mcplisten

Drain all pages from MCP SDK paginated list methods — get every tool, resource, and prompt without cursor management.

Python developers using the MCP Python SDK's Client.list_tools() only receive the first page of results when a server paginates its responses. mcplisten transparently drains every page so you get the complete inventory.

Quick Start

pip install git+https://github.com/prasad-a-abhishek/mcplisten.git

from mcplisten import list_all_tools

# Async usage (with MCP SDK Client)
from mcp import Client
async with Client(server_params) as client:
    tools = await list_all_tools(client)  # drains all pages
    print(f"Got {len(tools)} tools")

# Sync usage (with SyncMCPClient from mcpsync)
from mcpsync import SyncMCPClient
from mcplisten import list_all_tools_sync

with SyncMCPClient(params) as client:
    tools = list_all_tools_sync(client)

⚡ Performance & Benchmarks

mcplisten is a thin pagination drain helper — it adds zero computational overhead beyond the network I/O of fetching additional pages. The synchronous wrapper spawns a fresh event loop per call (thread-safe for single-shot use).

python3 benchmarks/run_benchmark.py

Benchmark environment: Python 3.11.15, Linux x86_64 (kernel 6.12.67), 4 vCPU. Note: mcplisten has no direct competitor — the MCP Python SDK itself does not provide list_all_* helpers (see modelcontextprotocol/python-sdk#2556).

Test Status

The repository ships with a 131-test pytest suite covering every spec acceptance criterion (22/22), concurrent-call safety, unicode round-trip, dict-shaped and pydantic-shaped responses, CLI end-to-end against a fake paginating stdio server, and zero-dependency enforcement. Run pytest tests/ -q to reproduce (≈1 second).

Spec: see spec.md for the canonical 22-acceptance-criteria contract that drives the test suite.

Why mcplisten?

The MCP Python SDK's ClientSession.list_tools(cursor=...) accepts a cursor parameter but requires callers to manually loop across pages. MCP gateways, Claude Code, Cursor, and the Inspector CLI all shipped with this single-page limitation — silently losing tools when servers register more tools than fit in one page.

mcplisten is the fix: a zero-dependency, zero-overhead wrapper that calls list_tools(cursor=None) then follows nextCursor until the server returns null. No SDK changes required. No configuration. Drop in.

Key Features

• list_all_tools(client) — drain all pages from list_tools(), returns list[dict]
• list_all_resources(client) — drain all pages from list_resources()
• list_all_prompts(client) — drain all pages from list_prompts()
• list_all_resource_templates(client) — drain all pages from list_resource_templates()
• list_all(client, method, items_key) — generic drain helper
• Sync wrappers — list_all_*_sync() variants for SyncMCPClient
• CLI — mcplisten list-tools -- <server-cmd>, list-resources, list-prompts
• Zero dependencies — pure Python stdlib, no third-party packages
• Full type stubs — Python 3.11+ with complete type annotations

CLI Reference

# List all tools from a stdio server
mcplisten list-tools -- python -m my_mcp_server

# List all resources
mcplisten list-resources -- python -m my_mcp_server

# List all prompts
mcplisten list-prompts -- python -m my_mcp_server

# Verbose mode — includes per-server tool counts
mcplisten list-tools --verbose -- python -m my_mcp_server

# Help
mcplisten --help
mcplisten list-tools --help

API Reference

list_all_tools(client) — async

from mcplisten import list_all_tools

tools: list[dict[str, Any]] = await list_all_tools(client)

Drains all pages from client.list_tools() (works with Client, ClientSession, and ClientSessionGroup).

list_all_tools_sync(client) — sync

from mcplisten import list_all_tools_sync

tools: list[dict[str, Any]] = list_all_tools_sync(client)

Synchronous wrapper for use with SyncMCPClient (from the mcpsync extra).

list_all(client, method, items_key) — generic

from mcplisten import list_all

result: list[dict[str, Any]] = await list_all(client, "list_tools", "tools")

Generic drain helper for any paginated list method. items_key is the dict key holding the items list in each page response.

CLI

mcplisten [--help]
  list-tools [--verbose] [--] <server-cmd>...
  list-resources [--] <server-cmd>...
  list-prompts [--] <server-cmd>...
  list-resource-templates [--] <server-cmd>...

Exits 0 on success (print JSON to stdout), non-zero on error (print message to stderr).

Limitations

• Cursor timing: if a server adds a tool between page 1 and page 2 of a drain, the caller may see duplicates. This is consistent with the MCP SDK's at-most-once semantics.
• No streaming variant: memory-constrained streaming (async generator) is a separate future feature.
• SDK version sensitivity: if the MCP SDK changes its pagination API (method signature or cursor field location), this library will break. Tests pin against a known SDK version.
• No parallelization: ClientSessionGroup drains each session sequentially, not in parallel.

Non-Goals

• Changing the underlying MCP SDK's pagination mechanism
• Server-side pagination implementation
• Authentication (passes through to the underlying client)
• Tool/result caching

License

MIT License — © 2026 Prasad Abhishek