- Home
- Skills
- APIs & Backend
- MCP Server Kit
More screenshots
Works with the AI Agents you already use
MCP Server Kit
Build a working, tested MCP server in Python or TypeScript that connects Claude to your own API, database or tools.
$19
MCP Server Kit
Example session with this skill installed
We have an internal CRM REST API for contacts. Build an MCP server so Claude Code can search contacts and read one in full. Python is fine.
- Read your context and instructions
- Compiled the mcp server kit
CRM contacts MCP server: built and tested
I started from the kit's REST API wrapper and pointed it at your CRM.
Tools
| Tool | Reads/writes | Returns |
|---|---|---|
search_contacts(query, page=1) | read | 20 summaries per page (id, name, company, email), total, has_more |
get_contact(contact_id) | read | one contact with phone, notes and last update; 40+ upstream fields trimmed away |
Ids must match ^[A-Za-z0-9_-]{1,64}$ and are URL-quoted, so model input never reaches the upstream path raw. Upstream 401, 404, 429 and 5xx come back as tool errors that tell Claude what to do next ("Wait 30 seconds", "Use search_contacts to get a valid id").
Tests
$ pytest -q test_rest_api_server.py
........ [100%]
8 passed in 0.59s
$ python mcp_smoke.py --stdio "python rest_api_server.py" --call search_contacts '{"query":"acme"}'
PASS stdout hygiene stdout carries only JSON-RPC
PASS handshake contacts 1.0.0, protocol 2025-11-25, 684 ms
PASS tools 2 tool(s) listed: get_contact, search_contacts
PASS resources 0 resource(s), 0 template(s)
PASS prompts 0 prompt(s)
PASS call search_contacts isError=False +structuredContent: { "total": 23, "page": 1, "has_more": true, ... }
6 pass, 0 warn, 0 fail
The tests run against a mocked upstream (success, pagination, 401, 404, 429, 503, timeout, a path-traversal attempt), so they need no network and no real key.
Register it
claude mcp add crm -e API_BASE_URL=https://crm.internal/v1 -e API_KEY=... -- "$PWD/.venv/bin/python" "$PWD/rest_api_server.py"
Then /mcp in Claude Code shows crm connected with 2 tools.
Next steps
- Keep
API_KEYout of the repo:.env(git-ignored) or the client config'senv. - To share it with the team over HTTP: set
MCP_AUTH_TOKENandMCP_ALLOWED_HOSTS, then run the security checklist.
Connects securely to your tools. The creator never sees your data.
About this skill
What it does
MCP Server Kit gets you from "I want Claude to talk to our system" to a server that works the first time it is plugged into a client, and keeps working once people use it.
It ships two templates that pass their own protocol tests over stdio and Streamable HTTP: one on FastMCP (Python), one on the official TypeScript SDK. Both show every pattern in one small file: tools with structured output, tool errors the model can recover from, a resource, a prompt, bearer auth compared in constant time, DNS rebinding protection, and a refusal to listen publicly without a token.
The smoke tester
mcp_smoke.py tests any MCP server, in any language, the way a client sees it:
- stdout pollution, at startup and during tool calls (the classic stray
print()that breaks every client) - schemas and parameter names the Claude API rejects, root-level
anyOf/oneOf, tool names over the 128-character limit - missing descriptions, undescribed parameters, missing annotations, missing server instructions
- HTTP 401, 403, 404, 405 and 421 translated into the actual cause
- CI-friendly exit codes and
--json
Also inside
scaffold.py: a new project from a template in one command- A REST API wrapper example with shared client, timeouts, upstream errors mapped to recovery hints, trimmed output and safe ids, plus 8 tests against a mocked upstream
- References: tool design, transports and auth (stateless vs stateful, bearer vs OAuth 2.1), exact client configs for Claude Code, Claude Desktop, Cursor and VS Code, troubleshooting tables, and a security checklist
Tested with mcp 1.27 and @modelcontextprotocol/sdk 1.32.
Built by Nex AI, nex-ai.be
The templates run HTTP stateless; server-initiated messages across requests (progress on long jobs, subscriptions) need a stateful setup, which is described but not templated. Uses a bearer token, not OAuth; the references explain when and how to move to OAuth 2.1. Client config formats change; the reference reflects them as of the version date.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 29 days ago
Requires
Permissions
Terminal to run the scripts and tests. Read/write files to scaffold a project. Network only for installing packages (pip, npm) and for HTTP calls to the MCP server URL you test. Environment variables for the server token and API keys you configure.