A small project todo board where every server-side MCP feature has a real job: tools that mutate state, resources that expose it, prompts that seed conversations, sampling that borrows the connected host's model, elicitation that asks the user, progress and logs while it works, and per-resource subscriptions that announce every change. It is a faithful port of the TypeScript SDK's examples/todos-server — think of it as the "polls app" of MCP servers: small enough to read in one sitting, real enough that nothing in it is contrived.
It serves both protocol revisions at once — 2026-07-28 and 2025-11-25 are negotiated per connection, from the same handlers — and both transports: stdio and Streamable HTTP.
From this directory:
# stdio — for hosts that spawn their servers as child processes
uv run python -m mcp_todos_server
# Streamable HTTP — for remote-style connections (default port 3000; --port or $PORT to change)
uv run python -m mcp_todos_server --transport streamable-httpOver stdio the server speaks on stdin/stdout (its own diagnostics go to stderr). Over HTTP it serves http://127.0.0.1:3000/mcp.
There is no era flag: both entries detect each connection's revision during the handshake, so a 2025-era client and a 2026-era client can talk to the same process — simultaneously, over HTTP.
Any mcpServers-style host can spawn it too:
The TypeScript SDK's reference host, cli-client, connects to the HTTP entry out of the box:
uv run python -m mcp_todos_server --transport streamable-http # terminal A, this repo
pnpm --filter @mcp-examples/cli-client start -- --server http://127.0.0.1:3000/mcp # terminal B, typescript-sdk repo| Server feature | Where it lives | Notes |
|---|---|---|
| Tools | add_task, add_tasks, list_tasks, complete_task |
plain CRUD; add_task also returns structuredContent against an outputSchema |
| Sampling | prioritize, brainstorm_tasks |
the server borrows the host's model; the host shows the request for approval first |
| Elicitation (form) | clear_done, brainstorm_tasks |
schema-driven forms; accept / decline / cancel all handled |
| Multi-round input_required | brainstorm_tasks |
theme+count form → optional custom-amount round → sampling round, as a resolver chain; the framework carries the recorded answers in sealed request_state |
| Progress | work_through_tasks, add_tasks |
paced per-task progress notifications via ctx.report_progress |
| Logging | every mutating tool, via log_info |
honours logging/setLevel on 2025 connections and the per-request log-level _meta opt-in on 2026-07-28 |
| Resources | todos://board, todos://tasks/{id} |
one concrete resource + a URI template; every task also appears in resources/list |
| Subscriptions | the board | resources/subscribe/unsubscribe handlers for 2025-era clients; subscriptions/listen streams (over HTTP) for 2026-07-28, consumed client-side with client.listen(...); every mutation notifies |
| list_changed | every mutation | resource list + resource updated notifications on both eras |
| Prompts + completions | plan-my-day, seed-board |
argument completion (project names, themes, task ids) wired to completion/complete via @mcp.completion() |
The two protocol eras differ in how interactive conversations travel: on 2025-era connections the wire carries pushed elicitation/create / sampling/createMessage requests; on 2026-07-28 the server returns input_required results and the client retries the call with the answers. The interactive tools (brainstorm_tasks, clear_done, prioritize) are written once, as resolver dependencies: a parameter annotated Annotated[T, Resolve(fn)] is filled before the tool body, and a resolver that returns Elicit(...) or Sample(...) has the framework put the question to the client over whichever transport the connection negotiated — so there is no era branch in any handler, and no round bookkeeping in the example. brainstorm_tasks shows the multi-round shape as a chain: a count/theme form, a conditional custom-amount form that only asks when the first answer was "custom", and a sampling round derived from the recorded answers.
| Env var | Effect |
|---|---|
REQUEST_STATE_SECRET |
Key for the sealed request_state (≥ 32 bytes). Unset, the SDK generates a per-process key — fine whenever a single process serves the whole flow. |
PORT |
HTTP port when --port isn't passed (default 3000). |
mcp_todos_server/
server.py transport entry: stdio by default, streamable HTTP behind --transport
todos.py the application: state, tools, resources, prompts, subscriptions — every feature above
This port is verified against the TypeScript todos-server by driving both over stdio and HTTP, on both protocol eras, through an identical scripted scenario (same tool calls, elicitation answers, and sampling replies): every tool result text, structured output, elicitation form, sampling request, progress sequence, and log line matches. Known, deliberate differences:
- JSON Schema style. Input schemas come from pydantic here and zod there, so cosmetics differ (pydantic emits
titles and$defsrefs for the nestedadd_tasksitems). The schemas are semantically identical. resources/listcomposition. The TypeScriptResourceTemplatehas alistcallback;MCPServerdoesn't, so this example overrides the low-levelresources/listhandler to append one entry per task (the same private-API pattern the everything-server uses forresources/subscribeandlogging/setLevel).subscriptions/listenover stdio. The Python SDK serves 2026-era listen streams on streamable HTTP only; over stdio a listen request is rejected. Board-change notifications over stdio therefore reach 2025-era subscribers only.- Legacy HTTP interactivity. The TypeScript server's per-request HTTP posture refuses push-style sampling/elicitation for 2025-era HTTP clients; the Python server's default Streamable HTTP mode is stateful, so those tools work on that leg here.
- Legacy HTTP fan-out. Pre-2026 board-change notifications go to the session that made the mutating call. Over stdio that is every subscriber; with several concurrent 2025-era HTTP sessions, the others don't hear about it. The subscription set is also process-wide, so a mutating session that never subscribed is notified once any other session has subscribed (the TypeScript entry scopes subscriptions per session and broadcasts via its handler notifier). Pre-2026 HTTP handshakes also advertise
listChanged: false— the SDK exposes no seam to change that on the HTTP path (stdio is patched, seeserve_stdio). - Cancellation granularity. When a client cancels
work_through_tasks, this SDK interrupts the handler at its nextawait(the in-flight pretend task stays open); the TypeScript server checks between tasks and finishes the in-flight one. - Elicited answers are schema-validated. The resolver framework validates every accepted form answer against the requested schema, so a non-conforming answer (an out-of-range custom count, a value outside the count enum) fails the call with a validation error; the TypeScript handlers parse manually and report those as polite "Nothing added" declines. On 2026-07-28 connections, both servers answer a client that never declared the sampling/elicitation capability with a structured
-32021protocol error naming the missing capability; on pre-2026 connections this server still answers-32021, where the TypeScript server reports anisErrortool result.
{ "mcpServers": { "todos": { "command": "uv", "args": ["run", "--directory", "/absolute/path/to/examples/servers/todos-server", "python", "-m", "mcp_todos_server"] } } }