See exactly how an MCP server works in under 100 lines of Python, then ship your own tools on it today.
The smallest complete MCP server we could write in plain Python: one file of protocol code, one file of tools, no SDK, no pip install. Copy the folder, rename it, replace the two example tools with yours, and connect it to Claude Code, Cursor or Claude Desktop. Includes mcp_probe.py, a tiny MCP client that tests any stdio server (yours or someone else's) the way a real client does.
The screenshot is real output from the kit’s own commands. Tested from a fresh download of this exact zip: ./run_tests.sh ran 11 tests OK on Python 3.13 and 3.10.
| File | What it is |
|---|---|
| server.py | entry point: name, version, instructions, tool table |
| myserver/tools.py | your tools: two examples (text_stats, convert_units) showing return values, input validation and ToolError |
| myserver/mcpcore.py | the protocol: JSON-RPC over stdio, server/discover, initialize, tools/list, tools/call, errors (under 100 lines) |
| mcp_probe.py | client: tries MCP 2026-07-28 server/discover, falls back to the initialize handshake, lists tools, calls one |
| tests/ | 11 tests: both protocol eras, error codes, your tools, the probe against a legacy-only server |
| .claude-plugin/, .mcp.json | so the folder also installs as a Claude Code plugin |
It speaks MCP 2026-07-28 (stateless: every request carries its version in _meta, server/discover, resultType and server info on results, cache hints on lists) and the handshake revisions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05 on the same process, as the spec's "dual-era" server. Newer clients (for example Claude Code's v2 MCP runtime) use the stateless path; older ones keep working.
1. Write a function in myserver/tools.py that takes keyword arguments and returns a dict. 2. Add it to TOOLS: "name": (function, "what it does and when to use it", {json schema properties}, ["required"], {"readOnlyHint": True}). 3. Add a test in tests/test_template.py, run ./run_tests.sh, then python3 mcp_probe.py --call name '{"arg": 1}' -- python3 server.py.
Raise ToolError("message") for problems the model should see and fix (bad input, not found). Anything else becomes a JSON-RPC internal error and is logged to stderr. Never print to stdout: stdout is the protocol channel.
claude mcp add my-server -- python3 /absolute/path/to/server.py # Claude CodeCursor: examples/configs/cursor-mcp.json into .cursor/mcp.json. Claude Desktop: merge examples/configs/claude_desktop_config.json.
Want a real server built on this core? See the GitHub Issue Triage MCP Server. More at https://youraitools.co/marketplace. License: LICENSE.txt (free, keep the notice).
Zip, 13 KB, v1.0, 15 files. Free to use, modify and share for any purpose, including commercial work, as long as this notice stays in copies. Not affiliated with or endorsed by Anthropic or the Model Context Protocol project.
Free Python MCP server template with no dependencies: stdio server for MCP 2026-07-28 and 2025 clients, two example tools, tests and a probe client.
10 Claude Code subagents for DevOps and SRE: incident triage, Terraform plan review, Kubernetes, CI fixes, alerts, postmortems. Guard hook blocks destroy. $19.
View DevOps & SRE Subagents for Claude Code10 Claude Code subagents for data teams: SQL and dbt review, query tuning, data quality, metrics, A/B tests. SQL guard blocks DROP and unfiltered DELETE. $19.
View Data & Analytics Subagents for Claude Code30 Claude Code subagents: code review, security, tests, DevOps/SRE, Terraform, Kubernetes, SQL, dbt, data quality. Guard hooks included. $39 ($57 separately).
View Claude Code Subagents Bundle