Skip to content

README.md

Latest commit

 

History

History
700 lines (526 loc) · 63.2 KB

README.md

File metadata and controls

700 lines (526 loc) · 63.2 KB
Raw

MCP Inspector CLI Client

CLI for the Inspector: connect, run a --method, disconnect. Invoked as mcp-inspector --cli.

Running the CLI

You can run the CLI via npx:

npx @modelcontextprotocol/inspector --cli node build/index.js

Supports tools, resources, and prompts (plus --method servers/list / servers/show to read catalog entries, and servers/add / servers/edit / servers/remove to change them, without connecting).

Coming from the v1 CLI? See the v1 → v2 migration guide — every v1 flag still exists, but exit codes, argument ordering, and the -- separator changed.

Examples

Basic usage

npx @modelcontextprotocol/inspector --cli node build/index.js

With a configuration file

npx @modelcontextprotocol/inspector --cli --config path/to/config.json --server myserver

List available tools

npx @modelcontextprotocol/inspector --cli node build/index.js --method tools/list

Call a specific tool

npx @modelcontextprotocol/inspector --cli node build/index.js --method tools/call --tool-name mytool --tool-arg key=value --tool-arg another=value2

Call a tool with JSON arguments

npx @modelcontextprotocol/inspector --cli node build/index.js --method tools/call --tool-name mytool --tool-arg 'options={"format": "json", "max_tokens": 100}'

List available resources

npx @modelcontextprotocol/inspector --cli node build/index.js --method resources/list

List available prompts

npx @modelcontextprotocol/inspector --cli node build/index.js --method prompts/list

Remote Servers

You can also connect to remote MCP servers using HTTP or SSE transports.

Connect to a remote MCP server (transport inferred from a /mcp or /sse path)

npx @modelcontextprotocol/inspector --cli https://my-mcp-server.example.com/mcp --method tools/list

There is no default transport: a URL whose path ends in neither /mcp nor /sse needs an explicit --transport http or --transport sse, as below. A target is either a positional command/URL or --server-url, never both, and --transport stdio cannot be used with a URL — either combination is refused rather than resolved by precedence.

Connect with Streamable HTTP transport

npx @modelcontextprotocol/inspector --cli https://my-mcp-server.example.com --transport http --method tools/list

Pass custom headers

npx @modelcontextprotocol/inspector --cli https://my-mcp-server.example.com --transport http --method tools/list --header "X-API-Key: your-api-key"

When a server is loaded from a --catalog/--config file, its per-server settings (headers, connection/request timeouts, OAuth, and roots) are applied to the connection — the same resolution the TUI uses. A --header flag overrides the file's headers for that run while leaving the file's timeouts and OAuth in place.

The file is the only durable way to give a run its roots: there is no roots flag, and --method roots/set applies only to that one short-lived connection. Roots configured for a server (the same field the web UI's Server Settings writes) are advertised at connect, so a server that asks for roots/list — @modelcontextprotocol/server-filesystem does, to learn its allowed directories — gets them.

Environment-variable semantics. MCP_CATALOG_PATH is honored only when no ad-hoc target is given (positional command, --server-url, or --transport) — so a shell that exports it can still run one-off ad-hoc invocations without hitting the catalog/ad-hoc conflict. MCP_STORAGE_DIR sets the storage directory used by the OAuth persist backend (<MCP_STORAGE_DIR>/oauth.json); the per-file MCP_INSPECTOR_OAUTH_STATE_PATH override still takes precedence over it. Every variable the CLI reads is listed in Environment variables.

HTTP proxy support

Connections to remote HTTP/SSE servers honor the conventional proxy environment variables: HTTPS_PROXY / HTTP_PROXY (and their lowercase forms) select the proxy, and NO_PROXY exempts hosts. This applies to every Node client — CLI, TUI, and the web backend — with no inspector-specific flag. It also covers OAuth discovery and token requests, which run through the same fetch.

Proxy routing is powered by the undici package (declared in the root manifest only; ^8.x requires Node >= 22.19.0, the Inspector's supported floor). It is imported lazily on the first request and only when a proxy variable is set, so runs without a proxy configured pay no cost.

Both halves of the pair come from userland undici — its fetch and its EnvHttpProxyAgent (#2067). The obvious-looking alternative is to hand the agent to Node's built-in fetch as a dispatcher, which is what v2.0.0–2.3.0 did. That couples two different copies of undici at the dispatcher handler interface, and that interface is not stable across majors: Node 22 embeds undici 6, Node 24 embeds 7, Node 26 embeds 8. A userland undici 8 agent handed a built-in undici 7 handler is rejected outright — fetch failed: invalid onRequestStart method — and the request never leaves the process. Keeping both sides inside one copy is what makes proxying work unchanged from the Node 22.19 floor through Node 26. (Node's own NODE_USE_ENV_PROXY is not a substitute: it is unsupported at our 22.19 engine floor.)

Because undici's Response is a different class from globalThis.Response, the proxied fetch re-wraps each response as a genuine global Response, streaming preserved. Without that, res instanceof Response is false for callers that test it — including the MCP SDK, whose OAuth error formatter would degrade every message to Raw body: [object Response].

undici is declared only in the root package.json, and every client's tsup config lists it as external. Both halves matter: tsup auto-externalizes what the nearest manifest declares, so without the explicit entry the web and TUI bundles inlined it — and a CommonJS package inlined into an ESM bundle throws Dynamic require of "assert" is not supported the first time it is used. npm run verify:bundle-externals is the durable guard.

Shell completion

--completion <bash|zsh|fish> prints a completion script for mcp-inspector on stdout and exits without connecting to anything. The first word completes the launcher's mode flags (--cli, --web, --tui); after --cli it completes every CLI flag, the --method names (tools/list, tools/call, servers/list, …) and the finite values of --transport, --log-level, --format and --protocol-era. Path flags (--catalog, --config, --cwd, --client-config) and the stdio target command fall back to file completion.

# bash — current shell, or persist it
source <(mcp-inspector --cli --completion bash)
mcp-inspector --cli --completion bash > ~/.local/share/bash-completion/completions/mcp-inspector

# zsh — current shell (after compinit), or save it on your $fpath
source <(mcp-inspector --cli --completion zsh)
mcp-inspector --cli --completion zsh > "${fpath[1]}/_mcp-inspector"

# fish
mcp-inspector --cli --completion fish > ~/.config/fish/completions/mcp-inspector.fish

The flag list is read from the CLI's own commander definition when the script is generated, so it cannot drift from --help; regenerate the script after upgrading to pick up new flags. Only --cli mode is completed — web and TUI flags are not.

Options

MCP server (which server to connect to)

Options that specify the MCP server (catalog/config file, ad-hoc command/URL, env vars, headers) are shared by the Web, CLI, and TUI and are documented in MCP server configuration: --catalog (writable catalog, seeded empty if missing; default ~/.mcp-inspector/mcp.json or MCP_CATALOG_PATH), --config (read-only session, errors if absent), --server, -e, --cwd, --header, --protocol-era (legacy/auto/modern; sets the era an ad-hoc run negotiates, or overrides a file's protocolEra), --transport, --server-url, and the positional [target...]. --catalog and --config are mutually exclusive, and neither combines with an ad-hoc target.

CLI-specific (what to invoke)

Option Description
--method <method> MCP method to invoke. Supports initialize (connect-only probe → {serverInfo, protocolVersion, capabilities, instructions}), tools/list, tools/call, resources/list, resources/read, resources/templates/list, prompts/list, prompts/get, logging/setLevel, skills/list, skills/get, resources/directory/read, plus catalog-only servers/list / servers/show and servers/add / servers/edit / servers/remove (no MCP connect; see Editing the server catalog). Stream / session-only methods (e.g. logging/tail) are rejected.
--rename <name> New name for the entry, with --method servers/edit only. Secrets stored under the old name move with it.
--tool-name <name> Tool name (for tools/call).
--tool-arg <key=value> Tool argument; repeat for multiple. Use key='{"json":true}' for JSON. Values are coerced (JSON-parsed, so count=1 becomes a number).
--tool-args-json <json> Tool arguments as a single JSON object (e.g. '{"zip":"10001"}'). Passed verbatim — no key=value coercion, so "012" stays a string. Mutually exclusive with --tool-arg.
--uri <uri> Resource URI (resources/read), directory URI (resources/directory/read), or skill URI (skills/get).
--cursor <cursor> Opaque pagination cursor for resources/directory/read — pass back the nextCursor from the previous page. The listing is not recursive and pages are not aggregated: SEP-2640 gives the cursor to the client, and descending is the caller's job.
--prompt-name <name> Prompt name (for prompts/get).
--prompt-args <key=value> Prompt arguments; repeat for multiple.
--log-level <level> Logging level for logging/setLevel (e.g. debug, info).
--metadata <key=value> General _meta entries (key=value); applied to all methods. The value is JSON-parsed when it parses, so trace={"id":"abc"} sends a real object and n=3 a number; anything that is not valid JSON is sent as the literal string. Merged over the server's persisted metadata from mcp.json, which is sent on every request without this flag (#2093).
--tool-metadata <key=value> Tool-specific _meta entries for tools/call. Same JSON-parsed value handling as --metadata.
--connect-timeout <ms> Connection timeout in ms. Defaults to 15000 for ad-hoc --server-url/target runs (so a black-holed host fails fast) and to the file-level connectionTimeout for --catalog/--config runs — 30000 when the file sets none. 0 disables the timeout.
--app-info Probe a tool's MCP App UI metadata without invoking it. With --method tools/call --tool-name <name>: prints one JSON line (hasApp, resourceUri, csp, permissions, domain, …) and exits 0 if the tool has an app or 2 (no_app) if not. With --method tools/list: emits NDJSON — one app-info line per tool over a single connection.
--advertise-apps Advertise the MCP Apps UI extension (io.modelcontextprotocol/ui) at initialize. Off by default, because the CLI cannot render an App and a server decides whether to return one from that advertisement. Set it when a server only exposes its App tools to a client that claims App support — typically alongside --app-info.
--strict With --method tools/list: report tool-schema portability problems in full (path, issue, suggested fix) on stderr, and exit 6 if any is error-severity. Without it, a one-line count is printed instead. See Schema portability.
--verify With --method skills/list or --method skills/get: run the SEP-2640 conformance, digest and frontmatter checks over the skills returned, emit one JSON report per skill on stdout, and exit 7 if any fails. See Skill verification.
--require-digests With --verify: exit 9 when a skill advertises no digests (resources: "dynamic"), instead of reporting it unverifiable and exiting 0. See Skill verification.
--skill-catalog-max-skills <n> With --verify (required; rejected without it, since it would bound nothing): the skills catalog budget, the most skills whose files one run reads (default 256). Positive integer; anything else is rejected before connecting. Overrides the server's skillCatalogMaxSkills from the catalog/config file — the same setting the web client exposes per server in Server Settings. Skills past the budget are reported incomplete (exit 8).
--skill-catalog-max-bytes <n> With --verify (required): the skills catalog budget, the most bytes one run reads across all skills (default 67108864, 64 MiB). Positive integer. Overrides the server's skillCatalogMaxBytes.
--format <text|json> Output format. text (default) pretty-prints the result. json emits a single JSON object on stdout ({ "result": … }, plus { "appInfo": … } as a sibling key for App tools) with no banners, so the whole output pipes cleanly into jq.
-q, --quiet Print only the result payload on stdout, or the error envelope on stderr on failure. Drops status lines, warnings, advisory summaries and a stdio server's own stderr; OAuth prompts a human must answer still print. See Quiet output.
--output <path> Write the result to a file instead of stdout. The parent directory must exist; an existing file is replaced. See Saving a result to a file.
--output-format <raw|json> Encoding of the --output file. json (default) is the whole result, pretty-printed. raw is the result's text, or the decoded bytes of its single image/audio/blob block; tools/call and resources/read only.
--load Load-test one tool: call it repeatedly and concurrently, then print a latency / error-rate summary in place of a result. Requires --method tools/call --tool-name <name>. See Load testing a tool.
--load-concurrency <n> With --load: the most calls in flight at once (positive integer; default 10).
--load-duration <sec> With --load: stop starting calls after this many seconds; calls already in flight still finish. Default 10 when --load-requests is not set.
--load-requests <n> With --load: stop after starting this many calls. With --load-duration too, whichever is reached first ends the run.
--load-rate <n> With --load: start at most this many calls per second across all workers, spaced evenly (default: unlimited).
--relogin Delete stored OAuth for this server URL from the shared store before connect; interactive login still only runs if the server requires auth. Requires an HTTP/SSE URL (rejected for stdio). Conflicts with --stored-auth-only / --use-stored-auth / --wait-for-auth / catalog short-circuits.
--no-revoke With --relogin, skip the RFC 7009 revocation request that would otherwise end the grant at the authorization server when the local state is deleted. The per-server oauth.revokeOnClear setting is the persistent form of the same opt-out; either one is enough to skip it. See Revoking on --relogin.
--stored-auth-only CI / non-interactive safe: never start interactive OAuth / step-up (and never auto-open a browser); use the shared store if present, otherwise fail immediately with auth_required. Prefer this over a bare pipe/CI run that would otherwise attempt interactive login.
--completion <shell> Print a bash, zsh or fish completion script and exit (no server connection). See Shell completion.

Revoking on --relogin

--relogin deletes this server's stored OAuth state so the next connect cannot silently reuse it. By default it now also revokes the grant at the authorization server, per RFC 7009 (#2144) — otherwise the delete is invisible to the AS, and the access token, plus the refresh token when one was issued, stay valid there until they expire on their own.

The request is built from the stored state before the delete and sent after it, so the local delete never waits on the network. That matters because the OAuth store is shared: holding this process's view of it across a five-second request would let another CLI or TUI write a fresh grant that this one then erased.

The request names the refresh token when there is one: RFC 7009 §2.1 asks the authorization server to invalidate the access tokens issued under the same grant, so a single request covers both.

It is best-effort and never changes the exit code. An authorization server that advertises no revocation_endpoint gets no request at all; a network error, a non-2xx, or the short timeout prints a one-line warning on stderr and the local delete proceeds either way.

Turn it off per run with --no-revoke, or per server with oauth.revokeOnClear: false in the catalog — either is enough, and neither can turn it on for the other. Disconnecting while still holding live tokens is a case worth reproducing when the server under test is the thing being debugged.

mcp-inspector --cli --server-url https://example.com/mcp --relogin --method tools/list
mcp-inspector --cli --server-url https://example.com/mcp --relogin --no-revoke --method tools/list

servers/show redacts secret-bearing fields (env values, sensitive headers, sensitive settings.metadata keys whose whole value is replaced whether or not it is structured, requestInit / eventSourceInit headers, oauthClientSecret). It does not scrub credentials embedded in a server url (userinfo or query tokens) or in stdio args — treat detail / raw URL fields as potentially sensitive before pasting into issues.

Editing the server catalog

servers/add, servers/edit and servers/remove write the same catalog file the web and TUI clients use — --catalog <path>, else MCP_CATALOG_PATH, else ~/.mcp-inspector/mcp.json — through the web backend's own /api/servers write path, run in-process. So the file comes out exactly as the web UI would write it: the atomic write, name validation, and the split of secret values (stdio env values) out of mcp.json into the secret store (OS keychain, or the fallback store). A web backend that is already running picks the change up through its file watcher. The server entry is named with --server <name> and described with the same flags an ad-hoc run takes: a positional command or URL, --server-url, --transport, -e, --cwd, --header, --protocol-era.

# Add a stdio server (its -e values go to the secret store, not mcp.json) and an HTTP one.
mcp-inspector --cli node build/index.js --method servers/add --server my-server -e API_KEY=…
mcp-inspector --cli --method servers/add --server remote --server-url https://example.com/mcp --header "X-Team: a"

# Edit: -e merges into the env and --cwd replaces it; a new command/URL replaces the
# transport; --header replaces the headers; --protocol-era sets the era; --rename renames.
# Anything not named (OAuth, timeouts, metadata, roots) is kept.
mcp-inspector --cli --method servers/edit --server my-server -e OTHER=1 --rename my-server-2

# Remove the entry and its stored secrets.
mcp-inspector --cli --method servers/remove --server remote

Each prints { ok, action, server, catalog } (plus previousName on a rename); --format json wraps it in { "result": … }. --config is refused, since it names a read-only session file. servers/add fails on a name that already exists, and servers/edit / servers/remove fail on one that does not. servers/remove rejects the entry-describing flags (a command/URL, --transport, -e, --cwd, --header, --protocol-era, --rename), and servers/edit with nothing to change is an error. When the secret store is in-memory only (a container with no keychain and nothing durable mounted), a write that supplies -e values is refused rather than losing them when the CLI exits, and so is --rename, since this process cannot see secrets held in another process's memory — set MCP_INSPECTOR_SECRET_STORE=file to keep them.

Quiet output (--quiet)

-q / --quiet reduces a run to its result: the payload on stdout on success, and the single-line error envelope on stderr on failure. It composes with --format: --format shapes stdout, and --quiet strips stderr of everything non-essential (the exceptions are in the table below).

mcp-inspector --cli node build/index.js -q --method tools/list | jq '.tools[].name'

What it suppresses:

Output Under --quiet
A stdio server's own stderr (startup banners, logs) Dropped
Schema portability: … Re-run with --strict for details. (tools/list) Dropped
The --verify one-line summary Dropped — a failing run's envelope carries the same text
Authorization complete. / Authorization complete. Retrying… Dropped
Warning: could not revoke the OAuth grant … (--relogin) Dropped
Advisory warnings from shared Inspector code: the [mcp-inspector] … secret-store notice (keychain fallback, memory caveat), ignored roots / OAuth-endpoint settings, lock and persistence trouble Dropped (all of console.warn is muted for the run). See secret storage for what the store notice would have said.
The result payload / NDJSON on stdout Kept
The error envelope on a non-zero exit Kept
The --strict report Kept — you asked for it, and it is the detail behind exit 6
The OAuth authorization URL, the step-up [y/N] prompt, "open it by hand" Kept — interactive login cannot finish without them

So a quiet run that needs an interactive login still shows what it must; for a run that must never prompt, combine --quiet with --stored-auth-only.

⚠️ Dropping the server's stderr and the advisory warnings also drops their explanation when something goes wrong. The CLI still exits non-zero with an envelope, but if the reason is not obvious, re-run without --quiet.

App probing (--app-info) and machine-readable output (--format json)

--app-info inspects a tool's MCP App UI posture without calling the tool, so a pipeline can decide whether to open a browser before touching one:

# Probe one tool. Exits 0 (has app) or 2 (no_app). One JSON line on stdout.
mcp-inspector --cli <server> --method tools/call --tool-name my_tool --app-info
# → {"hasApp":true,"toolName":"my_tool","resourceUri":"ui://…","csp":{…},"permissions":{…},"prefersBorder":true,"resourceMimeType":"text/html"}

# Probe every tool at once — NDJSON, one line per tool, single connection.
mcp-inspector --cli <server> --method tools/list --app-info | jq -c 'select(.hasApp)'

The CLI does not advertise the MCP Apps UI extension by default, since it cannot render an App. A server that registers its App tools only for a client that advertises Apps support will therefore show no app to a bare probe; add --advertise-apps to claim that support for the probe:

mcp-inspector --cli <server> --method tools/list --app-info --advertise-apps

Exit semantics: a tool that has an app exits 0; one with no app exits 2 (no_app); a missing tool exits 5 (tool_not_found) — distinct so a typo isn't mistaken for "no app". A probe failure (an unreadable UI resource, or a malformed _meta.ui.resourceUri) is tolerated and reported in a resourceError field rather than aborting — so in tools/list --app-info one bad tool never kills the rest of the listing.

--format json wraps any method's output in a single stdout envelope with no banners, so App tools and plain tools both pipe cleanly into jq:

mcp-inspector --cli <server> --method tools/call --tool-name my_app_tool --format json
# → {"result":{…tool result…},"appInfo":{"hasApp":true,"resourceUri":"ui://…",…}}

tools/list --app-info always emits NDJSON (one raw app-info object per line) regardless of --format — the per-tool list shape is fixed. --format json only reshapes the single-result paths (tools/call, tools/list without --app-info, etc.) into the {result[, appInfo]} envelope.

A tools/call that returns isError:true still prints its payload but exits 5 (tool_is_error) so && chains don't proceed on a failed call.

Saving a result to a file (--output)

--output <path> writes the method's result to a file instead of stdout, so a script can archive or diff results without shell redirection (#2431). --output-format picks how the file is encoded:

--output-format File contents
json (default) The whole result, pretty-printed with two-space indentation — the same shape the web client's exports download.
raw The result's own payload: the text of every text-bearing block (text blocks and embedded text resources), joined by newlines. A result with no text and exactly one binary block (image, audio, or resource blob) is written as that block's decoded bytes, so an image tool saves as a viewable file. Anything else has no single raw form and fails with output_not_raw. Only tools/call and resources/read have a raw form. 
# Archive the whole result as JSON.
mcp-inspector --cli <server> --method tools/call --tool-name echo --tool-arg message=hi --output echo.json

# Save just the text a tool returned.
mcp-inspector --cli <server> --method tools/call --tool-name summarize --output summary.txt --output-format raw

# Save an image tool's output as the image itself.
mcp-inspector --cli <server> --method tools/call --tool-name render_chart --output chart.png --output-format raw

The flag is --output-format rather than a new --format value because --format already shapes stdout, and the two compose: in text mode stdout stays empty and a one-line Wrote N bytes (json) to <path> confirmation goes to stderr (suppressed by -q, like any other status line); under --format json stdout still carries exactly one envelope, with output in place of result — {"output":{"path":"echo.json","format":"json","bytes":123}}, plus appInfo / schemaFindings when they would otherwise appear.

Exit codes are unchanged: a tools/call that returns isError:true is still written and still exits 5. A file that cannot be written (missing directory, no permission) exits 1 with envelope code output_write_failed.

--output is rejected where there is no single result to write — --app-info, --load, --verify, --list-stored-auth, --print-handoff, and servers/list / servers/show — and --output-format is rejected without --output, rather than either being silently ignored.

Load testing a tool (--load)

--load points the Inspector's own connection — transport, headers, OAuth and all — at one tool and keeps calling it, so seeing how a server behaves under concurrent or sustained traffic does not mean rewriting that setup in a separate script (#2439).

# 20 calls in flight for 30 seconds
mcp-inspector --cli <server> --method tools/call --tool-name search \
  --tool-arg query=mcp --load --load-concurrency 20 --load-duration 30

# exactly 500 calls, started at no more than 50 per second
mcp-inspector --cli <server> --method tools/call --tool-name search \
  --tool-arg query=mcp --load --load-requests 500 --load-rate 50

Every call sends the same arguments and _meta. When the run ends, one summary is printed on stdout (wrapped in { "result": … } with --format json):

{
  "tool": "search",
  "concurrency": 20,
  "durationMs": 30000,
  "requests": 41210,
  "ok": 41187,
  "toolErrors": 0,
  "errors": 23,
  "errorRate": 0.0006,
  "elapsedMs": 30012.4,
  "throughput": 1373.12,
  "latencyMs": { "min": 2.1, "mean": 14.5, "p50": 12.9, "p90": 22.4, "p99": 61.3, "max": 30004.8 },
  "errorMessages": [{ "message": "Request timed out", "count": 23 }]
}
  • toolErrors counts results carrying isError: true — the tool ran and reported a failure. errors counts calls that never produced a result: protocol errors, request timeouts, a dropped connection. errorRate is both together over requests, rounded to four places.
  • errorMessages groups the thrown messages by frequency. The first ten distinct messages are kept; any further ones are counted under (other).
  • Latency is per call, client-side, in milliseconds, with nearest-rank percentiles. It includes everything the Inspector does around the request, which is negligible next to a network round trip but not next to a sub-millisecond stdio echo.
  • Bounds. --load-duration stops starting calls; anything in flight is allowed to finish, so elapsedMs can exceed the window by up to one call. With --load-rate, starts follow an even 1/rate grid; a timer that wakes a little late keeps the grid, so the average rate holds. A slot missed entirely — every worker busy, or a stalled event loop — is dropped rather than made up later, so a stall is never followed by a catch-up burst. A slot that falls past the window is never started. Node's timers resolve to about a millisecond, so a rate in the high hundreds per second or above is reached only approximately (1000/s measured ~700 against a local stdio server); compare requests against what you asked for.

All calls share one connection, so this measures how a server copes with concurrent requests on a single session — not with many sessions. A slow call is bounded by the server's request timeout (requestTimeout in the catalog/config file).

A load run does not prompt for OAuth step-up: a tool that needs scopes the session lacks fails every call, and the challenge shows in errorMessages. Run one plain tools/call of the tool first to grant them, then load it.

The tool is looked up once before the clock starts, so a missing tool fails the run with exit 5 (tool_not_found) instead of reporting a 100% error rate. Otherwise a completed run exits 0 whatever its error rate — the summary is the result. --load is rejected with any other method and with --app-info, --output, --output-format, --list-stored-auth and --print-handoff, and each --load-* flag is rejected without --load.

Schema portability (--strict)

A tool schema can be perfectly legal JSON Schema and still be refused by the client the server is meant to run against. --strict (with --method tools/list) names those constructs — path, what is wrong, and a concrete fix — on stderr, and exits 6 if any is error-severity:

mcp-inspector --cli <server> --method tools/list --strict

The complete report against the unportable-schemas-http.json showcase — one block per finding, then the summary. On a non-zero exit the shared error handler adds one more stderr line after this, the ErrorEnvelope ({"error":{"code":"schema_unportable",…}}):

Error: tool "get_temp"
  Path: outputSchema.properties.data
  Issue: Bare `true` used where a schema object is expected.
  Suggestion: Declare what the value actually is — e.g. `{"type": "object", "additionalProperties": true}` for a free-form object. That is a deliberate change of contract, not an equivalent rewrite: `true` accepts any JSON value at all.

Warning: tool "echo"
  Path: inputSchema.properties.show_ids
  Issue: `type` is an array (["null","boolean"]). The array form is legal JSON Schema, but several MCP clients read `type` as a single string and either reject the tool or drop the constraint.
  Suggestion: Split it into `anyOf` branches, each with a single `type` — `{"anyOf": [{"type": "null"}, {"type": "boolean"}]}`. (Making the property optional instead is a different contract: absent is not the same as `null`.)

Warning: tool "echo"
  Path: inputSchema.properties.opts
  Issue: Schema carries no validation keyword at all, so it accepts any value — the object-literal spelling of a bare `true`.
  Suggestion: Declare what the value actually is — e.g. `{"type": "object", "additionalProperties": true}` for a free-form object. That is a deliberate change of contract, which is the point: as written it constrains nothing.

Warning: tool "add"
  Path: inputSchema.properties.a
  Issue: `$ref` points outside this document (`https://example.com/schemas/number.json`). Clients do not fetch remote schemas, so the constraint is dropped or the tool is rejected.
  Suggestion: Inline the referenced schema, or move it into `$defs` and reference it as `#/$defs/<name>`.

1 error, 3 warnings across 3 tools.

Note the shape of that suggestion. Where a replacement genuinely narrows the schema it says so, rather than implying an equivalent rewrite — and where an equivalent exists it gives that instead: an array-form type becomes anyOf branches (not "drop it from required", since absent is not the same as null), and a bare false becomes {"not": {}} (not "delete the entry", which would permit the property under the default additionalProperties).

Severity decides the exit code, not the report: errors are constructs a shipping MCP client refuses outright (a bare true or false where a schema object belongs), warnings are ones handled unevenly (an array-form type, a remote $ref, a schema carrying no constraining keyword at all). Only errors fail the run — a --strict that failed on warnings would be unusable as a CI gate against servers that are in fact fine.

There is deliberately no "inputSchema must be an object" check, even though MCP requires one. The SDK types inputSchema with type: literal("object"), so a tool whose input root is anything else fails ListToolsResultSchema and is dropped from the list before the lint could see it — it is reported through the malformed-items path instead. A rule that cannot fire would only make this documentation claim a check the CLI does not perform.

The report goes to stderr, so the result on stdout stays parseable. Under --format json the findings are folded into the same envelope instead ({"result":…,"schemaFindings":[…]}), so a caller reads one document rather than correlating two.

Without --strict nothing changes except one line. A tools/list whose schemas have findings prints a single stderr summary — Schema portability: 1 error, 3 warnings across 3 tools. Re-run with --strict for details. — and still exits 0. A clean list prints nothing at all.

This is deliberately not a JSON Schema validator. A census of 617 public servers found zero that fail the SDK's own parser, so a conformance check would report nothing on essentially every real server; what bites is the narrower subset each consumer accepts, which is what these rules encode. The same verdict drives the TUI's tool detail pane and the web Tools tab — all three read core/json/schemaLint.

CLI-specific (OAuth for HTTP servers)

The CLI runs the same loopback callback server as the TUI (http://127.0.0.1:6276/oauth/callback by default).

CLI (mcp-inspector --cli): on connect 401 or mid-session interactive auth (re-login / step-up), it:

  1. Starts the callback listener on --callback-url (or MCP_OAUTH_CALLBACK_URL)
  2. Prints the authorization URL to stderr (OSC 8 hyperlink when stderr is a TTY) and opens the default browser when allowed. Default: open on a stderr TTY (so 2>&1 | tee prints a plain URL into the log but does not auto-open); plain URL only when stderr is non-TTY. MCP_AUTO_OPEN_ENABLED=false never opens; =true forces open even on a non-TTY (same as the web launcher).
  3. Waits for the browser redirect, exchanges the code, and retries connect or the failed RPC

Interactive OAuth (connect-time or mid-RPC) requires a TTY on stdin or stderr (so 2>&1 | tee still works — stdin stays a TTY), or MCP_AUTO_OPEN_ENABLED=true. When neither stdin nor stderr is a TTY and that env is unset (typical CI), the CLI fails fast with auth_required instead of waiting up to 15 minutes on the loopback callback. Use --stored-auth-only for non-interactive runs that should only consume the shared store.

Step-up (standard OAuth): when an RPC needs extra scopes, the CLI prompts on stderr: Proceed with step-up authorization? [y/N]. y continues (including piped stdin — echo y | … or printf y | …); N or EOF with no answer (< /dev/null / Ctrl-D) declines. Piped answers must be newline-terminated, or stdin must close — a bare y held open without \n or EOF is not flushed as a line and times out. A non-TTY stdin that never sends a line within 5 seconds fails with auth_required (timed out, not the same as an explicit N). Answering y only confirms step-up — the following browser/loopback OAuth can still wait up to 15 minutes; for headless CI prefer --stored-auth-only with tokens already in the store. EMA step-up re-mints silently (no prompt).

Shared OAuth storage: the CLI reuses tokens stored by other Inspector clients — indexed by the shared ~/.mcp-inspector/storage/oauth.json, with the tokens themselves held in the secret store. That is passive storage sharing, not launching another app.

Shared with TUI (config only, not interactive login):

  • Per-server OAuth fields from mcp.json (static client, EMA resource credentials, scopes)
  • Install-level settings from ~/.mcp-inspector/storage/client.json (or --client-config / MCP_CLIENT_CONFIG_PATH) — EMA IdP, CIMD
  • CLI flags --client-id, --client-secret, --client-metadata-url override client.json when set
  • Keychain-backed secrets in mcp.json are rehydrated on catalog load (same as TUI)

OAuth callback URL

Surface Default callback
Web http://localhost:6274/oauth/callback
TUI http://127.0.0.1:6276/oauth/callback (interactive — callback server)
CLI http://127.0.0.1:6276/oauth/callback (interactive — same callback server as TUI)

Register http://127.0.0.1:6276/oauth/callback on static or enterprise IdPs that require pre-registered redirect URIs before using the TUI or CLI. Override with --callback-url or MCP_OAUTH_CALLBACK_URL. Only one process should bind the default port at a time.

Flags

Option Env Description
--client-config <path> MCP_CLIENT_CONFIG_PATH Install-level client config (default: ~/.mcp-inspector/storage/client.json).
--client-id <id> — OAuth client ID (static client); overrides client.json.
--client-secret <secret> — OAuth client secret; overrides client.json.
--client-metadata-url <url> — CIMD metadata URL; overrides client.json.
--callback-url <url> MCP_OAUTH_CALLBACK_URL Redirect URI sent to the authorization server (default: http://127.0.0.1:6276/oauth/callback). Must bind a loopback host (localhost / 127.0.0.0/8 / [::1]) — the listener receives the authorization code over plaintext http, so a non-loopback host hard-errors (no opt-in; use a port-forward if the browser is elsewhere).
— MCP_AUTO_OPEN_ENABLED Browser auto-open and non-TTY interactive-OAuth admit: true (allow interactive OAuth without a TTY and force-open the browser — same as the web launcher), false (never open), unset (open on a TTY unless VITEST is set). The authorization URL is always printed first; an open that fails, or does not launch within 5s, adds a stderr line saying to open it by hand. For CI that must not hang, prefer --stored-auth-only.

Example — list tools on an OAuth-protected server using stored tokens and CIMD from the command line:

npx @modelcontextprotocol/inspector --cli --catalog mcp.json --server my-http-server \
  --client-metadata-url https://example.com/.well-known/oauth/client-metadata.json \
  --method tools/list

See EMA / enterprise-managed auth and OAuth smoke testing (§3 Stytch/CIMD; §5 mid-session manual validation — CLI C1–C2).

Stored-auth (web → CLI handoff)

For the common case where OAuth was already completed in the web inspector on the same machine, the CLI can reuse the resulting token instead of running its own interactive flow. It reads the shared OAuth state (the oauth.json the web backend writes, joined with the tokens in the secret store) and injects Authorization: Bearer <token> for --server-url.

Option Description
--use-stored-auth Read the stored auth for --server-url and inject Authorization: Bearer. When a refresh_token is stored, the CLI runs the OAuth refresh grant first and injects the fresh access token (persisting the rotation); otherwise it injects the stored access token. Exits 3 (no_stored_token) — listing the stored keys — when nothing matches. Requires --server-url.
--wait-for-auth <sec> Poll the OAuth state file (500 ms interval) until an access token for --server-url appears, then inject it. Times out at <sec> with exit 3 (auth_wait_timeout). Use after handing off to a human to complete OAuth in a browser. Unlike --use-stored-auth, this injects the freshly-landed access token directly (a token that just completed the browser flow is not expired), so it does not run the refresh grant.
--list-stored-auth Print { oauthStatePath, storedServerUrls } (the server keys that currently have a token) and exit. No server connection is made.
--print-handoff Print a JSON handoff block (deepLink, portForwardCmd, oauthStatePath, apiToken) for --server-url and exit — everything a script/remote VM needs to drive the browser-side OAuth dance. The deepLink is https when both INSPECTOR_TLS_KEY_PATH and INSPECTOR_TLS_CERT_PATH are set, matching a web UI served over TLS. Requires --server-url.

State-file resolution follows MCP_INSPECTOR_OAUTH_STATE_PATH → <MCP_STORAGE_DIR>/oauth.json → ~/.mcp-inspector/storage/oauth.json — the same precedence the rest of the Inspector uses, so the CLI and web backend agree on the file. Server keys are canonicalised with new URL().href (the scheme the web store writes), so a trailing-slash or case mismatch between the URL a human opened and the one the agent passed still resolves.

Token refresh (#1665). When the stored server state carries a refresh_token (plus the clientInformation the web inspector persists after a completed flow), --use-stored-auth runs the SDK's OAuth refresh_token grant to mint a fresh access token before connecting, then writes the rotated tokens back to the state file (owner-only 0o600, via the shared store writer) so web and CLI stay consistent. The write-back updates only the refreshed issuer slot's tokens, against the entry as it is on disk at write time, so a same-server change made while the grant was in flight (say, a web OAuth flow against another authorization server) is kept rather than overwritten, and a slot whose tokens were cleared or replaced meanwhile is left as it is (#2574). The auth-server metadata is reused from the stored state, or discovered from --server-url when absent. This covers both an absent and an expired stored access token — the persisted blob carries no expiry, so the refresh token is treated as the durable credential. If the refresh fails (revoked token, transient auth-server error) and a stored access token is also present, the CLI falls back to injecting that token rather than hard-failing; only when there is nothing to fall back on does it exit 3 (auth_required, envelope refresh_failed). Without a stored refresh_token the access token is injected as-is, and a stale one surfaces as an HTTP 401 → exit 3.

Because there is no stored expiry, a refresh_token is refreshed on every --use-stored-auth run. Two consequences with rotating refresh tokens: two concurrent invocations against the same state file race the single-use token (one wins), and a crash between a successful grant and the write-back leaves the rotated token unsaved. Both are narrow; re-authorize in the web inspector to recover.

Short-circuit modes. --list-stored-auth and --print-handoff each print their output and exit without connecting to a server; they ignore the method/target flags. They are mutually exclusive — if both are passed, --list-stored-auth takes precedence.

The deepLink is the canonical web deep-link (#1576) — http://<host>:<port>/?serverUrl=<url>&transport=<http|sse>&autoConnect=<token> — so navigating it in a browser reaches a connected inspector in one shot. transport is derived from the resolved server (--transport, else auto-detected from the URL path: /sse → sse, else http), not hardcoded. autoConnect is set to MCP_INSPECTOR_API_TOKEN; when that env var is unset the link is still emitted but a note field flags that the web app's autoConnect gate will reject it until the inspector is launched with a known token.

# On a remote VM: print what a human needs to complete OAuth in their browser.
mcp-inspector --cli --server-url https://api.example/mcp --print-handoff

# Then block until the token lands and run the call with it.
mcp-inspector --cli --transport http --server-url https://api.example/mcp \
  --wait-for-auth 120 --method tools/list

Skill verification (--verify)

SEP-2640 puts real obligations on whoever consumes a skill: verify each fetched file against the digest its manifest advertised, check that the served SKILL.md's frontmatter matches the one the listing advertised, and honour the per-skill limits. --verify runs all of them over a whole catalog and turns the answer into an exit code, so a server author can gate CI on it:

mcp-inspector --cli <server> --method skills/list --verify

Stdout is NDJSON, one report per skill, in listing order:

{
  "uri": "skill://tampered-notes/SKILL.md",
  "name": "tampered-notes",
  "conformance": [],
  "frontmatter": [],
  "files": [
    { "uri": "skill://tampered-notes/SKILL.md", "status": "verified", "…": "…" },
    { "uri": "skill://tampered-notes/notes.md", "status": "mismatch", "…": "…" }
  ],
  "ok": false
}

Stderr gets a one-line summary, so a reader who piped stdout into jq still sees the verdict — and then, on a failing run, the ordinary ErrorEnvelope line that every non-zero exit writes. Two stderr lines on failure, one on success, which is the same shape --strict produces and is why the envelope is not suppressed here: a caller branching on .code should not have to special-case this command. --method skills/get --uri <skill> verifies exactly one skill, in the same shape.

What fails the run. Four outcomes, because "this skill is wrong", "this skill could not be fully checked" and "this skill offered nothing to check" are different answers:

outcome Exit When
verified 0 Everything was checked and everything passed.
failed 7 Something SEP-2640 makes a MUST was broken — an error-severity finding, a digest or size mismatch, or an unreadable manifest file.
incomplete 8 Nothing checked was wrong, but the read bounds stopped the walk before it finished. See incomplete in the report for the reason.
unverifiable 0, or 9 with --require-digests Nothing checked was wrong, but resources is "dynamic": no digest was advertised, so nothing was hashed. SKILL.md is still read for the frontmatter cross-check.

When a catalog mixes them, the exit code is the loudest: 7 over 8 over 9.

The run is bounded, and says when a bound bit. Three limits, all reported as incomplete (8) rather than as a pass or a failure, because an entry that was not read has not been cleared of anything:

Bound Limit Why
Per skill 512 manifest entries / 16 MiB SEP-2640's own interoperability limits.
Per skill, on the wire 16 MiB actually served The declared sizes are server-controlled; this one cannot be lied past.
Per run 256 skills / 64 MiB SEP-2640 bounds a skill and deliberately does not bound a catalog. Every entry costs at least one resources/read, so without this a large listing — hostile or merely big — is unbounded work against the tool inspecting it.

The run bound is this tool's, not the spec's, and it is configurable per server: set skillCatalogMaxSkills / skillCatalogMaxBytes on the server's entry in mcp.json, or edit them under Skills in the web client's Server Settings (see the configuration reference). Both must be positive integers — there is no unlimited value, since the bound is what makes the run terminate. A skill past it is still reported, with its static conformance findings and an incomplete reason saying nothing about its files was checked; verify it on its own with --method skills/get --uri <skill> to get a verdict for it.

A warning never produces 7. That distinction matters most for resources: "dynamic", which is a conforming wire form for generated content: it means integrity cannot be verified, which is worth reporting, but failing CI for it would tell server authors their valid skill is broken. It is not verified either — SEP-2640 calls such a SKILL.md "unverifiable", and reporting it with the same outcome as a skill whose every file hashed clean gave CI no way to tell them apart (#2405). So it gets its own outcome, and the stderr headline says how many skills advertised no digests. SEP-2640 also says "Hosts MAY decline to load such skills"; a CI job standing in for such a host passes --require-digests to turn that outcome into exit 9.

Three checks, three different jobs, and the second is the one nothing else covers:

  • Conformance — structural checks against the entry as listed (name grammar, the name/URI invariant, digest and size formats, manifest completeness, the interoperability limits).
  • Frontmatter — the served SKILL.md's own YAML frontmatter, compared field by field against the frontmatter the listing advertised. A digest cannot cover this: it is taken over the bytes the server served, so it proves the file was not altered in transit and says nothing about whether the listing described it honestly. A server can advertise one description, serve another, and pass every digest check.
  • Files — each manifest entry fetched and hashed. Reads are sequential: a conforming manifest may declare 512 entries, and a parallel walk would open 512 resources/read calls against the server under test.

A read failure is recorded against the file it happened on and the walk continues, so one unreadable file never hides the findings after it.

Exit codes & error envelopes

Every non-zero exit maps to a stable failure class, so a programmatic caller (CI, a script, an agent) can branch on why the CLI failed without scraping prose from stderr:

Code Meaning
0 Success.
1 Usage / unexpected error (the catch-all).
2 No MCP App found on the tool (--app-info probe).
3 Server requires authentication (401/403 or a typed SDK auth error).
4 Server unreachable (DNS, connection refused, timeout, fetch failed).
5 Tool error (tools/call returned isError:true, or the tool was not found).
6 --strict found an error-severity tool-schema portability problem (schema_unportable — the schema is valid JSON Schema, just not portable).
7 --verify found a SEP-2640 violation (skills_nonconformant — a conformance error, a digest or size mismatch, or an unreadable manifest file).
8 --verify could not check the whole catalog (skills_incomplete — the read bounds stopped the walk). The server broke no MUST: the 512-entry and 16 MiB limits are SHOULD NOT, and hosts may support more. A job that tolerates oversized catalogs can allow 8 and still fail on 7.
9 --verify --require-digests found a skill that advertised no digests (skills_unverifiable — resources: "dynamic"). Only produced under --require-digests; without it such a skill is reported unverifiable and the run exits 0.

On any non-zero exit the CLI also writes a single JSON line to stderr — the ErrorEnvelope:

{
  "error": {
    "code": "auth_required",
    "message": "Unauthorized",
    "status": 401,
    "url": "https://api.example/mcp"
  }
}

The code is a stable identifier for the failure class; message is the human-readable error; cause, status, and url are included when known. Because it is one line, a caller can parse it with 2>&1 | tail -1 | jq .error.

Why use the CLI?

For a copyable connect → list → call → assert workflow built on the flags above, see Smoke-testing an MCP server.

While the Web Client provides a rich visual interface, the CLI is designed for:

  • Automation: Ideal for CI/CD pipelines and batch processing.
  • AI Coding Assistants: Provides a direct, machine-readable interface (JSON) for tools like Cursor or Claude to verify changes immediately.
  • Log Analysis: Easier integration with command-line utilities (like jq) to process and analyze MCP server output.

Development

Like the other clients, the CLI self-validates from its own folder:

npm run check          # format:check && lint && typecheck  (no tests)
npm run validate       # check && test  (fast; no coverage gate)
npm test               # build test-servers + binary, then run all tests
npm run test:coverage  # build + tests under the per-file ≥90 coverage gate

The CLI's test / test:coverage build the binary first (out-of-process e2e.test.ts spawns it). check is format:check && lint && typecheck and validate is check && test, with no separate build step (pretest builds). Repo-root validate:cli delegates to validate, and the root local:validate — the first stage of npm run local:gate — runs check instead, so the gate runs the suite once, under coverage / coverage:cli (#2341), matching AGENTS.md.

Tests run the CLI in-process (importing runCli()) so src/ is measured under coverage, with a thin out-of-process spawn layer for the real binary. See __tests__/README.md for details.