client-smoke — скилл для Claude Code: смоук-тест MCP-сервера

---
name: client-smoke
description: "Drive a locally built server with a real client: the Inspector CLI (scripted), the Inspector web UI (by hand) and an LLM client. Use when smoke-testing a server change; when calling a tool, reading a resource or getting a prompt from the command line; when choosing a protocol era; or when the Inspector CLI will not connect to, or misreads, a server command."
disable-model-invocation: false
---

# Client smoke

A server change is checked against real clients before its PR opens: the MCP
Inspector and an LLM client. This skill is how to run each one against a server
built from this checkout, and what the output means. What the PR must then
record is in `/pr-flow` (client evidence); building and starting the server is
`/local-dev`; automated tests are `/testing`.

Three clients, in the order to reach for them:

| Client | Driven by | Use it for |
| --- | --- | --- |
| Inspector **CLI** | A command. This is the **scripted path**, and the one an agent runs | One request, one JSON result: the evidence that goes in a PR |
| Inspector **web** UI | A person, by hand. Nothing in this repo automates a browser | What only a UI shows: notifications arriving, elicitation and sampling prompts, a long-running tool's progress |
| An **LLM client** | A prompt | Whether a model picks the tool and uses its result, which the PR checklist asks for |

Build first (`npm run build -w src/<server>`). Every command below runs the
last build of a TypeScript server, not your edit.

## The Inspector CLI

```sh
npx -y @modelcontextprotocol/inspector --cli <server command> [-- ] <inspector options>
```

It starts the server command over stdio (or connects to a URL), sends one
request, prints the result and exits. Pin the Inspector version in anything you
keep (`@modelcontextprotocol/inspector@<version>`); the commands here were run
on 2.9.0. `--cli --help` is the reference for the options.

```sh
INSPECT="npx -y @modelcontextprotocol/inspector --cli"

# List, then call, on a TypeScript server
$INSPECT node src/everything/dist/index.js stdio --method tools/list --format json
$INSPECT node src/everything/dist/index.js stdio \
  --method tools/call --tool-name echo --tool-arg message=hello --format json
# → {"result":{"content":[{"type":"text","text":"Echo: hello"}]}}

# Arguments that are not strings: pass them as JSON
$INSPECT node src/everything/dist/index.js stdio \
  --method tools/call --tool-name get-sum --tool-args-json '{"a":2,"b":3}' --format json

# Resources and prompts
$INSPECT node src/everything/dist/index.js stdio \
  --method resources/read --uri demo://resource/dynamic/text/1 --format json
$INSPECT node src/everything/dist/index.js stdio \
  --method prompts/get --prompt-name args-prompt --prompt-args city=Paris --format json

# A server that takes a positional argument
$INSPECT node src/filesystem/dist/index.js /abs/dir \
  --method tools/call --tool-name list_allowed_directories --format json

# A Python server: note the `--` (see below)
$INSPECT uv --directory src/time run mcp-server-time -- \
  --method tools/call --tool-name get_current_time --tool-arg timezone=UTC --format json
$INSPECT uv --directory src/git run mcp-server-git --repository "$PWD" -- \
  --method tools/list --format json
```

`--format json` prints a result as one JSON object, `{"result": …}`, on
**stdout**, and a failure as `{"error": …}` on **stderr**. The server's own
stderr (its startup banner) passes through on stderr too. So capture stdout
alone for the result, never `2>&1` into a JSON parser: a tool error produces
both objects, one on each stream (see the table below).

### How the command line is split

The server command comes **first** and the Inspector's options **after** it.
With no `--`, the server command ends at the first word that starts with `-`.
That rule produces the three failures people actually hit:

- **A server command with its own flags needs `--` before the Inspector's
  options.** `uv --directory …` and `mcp-server-git --repository …` both have
  one. Without the separator the command is cut at `--directory`, and what runs
  is a bare `uv`, which prints its help and exits. With it, everything before
  `--` is the server command and everything after is the Inspector's.
- **Putting an Inspector option before the server command loses the command.**
  `--cli -e KEY=VALUE node …` leaves no server command at all, and the CLI
  falls back to your personal catalog (`~/.mcp-inspector/mcp.json`). With
  several servers in it that fails with `Multiple servers found in config
  file. Please specify one with --server`. Either way it is no longer talking
  to your build, so move the option after the command.
- **`Method is required`** means the options ended up on the wrong side of a
  `--`.

### Environment variables

The spawned server gets only the SDK's default safelist from your shell
(`PATH`, `HOME`, `USER` and the like), not the variables you set. Setting
`MEMORY_FILE_PATH=… npx …` has no effect on the server; pass it with `-e`:

```sh
$INSPECT node src/memory/dist/index.js \
  -e MEMORY_FILE_PATH=/abs/tmp/graph.jsonl \
  --method tools/call --tool-name read_graph --format json
```

### Exit codes and errors

The exit code is the verdict, so check it rather than reading the text:

| Outcome | stdout | stderr | Exit |
| --- | --- | --- | --- |
| The request succeeded | `{"result": …}` | | 0 |
| The tool ran and returned `isError: true` | `{"result":{"content": …, "isError": true}}` | `{"error":{"code":"tool_is_error", …}}` | 5 |
| No such tool | | `{"error":{"code":"tool_not_found", …}}` | 5 |
| An HTTP server could not be reached (refused, DNS, timeout) | | `{"error":{"code":"unreachable", …}}` | 4 |
| Bad arguments, a failed version negotiation, or a stdio server that exited | | `{"error":{"code":"error", …}}` | 1 |

These are the codes observed on 2.9.0 for the cases this repo's servers
produce; the CLI defines others (schema and skill checks, authentication)
that its `--help` describes.

A tool error is a **successful protocol exchange**: the result object is still
printed on stdout, and it is what to quote when the change is about an error a
tool returns.

### `everything` over HTTP

Start the transport in one terminal (or in the background), then give the CLI
the URL. The transport is inferred from the path: `/mcp` is Streamable HTTP,
`/sse` is HTTP+SSE.

```sh
(
  URL=http://localhost:3917/mcp
  # Refuse a port something already answers on: it would be tested instead of this build.
  curl -s -m 2 -o /dev/null "$URL" && { echo "port 3917 is already in use" >&2; exit 1; }
  PORT=3917 node src/everything/dist/index.js streamableHttp >/dev/null 2>&1 &
  SERVER_PID=$!
  trap 'kill "$SERVER_PID" 2>/dev/null' EXIT    # stops it however the subshell ends
  READY=
  for _ in $(seq 50); do                         # at most ten seconds
    kill -0 "$SERVER_PID" 2>/dev/null || { echo "server exited before listening" >&2; exit 1; }
    curl -s -m 2 -o /dev/null "$URL" && { READY=1; break; }
    sleep 0.2
  done
  [ -n "$READY" ] || { echo "server did not listen within ten seconds" >&2; exit 1; }
  $INSPECT "$URL" --method tools/call --tool-name echo --tool-arg message=over-http --format json
)
```

Each part of that is there for a failure it prevents. The first check refuses
a port that already answers, because a stale listener there would be the
server under test. Without the wait, the CLI can reach the port before the
server is listening and exit 4. The wait is bounded and checks that this
server is still alive, so one that fails to bind (the build is missing, the
port was taken in between) ends the run instead of hanging it. The subshell's `EXIT` trap stops the server by pid even when the CLI call
fails; a listener left on the port makes the next run talk to the old build.
By hand in two terminals, the equivalent is waiting for the `listening on
port` line and stopping the server with Ctrl-C.

A change to a tool, resource or prompt of `everything` is checked over stdio
and Streamable HTTP. Check HTTP+SSE as well when the change touches transport
or session code; it is deprecated but must keep working.

## Protocol eras

The Inspector connects in one of three eras, chosen with `--protocol-era` (the
CLI) or **Server Settings → Protocol Era** (the web UI):

| Era | What the Inspector does |
| --- | --- |
| `legacy` (the default) | A plain `initialize`. Negotiates 2025-11-25 or older |
| `modern` | Pins 2026-07-28 through `server/discover`, with no fallback |
| `auto` | Probes `server/discover`, and falls back to `initialize` |

**Every server serves both eras**: the Python servers since #4853, the
TypeScript servers since #4852 (stdio through the SDK's `serveStdio`, and
`everything`'s Streamable HTTP endpoint through `createMcpHandler` for
2026-07-28 requests). Run every command twice, with `--protocol-era legacy`
and `--protocol-era modern`, and expect both to succeed. A 2026-07-28 result
carries `io.modelcontextprotocol/serverInfo` in its `_meta`, which is the
quick way to see which era answered. Do not use `auto` for evidence: it does
not say which era answered.

`everything` lists some tools on one era only (the task tools on 2026-07-28,
the two `toggle-*` tools on the 2025 era), so a `tools/list` that differs
between the two runs is expected.

The CLI declares no client capabilities, so a tool that asks the client for
input (elicitation, sampling, roots) cannot run from it: on the 2025 era
`everything` does not list it (`tool_not_found`, exit 5), and on 2026-07-28
the SDK refuses the call because the request did not declare the capability
(`Cannot request input 'form' (elicitation/create): the request's client
capabilities do not declare the required capability`, exit 1). Drive those
from the web UI. On 2026-07-28 the CLI does follow a task handle: a call to
`simulate-research-query` polls the task and prints its final result.

## The Inspector web UI

The web client is the hand-driven path. It needs a person at a browser; this
repo configures no browser automation, so an agent prepares the command and
the steps and asks for what was observed.

```sh
npx -y @modelcontextprotocol/inspector node src/everything/dist/index.js stdio
npx -y @modelcontextprotocol/inspector -- uv --directory src/time run mcp-server-time
npx -y @modelcontextprotocol/inspector --transport http --server-url http://localhost:3001/mcp
```

⚠️ **The web launcher reads `--` the other way round from the CLI.** Here the
Inspector's own options come **before** `--` and the server command **after**
it; in `--cli` mode the server command is before `--` and the options after.
A line copied from one mode to the other starts the wrong thing. Pass a
server's environment with `-e KEY=VALUE` before the `--`.

Use the web UI for what a single request cannot show: `toggle-simulated-logging` and
`toggle-subscriber-updates` (notifications over time),
`trigger-elicitation-request` and `trigger-sampling-request` (the client is
asked something), `trigger-long-running-operation` (progress), and a
filesystem server's response to a Roots change.

Evidence from the web UI is what was clicked and what appeared, in words. This
repo does not attach screenshots.

## An LLM client

The PR checklist asks whether the change was tested with an LLM client. Any
client that can launch a stdio server works; each server's `README.md` has the
config blocks for Claude Desktop and VS Code, and `/local-dev` shows how to
point one at the local build.

Claude Code can do it in one headless command, which makes it the scriptable
option. Write a config that names the local build with an **absolute** path:

```json
{
  "mcpServers": {
    "everything": {
      "command": "node",
      "args": ["/abs/path/to/servers/src/everything/dist/index.js", "stdio"]
    }
  }
}
```

```sh
printf '%s' "Call the echo tool of the everything server with the message 'llm-smoke' and reply with exactly the text it returned." \
  | claude -p --mcp-config /abs/path/to/config.json --strict-mcp-config \
      --allowedTools "mcp__everything__echo" --max-turns 3
# → Echo: llm-smoke
```

`--strict-mcp-config` makes the file the only source of servers, so the answer
cannot have come from a published copy configured elsewhere. A tool is
addressed as `mcp__<server name in the config>__<tool name>`, and only the
tools listed in `--allowedTools` can be called without a prompt. Keep the
config file outside the worktree.

Claude Code, run this way, has no era switch and does not print the protocol
version it negotiated. `/pr-flow` asks for the LLM client in both eras as well;
where the client gives no way to choose, record the client and its version and
say that the era could not be selected, rather than labelling the run with an
era it was not shown to use. `everything` shows which era a client
negotiated through its tool list: a 2026-07-28 client sees
`simulate-research-query` and no `toggle-*` tools, a 2025-era client the
reverse. Claude Code 2.1.296 negotiates 2026-07-28 against these servers.

## What to write down

For each client: the command or the prompt, and the result trimmed to what
shows the change. For a bug fix, run the same CLI command against `v2/main`
and against the branch, and keep both outputs. Where the evidence goes in the
PR, and what a change with no client-visible surface records instead, is
`/pr-flow`.