Skip to content
MCP servers with an agent· Guide 9 of 10Intermediate8 min readUpdated

Test MCP clients with the time and everything servers ​

The reference everything server exists to test MCP clientsMCP client: The part of an AI application that holds one connection to one MCP server: it negotiates the protocol version and capabilities, then sends that server's requests and receives its replies.: it has a tool, resource or prompt for nearly every protocol feature, including progress notifications, elicitation, sampling, list changes and both error shapes. The reference time server is the opposite, three small read-only tools with structured output, which makes it a good first real server. Neither needs an account, a key or a network service, so they are the quickest way to find out what a client actually supports before you connect anything that matters.

Status of both servers (November 2026) ​

Both are current reference servers in modelcontextprotocol/servers, not among the servers moved to servers-archived. The repository describes its servers as "reference implementations to demonstrate MCP features and SDK usage", "not as production-ready solutions".

ServerPackageLatest releaseRuntime
everything@modelcontextprotocol/server-everything (npm)2.3.0, 2026-11-11Node.js 24 or later
timemcp-server-time (PyPI)2026.11.11, 2026-11-11Python 3.10 or later, usually through uvx

Both are released often. Pin the version you tested, so a new release does not change your test fixture under you.

Run them ​

bash
# everything: stdio (the default) or Streamable HTTP on http://127.0.0.1:3001/mcp
npx -y @modelcontextprotocol/server-everything@2.3.0 stdio
npx -y @modelcontextprotocol/server-everything@2.3.0 streamableHttp

# time: stdio only, with a fixed default timezone
uvx mcp-server-time==2026.11.11 --local-timezone=UTC

The everything server's HTTP mode binds to 127.0.0.1 and accepts only localhost browser origins by default; HOST=0.0.0.0 exposes it, which you should not need. In a container, the time server cannot see the host timezone, so it reads LOCAL_TIMEZONE and falls back to UTC.

What the everything server exercises ​

Version 2.3.0 serves two protocol eras from one process. A client that opens with initialize gets the 2025-era behaviour; a client that opens with server/discover gets MCP 2026-07-28, where capabilities travel with each request. That changes what you see. Opened with initialize and an empty capability set, it listed 16 tools, 4 prompts, 8 resources and 4 resource templates. Tools that need a client capability (elicitation, sampling, roots) appear only if the client declared it at initialize; on 2026-07-28 they are always listed and a call is refused (-32021) when the request does not declare what it needs.

Here is a checklist, one feature per line, with the thing to call:

Client featureCallWhat a working client shows
Tool call, input validationget-sum with {"a": 2, "b": "x"}A tool result with isError: true: Input validation error: ... b: Invalid input: expected number, received string
Unknown tooltools/call for nopeA JSON-RPC error, code -32602, Tool nope not found
Structured contentget-structured-contentstructuredContent that matches the tool's outputSchema, plus the same JSON as text
Content typesget-tiny-image, get-annotated-message, get-resource-links, get-resource-referenceAn image block, content annotations (priority, audience), resource_link items and an embedded resource
Progresstrigger-long-running-operation with {"duration": 2, "steps": 2} and a progressTokenTwo notifications/progress (1 of 2, 2 of 2), then the result
Cancellationthe same tool with a long duration, then cancelThe server stops its timer and sends no more progress
List changestrigger-list-changed with {"list": "tools"}notifications/tools/list_changed, and the client lists tools again
Resourcesresources/list, resources/templates/list, read demo://resource/dynamic/text/1Text and blob contents; a read of demo://resource/dynamic/text/abc fails with -32602 and the URI in data
Promptsargs-prompt with city (required) and stateRendered messages; completable-prompt also answers completions
Elicitationtrigger-multi-step-elicitationTwo questions in a row (a name, then a colour from three choices)
Samplingtrigger-sampling-requestA request for the client's model, if the client allows it
Loggingtrigger-log-messagesOne log message per level at or above the level the client set

Two tools deserve care even in a test:

  • get-env returns the server's environment as JSON. Values of credential-like names (*_KEY, *_TOKEN, *_PASSWORD and similar) are replaced with "<redacted>" unless GET_ENV_INCLUDE_SECRETS is set to a true value (1, true, yes or on). Leave it unset: with it, secrets go straight into the model's context.
  • gzip-file-as-resource fetches a URL. It is annotated openWorldHint: true and is not read-only. The README says it refuses non-public addresses and follows at most five redirects, but that the address check does not catch DNS rebinding. Restrict it with GZIP_ALLOWED_DOMAINS, or deny the tool in the client.

What the time server exercises ​

The time server has three tools, all annotated read-only, idempotent and closed-world, each with an output schema:

ToolArgumentsReturns
get_current_timetimezone (optional)timezone, datetime, day_of_week, is_dst
convert_timetime as HH:MM (required), source_timezone, target_timezonesource, target, time_difference, later_occurrence
get_time_agotimezone, days_ago, hours_ago, minutes_ago, seconds_agothe time that far before now

Converting 16:30 from America/New_York to Asia/Tokyo on 2026-11-16 returned "datetime": "2026-11-17T06:30:00+09:00" and "time_difference": "+14.0h", as structured content and as text. Its errors are tool results, not protocol errors: "timezone": "europe/warsaw" returns isError: true with Error executing tool get_current_time: Invalid timezone: 'No time zone found with key europe/warsaw', because IANA names are case-sensitive, and "time": "25:99" returns Invalid time format. Expected HH:MM [24-hour format]. An unknown tool name is a -32602 JSON-RPC error. It has no prompts or resources; prompts/list comes back empty.

That makes it a good agent test. A model that sends europe/warsaw should read the error and retry with Europe/Warsaw. If your agent instead stops or invents an answer, the problem is in how it handles tool errors, not in the server.

Its answers depend on the clock, so a test that compares results must either use a recording or check the shape (datetime is a string with an offset) rather than the value.

Run them in L2M ​

L2M is a self-hosted workflow builder and agent runtime, and its MCP client is what you are testing. Register each server under MCP servers → + Register server with the stdio_mcp adapter:

json
{ "command": "npx", "args": ["-y", "@modelcontextprotocol/server-everything@2.3.0", "stdio"], "timeoutMs": 60000 }
json
{ "command": "uvx", "args": ["mcp-server-time==2026.11.11", "--local-timezone=UTC"], "timeoutMs": 30000 }

The sample deployment catalog, samples/mcp-catalog.example.json, has an approved time server entry run with uvx, pinned to an older release (mcp-server-time==0.6.2). With MCP_CATALOG_ENFORCE=true a registration must use exactly the pinned package, so update the pin in your catalog to the version you tested (deployment catalog). Then open MCP → Playground: pick a tool, fill in the form generated from its schema and read the JSON-RPC exchange panel, which shows the tools/call request and the result as L2M received them. The Resources and Prompts panels read demo:// resources and render the everything server's prompts (playground).

Running the everything server through L2M's stdio client turns the checklist into concrete answers:

  • Era. With the default protocol mode (legacy), L2M opens with initialize. It always declares elicitation, so the server adds trigger-elicitation-request and trigger-multi-step-elicitation to the 16 tools above; workflow runs also declare sampling (and refuse it per server, see below), which adds trigger-sampling-request. With "protocolMode": "auto" L2M agrees on 2026-07-28 and sees 22 tools, including the task and roots tools (MCP 2026-07-28 support).
  • Progress. L2M sends a progressToken and receives both progress notifications. In a run they appear on the run tape as tool progress (progress and cancellation).
  • List changes. trigger-list-changed arrives as a list-changed event, and the registry re-probes the server a couple of seconds later.
  • Elicitation. trigger-multi-step-elicitation asks its two questions and completes with Ada's favorite color is red. once both are answered. trigger-elicitation-request fails: its form includes a multi-select field, and L2M accepts only flat string, number, integer, boolean and enum fields, so it reports nested schemas are not allowed. In a run, an attended run shows the questions as a form; unattended runs (webhooks, schedules) decline at once, and the playground declines too (elicitation).
  • Sampling. Off by default for every server. Until an admin turns it on for the everything server, trigger-sampling-request gets Sampling is off for MCP server ... (sampling).
  • Roots. L2M does not declare the roots capability, so get-roots-list is not listed on the 2025-era connection.

None of this needs a model. A workflow with one MCP Tool node per tool, each with a fixed argsTemplate, walks the checklist on every run.

Keep it least-privilege ​

These are test servers, but they still run as processes on the API host. Optionally:

  • Deny get-env and gzip-file-as-resource with MCP tool policy rules on the registered everything server, so no agent sees them.
  • Run the everything server in a container sandbox with no network. gzip-file-as-resource then cannot fetch anything, and the rest of the checklist still works. Without network npx cannot download the package, so use an image built from src/everything/Dockerfile in the servers repository, with a version tag (the sandbox refuses latest). Its entrypoint is node dist/index.js and it reads the first argument as the transport, while L2M appends the registered command and arguments after the image name. Register the command as just stdio with no arguments; npx -y ... there would reach the server as an unknown transport and stop it.
  • Keep sampling off unless you are testing it, and then cap it with Max tokens per request and Max requests per tool call.

Record them for CI ​

Both servers make good fixtures for a client test suite:

  1. Turn on Record MCP calls and run the checklist workflow once. The cassetteCassette: A recording of the outside calls a run made, with their answers, so a later test run can be answered from it instead of calling the real services. holds every request, result and error.
  2. Replay the run in strict mode after changes. The answers come from the cassette, so the time server's clock stops mattering and neither server has to be installed in CI.
  3. Export each server's snapshot. When a new release comes out, run l2m mcp check --snapshot time.snapshot.json --stdio "uvx mcp-server-time==<new version>" in CI before you bump the pin, so a release that removes a tool or breaks its arguments fails the build.

Test MCP servers in CI has the full commands. For a server of your own that works the same way, see build a minimal MCP server.

Try it in L2M ​

The quickstart workflow attaches two tools of the built-in mock MCP server, get_current_time and calculator, to an agent's tool port as MCP Tool nodes. Its chat model is the echo provider, which never asks for a tool, so the demo shows the wiring only. The checklist workflow above is shaped differently: there the MCP Tool nodes are steps in the flow itself, each calling its tool with its fixed argsTemplate, so no model is involved.

Further reading ​

Last updated:

This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings