| Filename | Latest commit message | Latest commit date |
|---|---|---|
519ecf49ee
A floated promise is invisible at review time — the call reads like an awaited one minus four characters — and the unhandled rejection it produces surfaces in a different test, in a different file, as a stack pointing at SDK internals. In #1947 two un-held `callTool` calls made `npm run ci` unpassable and took a full investigation to attribute; the fix was two lines. `@typescript-eslint/no-floating-promises` catches the class outright, so enable it at `error` everywhere. The rule is type-aware and `tseslint.configs.recommended` is not, so each scope's config gains a parser project. The four clients name their own tsconfig projects (both of them — `src` and `__tests__`/`server` live in different ones). The root scope has no tsconfig at all, so add `tsconfig.lint.json`: a lint-only project covering `core/**`, `test-servers/src/**`, and `vitest.shared.mts`, emitting nothing and gating nothing. It sets `moduleResolution: bundler` deliberately — under NodeNext `core/`'s extensionless imports fail to resolve, every import degrades to `any`, and the rule silently stops seeing promises at all. Measured surface, and the fixes: - core (2) — `stream.close()` in two Hono `onAbort` listeners. Hono invokes subscribers with a bare `subscriber()`, so a promise returned from the listener is floated one layer up; `closeAbortedStream` owns the rejection where it can actually be owned. - tui (5) — `handleDisconnect` had no catch of its own and its only caller is a key handler that cannot await, so a failing disconnect was a live unhandled rejection; it now surfaces the message the way `handleConnect` does. The rest are `void` at callers that genuinely cannot await, over callees that already own their failures. - web (92) — 83 are Storybook `expect(...)`, which is instrumented and returns a promise, now awaited (`scrollAreaStoryAssertions` is async for the same reason). Two tests were asserting against a provider call they never awaited. `open(url)` in both web servers now warns instead of escaping a synchronous callback. - cli, launcher — already clean. Type-aware linting costs time: `clients/web`'s `eslint .` goes from ~8s to ~19s, inside `validate`. Narrowing the loaded projects is the lever if that needs reducing later; dropping the rule is not. Closes #1959 Signed-off-by: cliffhall <cliff@futurescale.com> |
||
| .. | ||
| __tests__ | ||
| scripts | ||
| src | ||
| eslint.config.js | ||
| package-lock.json | ||
| package.json | ||
| README.md | ||
| tsconfig.json | ||
| tsconfig.test.json | ||
| vitest.config.ts | ||
MCP Inspector Launcher
The launcher is the package that provides the global mcp-inspector binary (e.g. when users run npx @modelcontextprotocol/inspector). It is not a separate user-facing app—it is the single entrypoint that selects and runs one of the clients (web, CLI, or TUI).
Responsibility
- Parse mode from a leading prefix of
--web(default),--cli, or--tuiimmediately after the script name. - Forward all following arguments unchanged (including tokens that look like mode flags).
- Dynamically import that app’s runner from
clients/{web,cli,tui}/build/index.js(relative to the launcher build output) and call it in-process (nospawn()).
All configuration parsing, config-file loading, and server setup are handled by the app runners and by core; the launcher does not interpret config or env vars.
Error reporting. A --cli failure is routed through the CLI's own error sink, so mcp-inspector --cli preserves the CLI exit-code map (1 usage, 2 no-app, 3 auth-required, 4 unreachable, 5 tool-error) and its machine-readable {"error":{…}} stderr envelope — the same as invoking the CLI bin directly. --web / --tui failures print a human-readable Error: <message> and exit 1 (append MCP_DEBUG=1 for the stack).
Web server-list flags (--web)
mcp-inspector --web chooses which server list the UI shows and whether it is
editable (see specification/v2_catalog_launch_config.md):
| Invocation | Server list | Editable in UI? |
|---|---|---|
mcp-inspector --web |
Default catalog ~/.mcp-inspector/mcp.json (seeded with the two sample servers if missing) |
Yes |
mcp-inspector --web --catalog <path> (or MCP_CATALOG_PATH=<path>) |
That file as the active catalog (same seed-if-missing behavior) | Yes |
mcp-inspector --web --config <path> |
That file as a read-only session — shown but never written, seeded, or migrated (safe for a foreign config) | No |
mcp-inspector --web --server-url <url> --transport http --header "Name: Value" (or a positional command) |
One ad-hoc server held in memory, connectable with the given --headers |
No |
Rules: --catalog and --config are mutually exclusive; neither combines with
an ad-hoc target or --header; --header requires an ad-hoc HTTP/SSE server
and is applied to that connection (it is no longer a warn-only no-op).
Seed contents are web-specific. When the web backend creates a missing
writable catalog it seeds DEFAULT_SEED_CONFIG (core/mcp/serverList.ts) — a
filesystem-server-default scoped to /tmp plus the canonical
everything-server-default — so a first launch has something to connect to.
The CLI and TUI seed an empty catalog instead; see the next section. A
read-only --config is never seeded on any surface.
CLI and TUI server-list flags (--cli / --tui)
The CLI and TUI use the same --catalog / --config vocabulary as --web,
resolved by the shared core/mcp/node/config.ts helpers:
| Invocation | Server list source |
|---|---|
mcp-inspector --cli / --tui (no source flag, no ad-hoc target) |
Default writable catalog ~/.mcp-inspector/mcp.json (created/seeded empty if missing) |
--catalog <path> (or MCP_CATALOG_PATH=<path>) |
That file as a writable catalog — seeded empty if missing |
--config <path> |
That file as a read-only session — served as-is, never written or seeded; errors if absent |
positional command / --server-url <url> |
One ad-hoc server |
Note the seed contrast with --web above: the CLI and TUI write an empty
{ "mcpServers": {} } (seedEmptyCatalog in core/mcp/node/config.ts), not
the web client's two sample servers — they are non-interactive or list-driven,
so sample entries would be noise rather than a starting point.
Rules (shared serverSourceConflict): --catalog and --config are mutually
exclusive, and neither combines with an ad-hoc command/URL target. The CLI/TUI
do not perform catalog CRUD yet — they are read consumers — so the
writable/read-only split currently surfaces only as seed-if-missing
(--catalog/default) vs error-if-missing (--config). Full writable
persistence is tracked in #1482 / #1432.
Production web build (--web, no --dev)
Prod --web serves static assets from clients/web/dist/, which only exists
after a build. In the published package dist/ always ships. In a fresh dev
checkout it is absent, so the runner builds it on demand the first time you
launch (vite build via npm run build:client, run in clients/web) instead
of serving a broken page (#1486). If that build can't run — e.g. dev
dependencies are missing — the launcher exits with an actionable error telling
you to run npm run build (from clients/web) or relaunch with --dev to use
the Vite dev server. --dev never needs dist/; it runs Vite directly.
CI and npm run validate (via the top-level smoke step) exercise this prod
path end-to-end with npm run smoke:web (scripts/smoke-web.mjs): it starts
mcp-inspector --web against the built dist/ and asserts GET / returns the
SPA (HTTP 200) with the injected __INSPECTOR_API_TOKEN__.
CLI and TUI smokes
The top-level smoke step also runs end-to-end smokes for the other two modes
through the built launcher artifact (beyond the --help checks in
smoke:launcher):
npm run smoke:cli(scripts/smoke-cli.mjs) — runsmcp-inspector --cliagainst the bundled stdio test server via a temp--catalogand assertstools/listreturns the server's tools, plus the--catalog/--configresolution paths (default-catalog seed-on-missing, read-only--configerror-without-seed,--catalog/--configconflict).npm run smoke:tui(scripts/smoke-tui.mjs) — launchesmcp-inspector --tui --catalog <temp>and asserts the Ink app renders its first frame within a timeout, then shuts it down (a shallow boot/render check, not full interaction).
Both build test-servers/build on demand if it is missing.
Development
Like the web client, the launcher self-validates from its own folder:
npm run validate # format:check && lint && build && test:coverage
This has no dependency on the other clients being built — it only checks the
launcher's own source. eslint.config.js is a Node-only flat config (the web
client's React/Storybook plugins stripped out), and the per-file coverage gate
covers parse-launcher-argv.ts (the pure arg-parsing logic); src/index.ts is
excluded as binary bootstrap and is instead exercised by the smokes above. The
repo-root validate:launcher simply delegates here (cd clients/launcher && npm run validate).
Publishing
The launcher provides the mcp-inspector bin for the single @modelcontextprotocol/inspector tarball. Packaging is a whole-repo concern — how the one-package/single-version tarball is assembled, the "files" allowlist invariants (no source maps, why clients/web/build needs .npmignore, why the cli/tui package.jsons ship), and the npm run pack:verify publish smoke — is documented in the root README.
Architecture
For how the launcher fits with the shared config processor and app runners, see the Launcher and config consolidation document in the repo root docs/ folder.