Build an MCP from a website

building-website-mcps helps an agent turn a website, web app, public API, or local CLI into an MCP server. Give it a service you are allowed to use, and it works through the boring but important parts: how to authenticate, which routes and actions exist, what is safe to expose, and which actions still need a human in the loop.

It does not treat a scraped endpoint as ready for production. The skill maps what it finds, tests it, and only exposes an action after there is real proof it works.

Install

Install it for Codex:

npx building-website-mcps install --target codex

Or install it for Claude Code:

npx building-website-mcps install --target claude

The installer puts the skill in your personal skills folder:

| Tool | Location | Use it as | | --- | --- | --- | | Codex | ~/.codex/skills/building-website-mcps | $building-website-mcps | | Claude Code | ~/.claude/skills/building-website-mcps | /building-website-mcps |

It will not replace an existing install unless you add --force. To use a different directory, add --dest <skills-directory>.

Claude Code uses the same standard SKILL.md layout for personal skills. See the Claude Code skills docs for its skill discovery rules.

What to ask for

Try one of these:

Use $building-website-mcps to build a safe MCP for this authorized service.

Map this public API, then give me a small MCP for the actions that are actually proven.

Turn this local CLI into an MCP. Do not use a shell or make up commands.

The skill keeps a record of its work, so a later agent can pick up where the last one stopped instead of starting over.

What you get

For an authorized target, the skill can produce:

• An auth plan that keeps credentials out of saved artifacts.
• An OpenAPI map for HTTP routes and an action graph for UI and CLI work.
• A compact MCP with tools for finding, understanding, planning, and running supported actions.
• A companion skill that explains setup and safe use to the next agent.
• Tests for normal use, error cases, full workflows, and response size and speed.

It keeps UI-only actions separate from stable HTTP or CLI operations. It also asks for confirmation before writes and keeps actions hidden until they pass fresh contract or end-to-end tests.

What it has been tested against

The repository includes three runnable local apps:

| Example | Covers | | --- | --- | | Public catalog | Anonymous reads and document downloads | | Session admin | Session recovery, CRUD, batch work, imports, exports, and confirmed writes | | Hybrid CLI | HTTP calls alongside typed local CLI commands |

These are real local HTTP and CLI processes. They are not mocked network calls.

Run the checks yourself

You only need Python 3.10 or newer:

WEBSITE_MCP_APPROVAL_KEY=local-fixture-key \
  PYTHONPATH=building-website-mcps/tests \
  python3 building-website-mcps/scripts/run_fixture_matrix.py \
  --output /tmp/website-mcp-matrix.json

python3 building-website-mcps/scripts/benchmark_mcp.py \
  --matrix /tmp/website-mcp-matrix.json \
  --output /tmp/website-mcp-benchmark.json --iterations 5

PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover \
  -s building-website-mcps/tests -p 'test_*.py' -q

There is also a cold-agent trace in building-website-mcps/evals/traces/. It comes from a real collaboration-agent run, but it is not independent external agent evidence and its timing and response-size fields were not instrumented.

Safety

Use this with public surfaces or accounts and data you are allowed to access. It does not bypass logins, CAPTCHAs, rate limits, or anti-bot protections. Saved artifacts use secret references instead of raw passwords, cookies, or tokens.

For the full build contract, see building-website-mcps/SKILL.md.