接入你的智能体
發布者自己的指令。不是 PlugMCP 端點。
用我們的金鑰與日誌託管一份副本。
npx -y @mockzilla/[email protected]Claude Code 沒有本機指令的安裝連結。複製下方指令並在終端機中執行。
MOCKZILLA_TOKEN=<secret> MOCKZILLA_MCP_URL=<value> MOCKZILLA_MANAGED_PORT=<value> claude mcp add io.github.mockzilla-mockzilla-mcp -- npx -y @mockzilla/mcp@0.2.28在終端機中執行。
{ "mcpServers": { "io.github.mockzilla-mockzilla-mcp": { "command": "npx", "args": [ "-y", "@mockzilla/[email protected]" ], "env": { "MOCKZILLA_TOKEN": "<secret>", "MOCKZILLA_MCP_URL": "<value>", "MOCKZILLA_MANAGED_PORT": "<value>" } } }}claude_desktop_config.json
遠端 URL 還是本機指令？ 兩種安裝方式的差異
來自發布者
MCP server for Mockzilla - an open-source API mock server for OpenAPI specifications. Let Claude Code, Claude Desktop, Cursor, or Gemini CLI install Mockzilla, inspect OpenAPI specs, and spin up realistic local mock APIs in seconds. No account required for local use.
Mock OpenAPI specs locally, or log in to deploy hosted mocks on mockzilla.org from your MCP client.
@mockzilla/mcp exposes two planes of tools to your MCP client.
Runs on your machine. No Mockzilla account needed.
From an agent you can:
.mockz archive.Ask your agent to log in, or ask for something hosted. The agent calls the login tool, your browser opens the Mockzilla login, and you pick an organization and read-only or read-and-write access. The bridge keeps the login on your machine and renews it by itself. Hosted tools are listed from the start; before you log in, the agent is told to log in first.
Agents can then:
Before logging in, only the local plane is exposed. Agents can still help users explore Mockzilla and run local mocks before they sign up.
You can use these directly from Claude Code, Claude Desktop, Cursor, or Gemini CLI once mockzilla is configured as an MCP server.
https://example.com/openapi.yaml expose?"POST /checkout to return a 402 response."openapi.yml into ./api."stripe-test and give me the live URL."One-liner, no config file editing:
claude mcp add -s user mockzilla -- npx -y @mockzilla/mcp@latest
-s user installs for your user account (available in every project).-s user to scope to the current project only.Edit ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"mockzilla": {
"command": "npx",
"args": ["-y", "@mockzilla/mcp@latest"]
}
}
}
Restart Claude Desktop after editing.
Easiest: Settings -> MCP Servers -> Add new MCP server and fill in:
mockzillanpx-y @mockzilla/mcp@latestOr edit ~/.cursor/mcp.json directly:
{
"mcpServers": {
"mockzilla": {
"command": "npx",
"args": ["-y", "@mockzilla/mcp@latest"]
}
}
}
Restart Cursor after editing.
One-liner, no manual JSON editing:
gemini mcp add -s user mockzilla npx -y @mockzilla/mcp@latest
-s user writes to ~/.gemini/settings.json (available in every project).-s user (or use -s project) to scope to the current directory's .gemini/settings.json.Or edit the settings file directly:
{
"mcpServers": {
"mockzilla": {
"command": "npx",
"args": ["-y", "@mockzilla/mcp@latest"]
}
}
}
Restart the Gemini CLI after editing.
@latest?Without @latest, npx caches the first resolved version and won't pick up new publishes. Pinning to @latest makes npx re-check the registry on every spawn, so a Claude / Cursor / Gemini restart is enough to upgrade. Trade-off: ~200 ms extra startup time.
These tools are always available and never leave the user's machine.
check_cli
Resolve Mockzilla on this machine: system PATH -> bridge cache -> go run invocation. Returns install options if nothing matches.
install_cli
Install Mockzilla into ~/.cache/mockzilla-mcp/. Methods: download (prebuilt from GitHub releases, default), go-install, go-run. Never touches system PATH.
bridge_status
Report the bridge's version, check npm for newer publishes, and surface upgrade steps.
info
Summarise an OpenAPI spec without serving it: {title, version, openapi_version, endpoint_count, paths}. Also reads .mockz packages.
lint
Find schemas in a spec that no value can satisfy, so a broken spec is caught before serving it: {clean, defect_count, defects}.
discover_specs
Scan a directory for OpenAPI specs and folders of static endpoint files. Returns a suggested_input for serve_locally.
mockzilla_docs_topics
List the Mockzilla docs, by category, with each topic's title and summary. The product docs from mockzilla.org, the engine docs and the mockzilla-codegen docs all ship inside the package, so no network or login is needed.
mockzilla_docs_read
Return the full markdown for one or more topics, or a whole category.
mockzilla_docs_search
Keyword search across all docs; returns top sections with snippets.
simplify
Drop or reduce union types, strip x-* extensions, and optionally cap optional properties per schema. Writes the simplified spec to disk and returns its path.
pack
Pack a directory of services into a .mockz archive that serve_locally can serve, even from a URL.
generate
Run mockzilla generate, the built-in mockzilla-codegen: Go models from a spec, plus a server, a client or MCP tools when asked. Reads codegen.yaml when the folder has one. Writes the files, lists them first with dry_run, or reports stale ones with check, along with the spec's errors and warnings. Needs CLI 2.11.0 or newer.serve_locally
Start a portable mock server on a free port. Accepts a spec file, directory, or public https URL. Returns {url, port, pid, services}. For a single spec, latency, errors, mount and context add delay, inject error responses, change the mount path, or set replacement values.
stop_locally
Stop a server started by serve_locally.
call_endpoint
Make an HTTP request and return {status, headers, body}, to show a mock's response. Localhost only unless allow_remote is set.
mock_endpoint
Quickly mock a single HTTP endpoint without an OpenAPI spec. Writes a static response into the managed mocks dir and (re)starts the shared server. Takes status and headers for a failure or a redirect with a real body (404 with an error payload, 201 with a Location); omit response for a body-less 204 or 304. Needs mockzilla 2.8.20 or newer.
list_mock_endpoints
List all endpoints currently mocked, plus the running server's URL and the Mockzilla UI URL.
clear_mock_endpoints
Wipe all mocks and stop the managed server.
request_history
List the requests the running server answered: method, URL, status, content type, latency, and whether the response came from an upstream. Pass id with service for one request's full headers and body. Reads the server's own history API, so no account is needed.
diagnose_requests
Explain what is wrong with the recorded traffic and where its data came from: a breakdown by origin, status and content type, latency p50/p95/max, and findings such as an upstream that failed and silently fell back to a generated mock, a 404 from a wrong mount prefix, or a JSON body under a non-JSON content type.
setup_replay
Configure replay for a service: record a response once, then serve it back for every matching request. Returns recording_scope saying what each endpoint is keyed by, since with no match fields every call to an endpoint shares one recording. Writes config.yml only inside the bridge's own mocks dir; for a folder from your own project it returns the YAML and where it goes.
list_replays
List a service's recordings, with the request values each one is keyed by.
check_github_deployable
Say whether a repository or folder would deploy a mock, and what is missing if not. Knows both kinds: portable service folders and a codegen Go server. Read-only.
list_github_repos
The user's repositories, flagging which already publish mocks, so the agent can ask which one to use.
publish_to_github
Push mocks to one of the user's own repositories and let the Mockzilla action deploy them at a host of their own, https://<label>.api.mockz.io. The label is the repository name, with -2, -3 if it is taken. No Mockzilla account needed: the first push registers the repository. Merges into an existing services folder rather than replacing it, and never overwrites an existing workflow.
wait_for_github_deploy
Wait for the workflow run and return the live URL, as the action reported it.
unpublish_from_github
Take the mocks down and free the simulation slot, optionally deleting the repository too.
login
Opens the Mockzilla login in the browser, where the user picks an organization and read-only or read-and-write access. Returns right away with the login link. The agent calls a hosted tool again once the user approves. The login is saved under ~/.config/mockzilla-mcp/, one per server URL, and renewed automatically.
logout
Revokes the connection and deletes the saved login. Log out and in again to switch organization or access.
@mockzilla/mcp lists the hosted tools from the start and forwards them to mockzilla.org's MCP endpoint once you log in. At the time of writing, the hosted surface includes:
get_contextlist_simslist_catalog_productsdeploy_mock_from_catalogdeploy_mock_from_specdeploy_mock_from_urlwait_for_deployRefer to the hosted server's docs or the MCP registry entry for the live tool list.
On a machine without a browser, such as CI or a remote server, set MOCKZILLA_TOKEN to an API key from the dashboard instead of logging in. The hosted tools are then available from the start.
The bridge keeps the CLI and mocks under ~/.cache/mockzilla-mcp/, and the login under ~/.config/mockzilla-mcp/:
~/.cache/mockzilla-mcp/
├── bin/mockzilla # downloaded or go-installed binary
├── config.json # { method, version, invocation? }
└── mocks/ # mock_endpoint persists static endpoints here
└── services/
└── <first path segment>/<rest of path>/<method>/
├── index.<ext> # the response body
└── meta.json # status and headers, written only when set
~/.config/mockzilla-mcp/
└── credentials.json # saved logins, one per server URL, readable only by you
rm -rf ~/.cache/mockzilla-mcp resets the CLI and all mocked endpoints. The login stays.rm -rf ~/.cache/mockzilla-mcp/mocks.PATH is never touched, so reset doesn't affect a separate brew install of Mockzilla.Recommended way to stay current:
@mockzilla/mcp@latest in your MCP client config so npx re-checks the registry on every spawn.bridge_status and tell me if @mockzilla/mcp is up to date."If it's stale, run:
npx clear-npx-cache @mockzilla/mcp
and restart your MCP client.
The Mockzilla CLI version is pinned by the bridge (via MOCKZILLA_VERSION in lib/install.js). Updating the bridge updates the pin; the next install_cli call brings the CLI itself up to date.
See CLAUDE.md for project conventions and a walkthrough of adding a new tool.
The bridge has two registries to keep in sync: npm (@mockzilla/mcp) and the MCP registry (server.json). Skipping the second one leaves discovery clients pinned to the previous tarball.
The usual way is to bump version in package.json, merge, and publish a GitHub release tagged v<version>. That covers both registries: see From GitHub Actions below.
To release by hand instead:
Bump version in package.json.
Run:
make publish-all
This will:
docs/ and hosted-tools.json from the published docs bundle, plus the engine docs at the pinned CLI version. Maintainers only: it needs DOCS_BUNDLE_CMD in local.mk.mock_endpoint against the real CLI, and the behavior checks that each tool reports only what the server really does (CLI-dependent ones skip when no CLI is installed).npm publish the new tarball.server.json.mcp-publisher in with the GitHub token from Keychain (see below).mcp-publisher publish against the MCP registry.Commit the server.json bump.
If you only want one side:
make publish for npm only.make publish-mcp for the MCP registry only (server.json is always re-synced from package.json first).mcp-publisher must be on PATH (brew install mcp-publisher or follow the installation docs).
A registry login lasts 5 minutes, so make publish-mcp logs in before every publish. It reads a GitHub token from macOS Keychain, because the browser login (mcp-publisher login github without a token) can't publish under io.github.mockzilla.
One-time setup:
Create a classic GitHub token with read:org and read:user.
Store it in Keychain. The command prompts for the token:
security add-generic-password -a "$USER" -s mcp-publisher-github -w
Publishing a GitHub release whose tag matches package.json runs .github/workflows/publish.yml, which builds the docs, runs the smoke tests and publishes to npm with provenance. It skips a version npm already has, so it is safe to run after make publish.
Two things have to be set up for it, both one-time:
mockzilla/mockzilla-mcp, workflow publish.yml. No token is stored anywhere.AWS_ROLE_ARN repository secret, set to the github_mcp_release_role_arn output of infra's live/aws/iam-github-oidc. The docs bundle is only in S3 - nothing serves it, and the CDN in front of that bucket caches for a day - so the release assumes a role that can read that one key..github/workflows/publish-mcp.yml starts on the same release and covers the other registry. It skips versions the MCP registry already has, waits up to 10 minutes for the version to appear on npm, then publishes server.json using GitHub OIDC, so no token is needed. It can still be started by hand from the Actions tab.
The two run in parallel rather than in sequence. The wait loop is what keeps that safe, and it is needed anyway: npm accepts a tarball before it can be read back.
mockzilla generateCopyright © 2026-present
Licensed under the MIT License
Mock OpenAPI specs locally, or log in to deploy hosted mocks on mockzilla.org from your MCP client.
{ "mcpServers": { "io.github.mockzilla-mockzilla-mcp": { "command": "npx", "args": [ "-y", "@mockzilla/[email protected]" ], "env": { "MOCKZILLA_TOKEN": "<secret>", "MOCKZILLA_MCP_URL": "<value>", "MOCKZILLA_MANAGED_PORT": "<value>" } } }}工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
發布者列出了以下環境變數：MOCKZILLA_TOKEN, MOCKZILLA_MCP_URL, MOCKZILLA_MANAGED_PORT。
Mockzilla 透過 stdio 在本機執行。
上次檢查時間：2026-11-09。 Mock OpenAPI specs locally, or log in to deploy hosted mocks on mockzilla.org from your MCP client.