issue                                                                  |
+| ----------------------------------------- | --------------------------------------------------- | ---------------------------------------------------------------------- |
 | `mcp-app-http.json` **(legacy era)**      | An MCP App (UI resource + app tool) in the Apps tab | [#1859](https://github.com/modelcontextprotocol/inspector/issues/1859) |
+| `app-elicitation-http.json` **(legacy era)** | An MCP App rendering a form elicitation           | [#1854](https://github.com/modelcontextprotocol/inspector/issues/1854) |
+| `mcp-app-domain-http.json` **(legacy era)** | An MCP App asking for a dedicated origin (`_meta.ui.domain`) | [#2056](https://github.com/modelcontextprotocol/inspector/issues/2056) |
 | `modern-mrtr-http.json`                   | A single MRTR round-trip                           | —                                                                      |
 | `mrtr-showcase-http.json`                 | Every MRTR preset in one server                    | [#1860](https://github.com/modelcontextprotocol/inspector/issues/1860) |
 | `modern-network-http.json`                | Network tab: `Mcp-*` headers + error taxonomy      | [#1628](https://github.com/modelcontextprotocol/inspector/issues/1628) |
@@ -148,11 +153,13 @@ Each config below is a ready-made server for exercising one feature by hand. Loa
 | `structured-output-http.json`             | Tools tab: a result's `structuredContent` section  | [#1908](https://github.com/modelcontextprotocol/inspector/issues/1908) |
 | `duplicate-tool-names-http.json`          | A `tools/list` that repeats a tool name            | [#1957](https://github.com/modelcontextprotocol/inspector/issues/1957) |
 | `nullable-fields-http.json`               | Tools tab: nullable (`anyOf` + `null`) arguments   | [#1928](https://github.com/modelcontextprotocol/inspector/issues/1928) |
+| `unportable-schemas-http.json` **(legacy era)** | Tool schemas a real client rejects, flagged in all three clients | [#1005](https://github.com/modelcontextprotocol/inspector/issues/1005) |
 | `rfc6570-templates-http.json`             | Resources tab: RFC 6570 resource-template expansion | [#1919](https://github.com/modelcontextprotocol/inspector/issues/1919) |
-| `advertised-extensions-http.json`         | Tool registration gated on advertised extensions   | [#1739](https://github.com/modelcontextprotocol/inspector/issues/1739) |
-| `logging-{legacy,modern}-http.json`       | Logging, both eras                                 | [#1629](https://github.com/modelcontextprotocol/inspector/issues/1629) |
-| `subscriptions-{legacy,modern}-http.json` | Resource subscriptions, both eras                  | [#1630](https://github.com/modelcontextprotocol/inspector/issues/1630) |
-| `tasks-{legacy,modern}-http.json`         | Tasks, both eras                                   | [#1631](https://github.com/modelcontextprotocol/inspector/issues/1631) |
+| `advertised-extensions-http.json`         | Tool registration gated on advertised extensions    | [#1739](https://github.com/modelcontextprotocol/inspector/issues/1739) |
+| `oauth-custom-resource-metadata-http.json` **(legacy era)** | OAuth discovery driven by the challenge's `resource_metadata` | [#2071](https://github.com/modelcontextprotocol/inspector/issues/2071) |
+| `logging-{legacy,modern}-http.json`       | Logging, both eras                                  | [#1629](https://github.com/modelcontextprotocol/inspector/issues/1629) |
+| `subscriptions-{legacy,modern}-http.json` | Resource subscriptions, both eras                   | [#1630](https://github.com/modelcontextprotocol/inspector/issues/1630) |
+| `tasks-{legacy,modern}-http.json`         | Tasks, both eras                                    | [#1631](https://github.com/modelcontextprotocol/inspector/issues/1631) |
 
 #### MCP Apps
 
@@ -162,6 +169,35 @@ Open the Apps tab, select `mcp_app_demo`, give it a title and click **Open App**
 
 For the scripted version of the same flow (`--app-info` probe → deep link → rendered widget), see [Reviewing an MCP App](./docs/mcp-app-review.md).
 
+#### An App's dedicated origin
+
+`mcp-app-domain-http.json` serves the same `mcp_app_demo` widget as above, with one addition: its UI resource declares `_meta.ui.domain`. Plain streamable-HTTP; connect with the **default (legacy)** protocol era.
+
+That field is how a server asks its host for a stable, dedicated origin. Without one, an App renders into a `srcdoc` frame sandboxed without `allow-same-origin`, so its document has an *opaque* origin and every request it makes carries `Origin: null` — which no CORS policy, OAuth callback, or API-key allowlist can admit ([#2056](https://github.com/modelcontextprotocol/inspector/issues/2056)).
+
+Open the Apps tab and run `mcp_app_demo`. The widget renders identically to `mcp-app-http.json` — the difference is not visual. Inspect the inner iframe in devtools: on this server it is served from `http://127.0.0.1:6278/app-document/<id>` and `location.origin` is that real origin, where on `mcp-app-http.json` it is `about:srcdoc` with an origin of `null`. (The host is whatever the Inspector bound to — `127.0.0.1` by default, an *address* rather than the name `localhost`, for the reason `resolve-bind-host.ts` documents.)
+
+The spec makes `domain`'s format **host-dependent**, and the Inspector owns no domain infrastructure — so it reads any non-empty value as a *request* rather than an address, and answers with a real loopback origin of its own. See [MCP App dedicated origins](./clients/web/README.md#mcp-app-dedicated-origins-metauidomain) for the full contract, including what the one shared origin does and does not isolate, and how every failure falls back to the default render rather than blanking the app.
+
+#### App-rendered form elicitations
+
+`app-elicitation-http.json` serves `app_choose_option` alongside the `choose_option_app` UI resource (`ui://demo/choose-option.html`). The tool sends a completely ordinary form `elicitation/create` — the only thing added is `_meta.ui.resourceUri` naming that app. Plain streamable-HTTP; connect with the **default (legacy)** protocol era.
+
+Run `app_choose_option` from the Tools tab. The server's app renders in a modal instead of the built-in elicitation form, and clicking **Option A**, **Decline**, or **Cancel** returns the standard `ElicitResult` straight to the server, which echoes it into the tool result.
+
+App rendering is selected only when **all four** conditions hold ([#1854](https://github.com/modelcontextprotocol/inspector/issues/1854), per [ext-apps#733](https://github.com/modelcontextprotocol/ext-apps/pull/733) / SEP-3118):
+
+1. the client advertises `elicitation.form`;
+2. the client advertises `extensions["io.modelcontextprotocol/ui"].mimeTypes` including `text/html;profile=mcp-app`;
+3. **both** the client and the server advertise the nested `elicitation` setting on that same extension;
+4. the request carries a valid absolute `ui://` URI in `_meta.ui.resourceUri`.
+
+Only the **web** client advertises the nested client-side setting, and only because it has a sandbox renderer to back it; the CLI and TUI advertise the MIME type (they know what an App is) but never claim they can resolve an elicitation through one, so the same server falls back to their native prompts. Turning **Server Settings → Advertised Extensions → MCP Apps UI** off, or turning form elicitation off, removes the claim on web too.
+
+Everything else falls back to the built-in elicitation form, by design: metadata that is absent or not an absolute `ui://` URI, a resource that fails to load, a sandbox or bridge that fails to initialize, an app that did not advertise `elicitation`, a request that times out, and any result that is not a valid `ElicitResult` for the requested schema. An explicit `decline` or `cancel` is **not** a fallback — it is a completed elicitation and goes back to the server as-is.
+
+> The Inspector speaks the ext-apps#733 wire protocol but does not yet consume its helpers: the released `@modelcontextprotocol/ext-apps` (1.7.5) predates that PR. `core/mcp/appElicitation.ts` and `clients/web/src/components/elements/AppRenderer/requestAppElicitation.ts` mirror it exactly and are marked for deletion in favour of the package's own exports once a release containing it ships.
+
 #### MRTR
 
 `modern-mrtr-http.json` serves the `mrtr_confirm` tool (preset `mrtr_confirm`, `createMrtrTool`) over the modern leg. Its handler returns `inputRequired(...)` embedding a form elicitation, so invoking it produces a real round-trip: `input_required` → the client fulfils the embedded elicitation and retries with a new id → `complete`.
@@ -252,6 +288,62 @@ Open the Tools tab and select `record_shipment`: `direction` must render as a **
 
 The **TUI** had the same gap and is worth checking against the same server (`--tui`, then test `record_shipment`): `direction` is a select, `quantity` an integer field, `express` a boolean. Both clients now share one collapse step — `normalizeNullableUnion` in [`core/json/nullableUnion.ts`](./core/json/nullableUnion.ts) — precisely so they cannot drift on which schemas they can render.
 
+#### Unportable tool schemas
+
+`unportable-schemas-http.json` serves four tools, three of whose advertised
+schemas carry constructs that are legal JSON Schema and are refused or
+mishandled by real MCP clients:
+
+| Tool | What it carries |
+| --- | --- |
+| `get_temp` | `outputSchema.properties.data` as a bare `true` — what Go's `jsonschema` package emits for `interface{}`, the case reported in [#1005](https://github.com/modelcontextprotocol/inspector/issues/1005) |
+| `echo` | an array-form `"type": ["null","boolean"]`, and an `opts` property that constrains nothing |
+| `add` | a property pointing at a remote `$ref` |
+| `get_weather` | nothing — left clean, so a flagged tool sits beside an unflagged one |
+
+Plain streamable-HTTP — connect with the **default (legacy)** protocol era.
+Every tool here is still **runnable**: the override replaces only the
+*advertised* schema, and the flagship bare-`true` rides `get_temp` (which
+returns structured content) rather than `echo` — see the ⚠️ caveat at the end
+of this section for why that placement matters.
+
+The Inspector is where a server author looks first, so a construct that will
+fail downstream is named here rather than passed through silently. All three
+clients report the same verdict from
+[`core/json/schemaLint.ts`](./core/json/schemaLint.ts), each with the room it
+has:
+
+```bash
+mcp-inspector --cli http://127.0.0.1:6603/mcp --method tools/list --strict   # exits 6
+```
+
+- **CLI** — `--strict` prints the full report (path, issue, suggested fix) on
+  stderr and exits `6` on an error-severity finding; without it, one summary
+  line. See [Schema portability](./clients/cli/README.md#schema-portability---strict).
+- **TUI** — the tools list marks `get_temp` with a red `!` and `echo`/`add`
+  with a yellow `?`; the detail pane lists each finding under **Schema
+  Portability**.
+- **Web** — the Tools sidebar row carries the same flag as a hover-labelled
+  icon, and selecting the tool shows a **Schema portability** section above the
+  argument form.
+
+This is deliberately **not** a JSON Schema validator. A census of 617 public
+servers (14,804 tool schemas) reported on that issue found **0** that fail the
+SDK's own `ListToolsResultSchema.safeParse`, so a conformance check would
+report nothing on essentially every real server. What bites is the narrower
+subset each consumer accepts, and each rule here is a construct known to be
+refused or degraded by a shipping client. The schemas are supplied through the
+test server's `rawToolSchemas` override, because the Zod-built presets cannot
+express any of them — which is the same reason a real server hits this only
+when its schemas come from another generator.
+
+⚠️ **If you add an `outputSchema` override of your own, put it on a tool that
+returns structured content.** A conforming client validates a tool result
+against the advertised output schema, so an override on a preset that returns
+none makes every call to it fail with "declares an output schema but returned
+no structured content" — a confusing thing to hit from a fixture. That is why
+the bare `true` rides `get_temp` here and not `echo`.
+
 #### RFC 6570 resource templates
 
 `rfc6570-templates-http.json` serves two resource templates straight out of [#1919](https://github.com/modelcontextprotocol/inspector/issues/1919) — `events_by_topic` (`foobar://events/{topic}`) and `events_by_query` (`foobar://events{?topic}`) — each echoing the URI it was matched against, plus a plain `foobar://events` resource (see below). Plain streamable-HTTP; connect with the **default (legacy)** protocol era.
@@ -264,23 +356,23 @@ Open the Resources tab and pick **events_by_topic**, then enter `foo/bar`. The r
 
 The web client and the TUI expand through one shared helper, [`core/mcp/uriTemplate.ts`](./core/mcp/uriTemplate.ts) — the web Resources form directly, the TUI via `InspectorClient.readResourceFromTemplate` — and both derive their **form fields** from its parser too, which is the half that makes the sharing real: a form submits values under the names it rendered, so a parser that mangles a name silently drops the value at expansion time. (The CLI is not a consumer: it has no template form, and its `resources/read` passes the already-expanded `--uri` straight through.)
 
-The SDK's `UriTemplate` is still used, but only to *validate* a template (constructing it is what rejects an unclosed expression). Its expander is not, because it is incomplete in five ways — each measured against the pinned SDK, not inferred:
+The SDK's `UriTemplate` is still used, but only to _validate_ a template (constructing it is what rejects an unclosed expression). Its expander is not, because it is incomplete in five ways — each measured against the pinned SDK, not inferred:
 
-| Shape | SDK behavior |
-| --- | --- |
-| `{a,b}` | raw-joins the values — no encoding, operator prefix dropped |
-| `{;id}` | `;` is missing from its operator list, so the variable parses as `;id` |
-| `{id:3}` | the prefix modifier is folded into the name, giving `id:3` |
+| Shape           | SDK behavior                                                                                                   |
+| --------------- | -------------------------------------------------------------------------------------------------------------- |
+| `{a,b}`         | raw-joins the values — no encoding, operator prefix dropped                                                    |
+| `{;id}`         | `;` is missing from its operator list, so the variable parses as `;id`                                         |
+| `{id:3}`        | the prefix modifier is folded into the name, giving `id:3`                                                     |
 | `{+v}` / `{#v}` | `encodeURI` mangles reserved `[`/`]` (`[::1]` → `%5B::1%5D`) and double-encodes pct-triplets (`%2F` → `%252F`) |
-| `{v}` | `encodeURIComponent` leaves the sub-delims `!'()*` bare, which RFC 6570 requires encoded |
+| `{v}`           | `encodeURIComponent` leaves the sub-delims `!'()*` bare, which RFC 6570 requires encoded                       |
 
 The `;` and `:3` rows are the ones a user sees directly: on the SDK's parse the form renders fields literally labelled `;id` and `id:3`. The `+`/`#` row is silent corruption rather than over-escaping — an IPv6 literal or an already-encoded path arrives at the server altered.
 
 A template that cannot be expanded at all — an out-of-grammar modifier (`{id:abc}`), or an expression declaring no variable (`{}`, `{a,}`, `{?}`) — **withholds the read** rather than sending something. Pick **events_malformed** (`foobar://events/{topic:abc}`) to see it: Read Resource is disabled, the reason is printed under the form, and the preview shows the template as the server declared it. The alternative is worse than it looks: `x://{}` would otherwise expand to `x://` with no inputs rendered, so the form's "everything required is filled" check passes vacuously and it reads a URI that is not the template the server published.
 
-Literals are pct-encoded on expansion too (RFC 6570 §3.1): `café/{var}` sends `caf%C3%A9/value`, not raw UTF-8 in the path — something the SDK's expander does not do either. And the *names* a template may use are RFC 6570's `varchar` plus a labelled tolerance for `-` and `~`: the conformance suite rejects `{default-graph-uri}`, but real servers publish such names and the SDK's matcher round-trips them, so the Inspector expands them and marks the variable `conforming: false` rather than refusing a resource that demonstrably works.
+Literals are pct-encoded on expansion too (RFC 6570 §3.1): `café/{var}` sends `caf%C3%A9/value`, not raw UTF-8 in the path — something the SDK's expander does not do either. And the _names_ a template may use are RFC 6570's `varchar` plus a labelled tolerance for `-` and `~`: the conformance suite rejects `{default-graph-uri}`, but real servers publish such names and the SDK's matcher round-trips them, so the Inspector expands them and marks the variable `conforming: false` rather than refusing a resource that demonstrably works.
 
-An **undefined** variable is what omits its expression — a variable defined as the empty string expands (`x{?q}` gives `x?q=`, `x{;q}` gives `x;q`, per RFC 6570 §3.2.7). The expander honors that distinction, so a caller such as `readResourceFromTemplate` can request either URI. Collapsing the two is a *form* concern, not a template one: both clients seed every declared variable with `""` and a text input cannot express "defined but empty", so each form drops its blanks (`definedValues`) on the way in.
+An **undefined** variable is what omits its expression — a variable defined as the empty string expands (`x{?q}` gives `x?q=`, `x{;q}` gives `x;q`, per RFC 6570 §3.2.7). The expander honors that distinction, so a caller such as `readResourceFromTemplate` can request either URI. Collapsing the two is a _form_ concern, not a template one: both clients seed every declared variable with `""` and a text input cannot express "defined but empty", so each form drops its blanks (`definedValues`) on the way in.
 
 Requiredness is a property of the **expression**, not the variable: RFC 6570 drops undefined names from a multi-name expression, so `{a,b}` with only `a` filled is expandable and a form must not block it. `requiredGroups` returns one entry per non-omittable expression and `hasRequiredValues` asks that each be satisfied by any one of its names — which no per-variable flag can express once a name recurs across expressions (`{a,b}{a,c}` is satisfied by filling `b` and `c`).
 
@@ -294,6 +386,23 @@ Requiredness is a property of the **expression**, not the variable: RFC 6570 dro
 
 This is the debugging knob for a server legitimately changing tool registration based on what the client advertises. Legacy stateful leg only — the modern per-request leg has no persistent `oninitialized`.
 
+#### OAuth `resource_metadata` at a non-default path
+
+`oauth-custom-resource-metadata-http.json` is an OAuth-protected server (combined AS + resource, DCR enabled) that serves its RFC 9728 protected-resource metadata document **only** from `/custom/protected-resource`, and advertises it on every 401:
+
+```http
+HTTP/1.1 401 Unauthorized
+WWW-Authenticate: Bearer resource_metadata="http://127.0.0.1:8082/custom/protected-resource"
+```
+
+The default `/.well-known/oauth-protected-resource` route is deliberately left unserved, so a client that ignores the advertised URL cannot discover the document at all. Plain streamable-HTTP — connect with the **default (legacy)** protocol era.
+
+Add the server, click **Connect**, and watch the Inspector's first protected-resource metadata request in the Network tab: it must go to `/custom/protected-resource`. On the broken build the challenge's `resource_metadata` was parsed and then dropped before the SDK's `auth()` ever saw it ([#2071](https://github.com/modelcontextprotocol/inspector/issues/2071)), so discovery probed locations derived from the MCP server URL, 404'd, and authorization failed for any server that puts the document somewhere other than the well-known path.
+
+The same server is worth running against `--cli` / `--tui`, which reach it by a different route: with no stored token in the legacy era the Inspector connects with no auth provider (so the SDK cannot open a browser before the callback server is listening), the 401 surfaces as the SDK's headerless `UnauthorizedError`, and the client calls `authenticate()` with no challenge in hand. The transport therefore *observes* every 401/403 passively, so the advertised URL is still available on that path.
+
+The value now rides the normalized `AuthChallenge` as a string — it has to be serializable, because the web client's challenge crosses the remote-backend boundary as JSON — and is converted to a `URL` at the OAuth boundary, where it is handed to `auth()` as `resourceMetadataUrl` and to the CIMD pre-registration probe, which runs *before* `auth()` and would otherwise do its own default-location discovery. A malformed value is ignored rather than surfaced, matching the SDK's own `WWW-Authenticate` parser: discovery falls back to the default locations instead of failing the whole authorization on a bad header. The callback leg needs nothing extra — SDK `auth()` persists the URL in its discovery state, so it survives both the web full-page redirect and the CLI/TUI loopback callback.
+
 #### Logging, both eras
 
 `logging-legacy-http.json` and `logging-modern-http.json` both serve `logging: true` plus a `send_notification` tool that emits a `notifications/message` at a chosen level. The legacy one is a plain streamable-HTTP server; the modern one sets `transport.modern: true`.
@@ -343,19 +452,22 @@ Each client self-validates from its own folder; the root scripts chain them. The
 | ----------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `npm run validate`                  | Runs the three durable guards first — `verify:format-coverage` (every tracked source file is format-gated), `verify:typecheck-coverage` (every one lands in a tsconfig project), `verify:dep-lockstep` (no dependency reaching one `tsc` program from two installs skews across them) — then `test:scripts` (the guards' own parser unit tests), then `validate:core` (the shared `core/` `format:check` + `lint` gate), then per client: `format:check` + `lint` + **`typecheck`** (cli/tui/launcher; web typechecks via `tsc -b` inside its `build`) + `build` + fast unit tests. The quick inner-loop check.                                                                                                                                                                                                                                                                                                                                                                                |
 | `npm run coverage`                  | The **per-file ≥90% gate** (lines/statements/functions/branches) under v8 instrumentation, per client. CI-enforced. For web this also runs the integration project and covers the shared `core/` runtime (including `core/json` and `core/client`).                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                      |
-| `npm run smoke`                     | End-to-end smokes through the built launcher (`--help` dispatch + prod cli/tui/web), plus two headless-Chromium smokes: a boot smoke that runs the prod web bundle and asserts a clean first render (no uncaught error — sync exception or unhandled rejection, how a Node built-in reaching the browser bundle manifests), and an **MCP Apps** smoke (`smoke:web:app`) that drives connect → open app → `data-app-status="ready"` against a composable App server, covering the sandbox proxy and UI-protocol bridge.                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
+| `npm run smoke`                     | End-to-end smokes through the built launcher (`--help` dispatch + prod cli/tui/web), plus three headless-Chromium smokes: a boot smoke that runs the prod web bundle and asserts a clean first render (no uncaught error — sync exception or unhandled rejection, how a Node built-in reaching the browser bundle manifests), and an **MCP Apps** smoke (`smoke:web:app`) that drives connect → open app → `data-app-status="ready"` against a composable App server, covering the sandbox proxy and UI-protocol bridge, and an **app-rendered elicitation** smoke (`smoke:web:elicit`) that drives one end to end — call the tool, answer inside the sandboxed app, see the app's `ElicitResult` reach the server — and then the same tool against a server that never advertised the capability, which must fall back to the native elicitation form.                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
 | `npm run verify:build-gate`         | Runs a real `vite build` with a Node built-in forced into the browser graph and asserts the build **fails** via the #1769 gate (which turns Vite's browser-externalization warning into a hard error). Guards against the warning phrasing drifting in a Vite bump and silently disabling the gate. Part of `npm run ci`.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
+| `npm run verify:bundle-externals`   | Guards the must-not-bundle invariant (#2067): for each tsup-bundled client it reads the **built** `build/` output and fails if any package that must stay external was inlined anyway. Candidates are the union of the client's own `external` array **and the root manifest's `dependencies`** — the latter because #2067 was a *missing* `external` entry, which a self-referential check would have passed. Detection is via esbuild's `// <path>` module banners, so it covers both shapes — a separate `<pkg>-HASH.js` chunk (what a dynamically `import()`ed CommonJS package produces) and a statically-imported package folded straight into `index.js`, which emits no chunk at all. A build with no banners fails as such rather than passing clean, so enabling `minify` cannot silently retire the check. `undici` was declared only in the root and `clients/cli` manifests, and tsup auto-externalizes only what the _nearest_ manifest declares, so the web and TUI bundles inlined 1.05MB of it — and CommonJS inlined into an ESM bundle throws `Dynamic require of "assert" is not supported` on first use, from a specifier no user-side install can satisfy. Reads the output rather than the config because those two disagreed for four releases. Part of `npm run ci`. |
 | `npm run verify:format-coverage`    | Parses the `format:check` globs out of every `package.json` (only those reachable from `validate`), enumerates all tracked source files, and **fails** listing any not covered by a glob — the durable guard for the "every first-party source file is format-gated" invariant (#1792). Runs first in `validate`.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
-| `npm run test:scripts`              | Table-driven unit tests (`node --test`) for the guard's own pure parsers (`scripts/lib/npm-scripts.mjs`, `scripts/lib/tsc-program.mjs` + the exported helpers of `verify-typecheck-coverage.mjs` and `verify-dep-lockstep.mjs`), one case per rule they encode, plus `scripts/lib/resolve-node-bin.test.mjs` — the cross-platform bin resolver (#1939), pinned against the real `bin`/`exports` shapes of the packages the scripts actually spawn. Runs in `validate` — and `verify:typecheck-coverage` guards *this* gate in turn (reachable from `validate`, non-empty test set, every test file matched by the `test:scripts` glob), since `node --test` silently skips a file its glob misses and still exits 0.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       |
+| `npm run test:scripts`              | Table-driven unit tests (`node --test`) for the guard's own pure parsers (`scripts/lib/npm-scripts.mjs`, `scripts/lib/tsc-program.mjs` + the exported helpers of `verify-typecheck-coverage.mjs` and `verify-dep-lockstep.mjs`), one case per rule they encode, plus two suites over shared `scripts/lib` helpers that no smoke can check itself: `resolve-node-bin.test.mjs` — the cross-platform bin resolver (#1939), pinned against the real `bin`/`exports` shapes of the packages the scripts actually spawn — and `announced-child.test.mjs` — the spawn/readiness ownership helper (#2000), which drives real `node -e` children to prove a child that never announces is still published to the caller before the timeout throws, and so is reachable by teardown rather than orphaned. Two more do the same: `mcp-app-flow.test.mjs` covers the shared MCP Apps flow (#2003) — the deep link's two CSRF gates and `appArgs` encoding, plus `driveAppFlow`'s failure branches against a stand-in page, all of which are dead code from the happy-path smokes' point of view and would otherwise surface only as opaque timeouts; and `ensure-test-servers.test.mjs` pins the [#2111](https://github.com/modelcontextprotocol/inspector/issues/2111) invariant — that `test-servers/build` is rebuilt **even when it already exists** — which no smoke can assert about itself, since one driving a stale fixture reports a product failure rather than a staleness one. Runs in `validate` — and `verify:typecheck-coverage` guards *this* gate in turn (reachable from `validate`, non-empty test set, every test file matched by the `test:scripts` glob), since `node --test` silently skips a file its glob misses and still exits 0.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       |
 | `npm run verify:typecheck-coverage` | The typecheck-coverage analog of the above (#1791): for each Node client (auto-discovered from disk — enrolled via its `typecheck` script's projects, or for a `tsc -b` client like `clients/web` via its `tsconfig.json` `references`) it runs those projects with `tsc --listFilesOnly`, unions them, and **fails** listing any tracked `.ts`/`.tsx`/`.mts`/`.cts` under the client that lands in no project (so a new top-level config/helper can't silently go untypechecked). It also requires, deny-by-default, the first-party TS no client owns (`test-servers/src`, the root `vitest.shared.mts`, all of `core/`, and any new top-level location) to land in some client project's tsc pass — so a `core` `*.tsx` web's projects don't reach is caught too. Also asserts the gate is wired (each client's typecheck pass — its `typecheck` script, or web's `tsc -b` — is reachable from its `validate`, and the root chain runs each client's `validate`). Runs in `validate`. |
 | `npm run verify:dep-lockstep`       | Guards the "one version per install-crossing dependency" invariant (#1896). v2 is not a workspace, so a client's test project compiles the shared first-party TypeScript — `core/`, `test-servers/src`, and the root-owned `vitest.shared.mts`, all of which resolve their dependencies from the **root** install — alongside the client's own sources, putting the same package in one `tsc` program twice. At the same version that's harmless; skewed, TypeScript must relate two structurally-distinct copies of every type, which for a recursive-generic surface is exponential (zod `4.3.6` vs `4.4.3` exhausted the 4GB tsc heap in `clients/web`). Derives its candidate set from **what actually enters each program** (#1965) — every client tsconfig project listed with `tsc --listFilesOnly` via the shared `scripts/lib/tsc-program.mjs`, each resolved `node_modules` file mapped to its owning install, keeping the packages that reach one program from two installs (a package whose declarations arrive only through another package's `.d.ts`, as `@modelcontextprotocol/sdk`'s do, is invisible to a scan of first-party imports). Prices each copy from the lockfile entry for the exact install path the program resolved, compares only the installs that met in one program, and **fails deny-by-default** on any disagreement not in the annotated `TOLERATED_SKEW` allowlist — empty today — with an allowlisted package tolerated only *within a major version*. Runs in `validate`.
-| `npm run ci`                        | **Mandatory pre-push command.** `validate` → `coverage` → `verify:build-gate` → `smoke` → Storybook. A true superset of GitHub CI.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       |
+| `npm run ci`                        | **Mandatory pre-push command.** `validate` → `coverage` → `verify:build-gate` → `verify:bundle-externals` → `smoke` → Storybook. A true superset of GitHub CI.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       |
 | `npm run pack:verify`               | Publish smoke — see [Publishing](#publishing).                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
 
 Per-client scripts exist too (`validate:web`, `coverage:cli`, `smoke:tui`, …), plus root `validate:core` / `format:core` for the shared `core/` package, `format:scripts` for the root `scripts/` tooling, and `format:shared` / `lint:shared` for the root "shared" surface (`test-servers/src/**`, `vitest.shared.mts`, the root `eslint.config.js`). Run `npm run format` before committing — the root `format` fixes `core/`, the root `scripts/`, the shared surface, and every client; `validate` runs the non-fixing `format:check` and fails CI on any unformatted file.
 
 **Linting is type-aware.** All five ESLint scopes (`clients/{web,cli,tui,launcher}` plus the root `core/` + shared gate) enable `@typescript-eslint/no-floating-promises` at `error`, so a promise that is neither awaited, returned, `.catch(…)`-terminated, nor explicitly discarded with `void` fails `lint` — and therefore `validate` ([#1959](https://github.com/modelcontextprotocol/inspector/issues/1959)). The rule needs type information, so each scope's config names a parser project; the root scope's is **`tsconfig.lint.json`**, a lint-only project covering `core/**`, `test-servers/src/**`, and `vitest.shared.mts`, which have no tsconfig of their own. It emits nothing and changes no typecheck — but a new first-party TS location added to the root lint scope must be added to its `include`. See **TypeScript instructions** in [`AGENTS.md`](./AGENTS.md) for when `void` is acceptable.
 
+**And lint has no warning tier.** Every `lint` script runs with `--max-warnings 0`, so a warning fails `validate` exactly as an error does ([#2085](https://github.com/modelcontextprotocol/inspector/issues/2085)) — a `warn`-level `react-hooks/exhaustive-deps` finding had otherwise let a stale-closure bug pass the mandatory pre-push gate and reach review. Fix the finding rather than silencing it; if a rule genuinely must be waived, use its inline disable comment with a one-line justification.
+
 For the full testing rules — the ≥90% per-file gate, where test files live, the unit vs. integration vs. storybook projects, and the `v8 ignore` policy — see [`AGENTS.md`](./AGENTS.md).
 
 ## Publishing
@@ -368,13 +480,13 @@ The root `package.json` `"files"` allowlist is the source of truth for the tarba
 
 - **No source maps.** The client bundlers set `sourcemap: false` (`clients/{cli,tui}/tsup.config.ts`, `clients/web/tsup.runner.config.ts`); Vite and the launcher's `tsc` already emit none. Maps are ~half the unpacked size and aren't needed at runtime — debug via `npm run dev` on the source.
 - **`clients/web/build` ships via `clients/web/.npmignore`.** `clients/web/.gitignore` lists `build/`, and npm's packlist honors that nested `.gitignore` over the root `"files"` allowlist — so the prod web-server runner was silently missing from the tarball while `clients/web/dist` slipped through (its `.gitignore` only lists `dist-ssr`). `clients/web/.npmignore` overrides the `.gitignore` for publishing so both `build/` (runner) and `dist/` (SPA) ship. The other clients don't need this — none ship a nested `.gitignore`.
-- **`clients/web/static` ships the MCP Apps sandbox proxy.** `clients/web/static/sandbox_proxy.html` is a committed source file (not a build artifact), read from disk at runtime by `clients/web/server/sandbox-controller.ts` as `<runner dir>/../static/sandbox_proxy.html`. It was missing from the root `"files"` allowlist entirely, so every published build failed the Apps tab with **"Sandbox not loaded"** ([#1859](https://github.com/modelcontextprotocol/inspector/issues/1859)) while working fine in the repo. Because the path is resolved _relative to_ `clients/web/build`, the directory must ship at that exact location — `pack:verify` asserts both the tarball entry and the installed-on-disk path.
-- **A dependency that renders React is bundled, not externalized.** An externalized package resolves its own `react` from wherever npm placed **it** in the consumer's tree, which is not necessarily where the bundle resolves ours — npm places a package beside a React satisfying *its* peer range, and those ranges are looser than ours. `ink-form` and `ink-scroll-view` declare `">=18"`, so a project holding React 18 satisfies them and gets them hoisted while the Inspector's React 19 nests underneath: two React copies, and the TUI dies with `TypeError: Cannot read properties of null (reading 'useState')` the moment a tool test form or a scroll view mounts ([#1952](https://github.com/modelcontextprotocol/inspector/issues/1952)). Both are therefore inlined by `clients/tui/tsup.config.ts` and are **not** root dependencies: the tarball ships their code inside `clients/tui/build/index.js` rather than having consumers install them. Bundling also pins their transitive deps to what this repo's install resolved (notably `ink-select-input@6` via `overrides`, which npm ignores for a package installed as a dependency). **`ink` is the one exception, on cost:** bundling it works but adds ~1.4 MB (`react-reconciler` and `yoga-layout` come along, plus a `createRequire` banner for the inlined CJS), so it stays external — *not* because its `">=19"` peer makes it safe, which it does not. What keeps that tolerable is the root `react` range: `"^19.0.0"` is deliberately open to the whole major so npm can dedupe our React with whatever React 19 a consumer pins, leaving an external `ink` on the same copy the bundle uses. **Narrowing that range reopens the bug for the renderer itself** — `clients/tui/__tests__/tsupConfig.test.ts` pins it to `ink`'s peer floor, and guards the rest of the split; see the [TUI README](./clients/tui/README.md#bundling-react-rendering-dependencies-must-be-inlined-1952).
+- **`clients/web/static` ships the MCP Apps sandbox proxy.** `clients/web/static/sandbox_proxy.html` is a committed source file (not a build artifact), read from disk at runtime by `clients/web/server/sandbox-controller.ts` as `<runner dir>/../static/sandbox_proxy.html`. It was missing from the root `"files"` allowlist entirely, so every published build failed the Apps tab with **"Sandbox not loaded"** ([#1859](https://github.com/modelcontextprotocol/inspector/issues/1859)) while working fine in the repo. Because the path is resolved _relative to_ `clients/web/build`, the directory must ship at that exact location — `pack:verify` asserts the tarball entry, the installed-on-disk path, and (since [#2003](https://github.com/modelcontextprotocol/inspector/issues/2003)) that a widget actually *loads through* it on the installed bin. Presence and reachability are different properties: a rename with a stale reader ships a file that is there and unusable.
+- **A dependency that renders React is bundled, not externalized.** An externalized package resolves its own `react` from wherever npm placed **it** in the consumer's tree, which is not necessarily where the bundle resolves ours — npm places a package beside a React satisfying _its_ peer range, and those ranges are looser than ours. `ink-form` and `ink-scroll-view` declare `">=18"`, so a project holding React 18 satisfies them and gets them hoisted while the Inspector's React 19 nests underneath: two React copies, and the TUI dies with `TypeError: Cannot read properties of null (reading 'useState')` the moment a tool test form or a scroll view mounts ([#1952](https://github.com/modelcontextprotocol/inspector/issues/1952)). Both are therefore inlined by `clients/tui/tsup.config.ts` and are **not** root dependencies: the tarball ships their code inside `clients/tui/build/index.js` rather than having consumers install them. Bundling also pins their transitive deps to what this repo's install resolved (notably `ink-select-input@6` via `overrides`, which npm ignores for a package installed as a dependency). **`ink` is the one exception, on cost:** bundling it works but adds ~1.4 MB (`react-reconciler` and `yoga-layout` come along, plus a `createRequire` banner for the inlined CJS), so it stays external — _not_ because its `">=19"` peer makes it safe, which it does not. What keeps that tolerable is the root `react` range: `"^19.0.0"` is deliberately open to the whole major so npm can dedupe our React with whatever React 19 a consumer pins, leaving an external `ink` on the same copy the bundle uses. **Narrowing that range reopens the bug for the renderer itself** — `clients/tui/__tests__/tsupConfig.test.ts` pins it to `ink`'s peer floor, and guards the rest of the split; see the [TUI README](./clients/tui/README.md#bundling-react-rendering-dependencies-must-be-inlined-1952).
 - **A single version number, read from the root `package.json`.** The Inspector ships as one package with one version, so only the **root** `package.json` carries a `version` — the four `clients/*/package.json`s deliberately have none. Every Node client (CLI, TUI, and the web backend) resolves the version through the shared `readInspectorVersion()` reader in `core/node/version.ts`, which walks up to the root manifest (always present in the tarball). No client `package.json` is read at runtime, so none needs to ship. The web **browser** can't read the filesystem; it gets its version from the backend via `GET /api/config` (see [#1639](https://github.com/modelcontextprotocol/inspector/issues/1639)).
 
 ### `npm run pack:verify` — publish smoke against the real tarball
 
-The `smoke:*` scripts run against the in-repo build tree, which is **not** the published package. `npm run pack:verify` (`scripts/pack-and-verify.mjs`) closes that gap: it builds, `npm pack`s the publishable tarball (asserting no source maps ship and that the runtime-required files are present), installs the tarball into a **clean throwaway consumer** — a fresh temp directory where it runs a real `npm install <tgz>` (pulls runtime deps, runs `postinstall`), exactly as `npx @modelcontextprotocol/inspector` would — and drives the installed `mcp-inspector` bin end to end: `--help` dispatch, a real `--cli tools/list` over stdio, and a prod `--web` boot that must serve `/` from the shipped `dist`. It catches "works in `--dev`, breaks under `npx …`" path/packaging failures. It requires network access (the install pulls deps), so it is a local / release check, **not** part of the fast `validate`/`ci` loop.
+The `smoke:*` scripts run against the in-repo build tree, which is **not** the published package. `npm run pack:verify` (`scripts/pack-and-verify.mjs`) closes that gap: it builds, `npm pack`s the publishable tarball (asserting no source maps ship and that the runtime-required files are present), installs the tarball into a **clean throwaway consumer** — a fresh temp directory where it runs a real `npm install <tgz>` (pulls runtime deps, runs `postinstall`), exactly as `npx @modelcontextprotocol/inspector` would — and drives the installed `mcp-inspector` bin end to end: `--help` dispatch, a real `--cli tools/list` over stdio, a prod `--web` boot that must serve `/` from the shipped `dist`, and — riding that same boot — an **MCP App rendered in headless Chromium** through the shipped sandbox proxy, connect → open app → `data-app-status="ready"` ([#2003](https://github.com/modelcontextprotocol/inspector/issues/2003)). That last step shares its flow with `smoke:web:app` via `scripts/lib/mcp-app-flow.mjs` rather than copying it, since the deep-link shape is the part that rots silently; the *client* comes from the install while the App test server stays a repo fixture. It catches "works in `--dev`, breaks under `npx …`" path/packaging failures. It requires network access (the install pulls deps), so it is a local / release check, **not** part of the fast `validate`/`ci` loop.
 
 ### Cutting a release
 
@@ -401,7 +513,7 @@ npm version minor --no-git-tag-version   # or major / patch; bump only, no tag
 
 **2. Merge `v2/main` → `main`** through the usual milestone-merge branch. It now carries the bump, so the release lands on `main` with the version already correct.
 
-Between steps 1 and 2 the two branches **do** differ, and that is expected, not drift: `v2/main` reads the version being built while `main` still reads the one currently released. What this ordering removes is *post-release* drift — once the milestone merge lands they agree again, and `v2/main` is never left **behind** `main`. If you see `v2/main` ahead of `main`, a release is in flight; if you see it behind, something went wrong.
+Between steps 1 and 2 the two branches **do** differ, and that is expected, not drift: `v2/main` reads the version being built while `main` still reads the one currently released. What this ordering removes is _post-release_ drift — once the milestone merge lands they agree again, and `v2/main` is never left **behind** `main`. If you see `v2/main` ahead of `main`, a release is in flight; if you see it behind, something went wrong.
 
 **3. Tag the `main` commit and draft the Release:**
 
@@ -417,7 +529,7 @@ git tag 2.3.0 origin/main && git push origin 2.3.0
 
 The release's target commit selects which workflow runs, so this only publishes when a release is cut from a commit carrying this (v2) workflow.
 
-**Why the bump goes on `v2/main` first ([#2010](https://github.com/modelcontextprotocol/inspector/issues/2010)).** It used to happen on the milestone-merge branch, which is cut from `main` — so the bump existed only *downstream* of `v2/main` and nothing carried it back. `v2/main` sat at `2.0.0` through both the 2.1.0 and 2.2.0 releases. That is not cosmetic: a branch cut from a milestone-merge branch silently carries the bump into an unrelated PR (this happened on [#2009](https://github.com/modelcontextprotocol/inspector/issues/2009), where a container bugfix arrived with a `2.0.0 → 2.2.0` diff), and anything reading the version in development — `readInspectorVersion()`, `--version`, `GET /api/config` — reported a version two releases old.
+**Why the bump goes on `v2/main` first ([#2010](https://github.com/modelcontextprotocol/inspector/issues/2010)).** It used to happen on the milestone-merge branch, which is cut from `main` — so the bump existed only _downstream_ of `v2/main` and nothing carried it back. `v2/main` sat at `2.0.0` through both the 2.1.0 and 2.2.0 releases. That is not cosmetic: a branch cut from a milestone-merge branch silently carries the bump into an unrelated PR (this happened on [#2009](https://github.com/modelcontextprotocol/inspector/issues/2009), where a container bugfix arrived with a `2.0.0 → 2.2.0` diff), and anything reading the version in development — `readInspectorVersion()`, `--version`, `GET /api/config` — reported a version two releases old.
 
 Do **not** "fix" a future drift by merging `main` back into `v2/main`. `main` carries the entire pre-v2 v1 history (retained through `ec5d8e13 chore: replace main's tree with v2` — ~230 commits `v2/main` does not have), so a back-merge grafts all of it into the develop branch's log permanently in order to deliver a two-file change. Bumping first means there is nothing to back-merge.
 
@@ -441,7 +553,14 @@ docker run --rm -p 127.0.0.1:6274:6274 -p 127.0.0.1:6275:6275 \
   ghcr.io/modelcontextprotocol/inspector
 ```
 
-Publish it on the **same port number** inside and out. The sandbox URL is handed to the browser via `/api/config` as `http://localhost:<container port>/sandbox`, so remapping it (`-p 9000:6275`) advertises a port the browser can't reach; use `-e MCP_SANDBOX_PORT=9000 -p 127.0.0.1:9000:9000` instead.
+**And `6278` if your app declares `_meta.ui.domain`.** That is the spec field a server uses to ask its host for a stable, dedicated origin — without one the app runs at an opaque origin and its requests carry `Origin: null`, which no CORS / OAuth-callback / API-key allowlist can admit. The Inspector answers the request with a real loopback origin on a third listener, `MCP_APP_ORIGIN_PORT` (default `6278`); apps that declare no `domain` never touch it. **Publish it if you use one** — this is the one failure that does not fall back: the backend publishes fine (its listener bound inside the container), so it hands the browser a URL on a port the browser cannot reach, and that app's frame stays **blank**. The cross-origin navigation failure is not observable from the page, so there is no opaque-origin fallback and no console warning here; those cover the failures the *backend* can see (no listener, a port that never bound, an older backend). See [MCP App dedicated origins](./clients/web/README.md#mcp-app-dedicated-origins-metauidomain) for the host-specific contract and its isolation trade-offs.
+
+```bash
+docker run --rm -p 127.0.0.1:6274:6274 -p 127.0.0.1:6275:6275 -p 127.0.0.1:6278:6278 \
+  ghcr.io/modelcontextprotocol/inspector
+```
+
+Publish each on the **same port number** inside and out. The sandbox URL is handed to the browser via `/api/config` as `http://localhost:<container port>/sandbox`, so remapping it (`-p 9000:6275`) advertises a port the browser can't reach; use `-e MCP_SANDBOX_PORT=9000 -p 127.0.0.1:9000:9000` instead. The same holds for the app origin: the URL a published app document is served from is built from the port the container bound, so remap with `-e MCP_APP_ORIGIN_PORT=9001 -p 127.0.0.1:9001:9001` rather than with `-p` alone.
 
 **Keep the `127.0.0.1:` prefix on the published port.** A bare `-p 6274:6274` publishes on **every host interface**, putting the Inspector on your local network. The container's `HOST=0.0.0.0` is a separate concern — it governs the _container's_ interfaces, not the host's — so the `DANGEROUSLY_BIND_ALL_INTERFACES` opt-in that guards a wildcard bind outside a container does not cover this. It matters more here than for an ordinary web app: the backend spawns processes on request, `GET /` embeds the API token into the served HTML, and a request arriving with **no** `Origin` header skips the origin allow-list entirely — so for any non-browser client the API token is the only guard. Publishing wider needs a real access-control boundary in front of the Inspector — a reverse proxy that authenticates, an SSH tunnel, a private network. Setting your own `MCP_INSPECTOR_API_TOKEN` does **not** substitute: `GET /` discloses whatever token is in use, so a custom one is harvested exactly as easily as a generated one.
 
@@ -455,6 +574,45 @@ docker run --rm -p 127.0.0.1:6274:6274 \
 
 The same volume also persists OAuth tokens and stored state, so an authorized server stays authorized across runs. Use `-e MCP_CATALOG_PATH=/some/other/path.json` to put the catalog somewhere else — mount a volume covering whatever directory you point it at. If you **bind-mount a host directory** instead of a named volume (`-v "$PWD/inspector-data:/home/node/.mcp-inspector"`), the directory keeps its host ownership, so on Linux add `--user "$(id -u):$(id -g)"` or `chown` it to uid `1000` — otherwise the non-root `node` user can't write and adding a server fails with `EACCES`.
 
+**Where secrets go, and how to make them survive (#1950).** The Inspector keeps the values it deliberately does _not_ write to `mcp.json` — an OAuth client secret, an enterprise IdP client secret, each stdio `env:` value — in the **OS keychain**. A container has no keychain (the published image has no D-Bus session), so on startup the Inspector probes for one and falls back, saying so in the logs and in a permanent footer at the bottom of the Client Settings and Server Settings dialogs. Which fallback you get depends on whether the directory it would write to is going to survive:
+
+| Situation                                                      | Store                                        | Secrets survive a restart? |
+| -------------------------------------------------------------- | -------------------------------------------- | -------------------------- |
+| Keychain reachable (a normal desktop install)                  | OS keychain                                  | Yes                        |
+| Container, **no volume** on `/home/node/.mcp-inspector`        | Memory                                       | No — session only          |
+| Container **with** that volume, or any host without a keychain | `~/.mcp-inspector/secrets.json`, mode `0600` | Yes                        |
+
+So the same volume that keeps your server list also switches secrets from session-scoped to durable — nothing extra to configure. The in-memory default for an unmounted container is deliberate: a file in the writable layer is discarded by `--rm` and by every image update, and promising durability it can't deliver is worse than declining to.
+
+**A file-backed store is unencrypted unless you give it a key.** Set `MCP_INSPECTOR_SECRET_KEY` and the file is encrypted with AES-256-GCM (the passphrase is stretched with scrypt against a per-file random salt). Without it the file is still `0600`, but the values are readable to anyone who can read the file — which the startup log and the settings footer both say, every session, in a warning tone:
+
+```bash
+docker run --rm -p 127.0.0.1:6274:6274 \
+  -v mcp-inspector-data:/home/node/.mcp-inspector \
+  -e MCP_INSPECTOR_SECRET_KEY="$MY_PASSPHRASE" \
+  ghcr.io/modelcontextprotocol/inspector
+```
+
+**Use a high-entropy passphrase — generated, not chosen.** The random salt stops an attacker precomputing a table across files; it does nothing against _guessing_, and the scrypt cost is deliberately low because the derivation runs on every read and write. Anyone who obtains `secrets.json` can therefore test candidate passphrases quickly and offline, so treat this value like any other credential rather than like a memorable password.
+
+Setting the passphrase later is safe — the next write upgrades an existing plaintext file in place. Until that write happens the existing values really are still readable, and the banner and footer keep saying so rather than reporting the file as encrypted the moment the variable appears. **Changing or losing the passphrase is not safe**: a file that can no longer be decrypted is read as empty and _refuses to be written_, rather than being silently replaced with a new one holding only your latest secret. Restore the original passphrase, or delete `secrets.json` and re-enter the values.
+
+The Inspector writes the file `0600` and re-tightens it at startup if something loosened it. If it _cannot_ — the file belongs to another user, or the mount is read-only — it says so in the log rather than continuing to describe the file as protected, since on that box the mode claim above is not true.
+
+**Two Inspectors, one file.** Within a process, mutations are serialized per file path, so a web session's own concurrent saves cannot lose each other. Across processes — a CLI run beside a web session — each mutation takes an exclusive lock on `secrets.json.lock` for the whole read-modify-write, using [`proper-lockfile`](https://github.com/moxystudio/node-proper-lockfile) (the same library npm itself locks with). The lock expires 10 seconds after its holder stops refreshing it, so an Inspector that is killed mid-save does not leave the file unwritable.
+
+Two running Inspectors are therefore genuinely serialized. What a lock file cannot make single-winner is the *takeover of a lock whose holder died* — that needs a compare-and-swap on a directory entry (`renameat2`) which Node does not expose, and it is what an earlier hand-rolled attempt failed three review rounds on. `proper-lockfile` does not close that race either. The window opens only after a holder dies without releasing.
+
+The Inspector adds one thing on top: every lock-directory removal the library makes on its behalf — on release, and from its exit handler — is guarded by a check that the directory is still the one it created (by inode and birth time, which survive the library's own refresh but not a delete-and-recreate). That matters because those removals are otherwise unconditional, so a holder whose lock had been replaced would delete the *winner's* lock on the way out, turning one compromised writer into two unprotected ones. It also surfaces the takeover as a warning. Treat all of this as **best-effort**: the guard is still a check followed by an act, so it makes the destructive case rare rather than impossible, and it rests on filesystem metadata that not every filesystem reports.
+
+Which is why, underneath the lock, each mutation still reads the file, applies its change, writes, then reads back and compares the whole map; if something wrote in between it re-applies onto what was left and retries, failing loudly after five lost rounds rather than returning as though the value were saved. That check is what still catches a clobber inside that window — and it covers what no lock can, since a lock only orders the writers that *take* it: an editor, a restored backup, or an Inspector older than this release.
+
+If another process holds the lock and will not let go, the save **fails** rather than going ahead unlocked — waiting past the stale window first, so a crashed Inspector resolves itself rather than failing everyone else's saves. Writing alongside a writer you can see is the one case where degrading would lose the secret it was trying to protect.
+
+It is also what covers the lock being unavailable. This store exists for boxes where the usual mechanism isn't there, so a directory that can't hold a lock file — a read-only `$HOME`, a mount owned by another uid — makes the save proceed unlocked with a warning, rather than turning every `set` into a failure on exactly the deployments the store was written for.
+
+Three env vars affect where the file lands. `MCP_INSPECTOR_SECRET_STORE=keyring|file|memory` picks the store outright, bypassing the probe. `MCP_INSPECTOR_SECRET_FILE` names the file. Failing both, the file follows `MCP_STORAGE_DIR` — the same variable that relocates OAuth tokens and `client.json` — so mounting a volume at your configured storage directory is enough to make secrets durable there.
+
 **Upgrading from an image before this fix?** Earlier images did not create `/home/node/.mcp-inspector`, so Docker created the volume's mount point as `root` and the non-root `node` user couldn't write to it. An **empty** volume repairs itself on the first run of a current image (Docker applies the image directory's ownership to an empty volume), but one that already has files in it keeps its old `root` ownership and still fails with `EACCES`. Fix it once:
 
 ```bash