Mock Anything
A working twin of any API, running on your machine in one command. No keys, no sandbox sign-up, no bill.
Point your app or agent at that URL instead of the real one and it behaves like the real API for the flows it covers: create a customer, charge a test card, get the signed webhook, refund it. None of it is real, and nothing leaves your machine.
Mock Anything works with any API, not just a fixed list:
Why
Agents write code that calls payments, SMS, email and developer APIs all day. Testing that today means real keys (the agent can spend money or text real people), sandboxes (accounts and keys for every provider, and many regional providers barely have one), or hand-written mocks that break the moment the agent does something you didn't expect.
A twin is different from a mock:
- It keeps state. POST a customer, then GET it back. Refund a payment and the remaining balance changes. Delete something and it's gone.
- It behaves like the provider. Real ID formats, list pagination, error shapes, status lifecycles (
requires_payment_method → succeeded, queued → sent → delivered).
- It sends webhooks, signed the way the provider signs them, so your verification code runs unchanged.
- It does the awkward parts: declined cards, 3-D Secure, bounced emails, undeliverable SMS, duplicate references, rate limits.
- Chaos mode (
--chaos) adds timeouts, 5xx errors, 429s, slow responses, duplicate and out-of-order webhooks. Seeded, so a failure you hit can be replayed.
Built-in packs
Hand-tuned twins, checked against each provider's official SDK in our test suite.
mockanything env <pack> prints the snippet for pointing the official SDK at the twin, for example:
Any other API
Give Mock Anything an OpenAPI 3.x or Swagger 2.0 description (a file, a URL, or just the API's name) and it builds a stateful twin by reading the spec:
/things + /things/{id} become a collection you can create, list, read, update and delete;/things/{id}/cancel-style paths become state changes, matched to the object's status values;- nested paths are scoped (issues in one repo don't show up in another);
- response envelopes, pagination style, error shape, auth and OAS 3.1 webhooks are read from the spec;
- paths the spec doesn't describe still work as generic REST (marked with
x-mockanything-fallback);
- read-only resources (nothing in the API creates them) return made-up sample objects shaped by the spec, so lookups and lists aren't empty.
See what was inferred, and change it:
How many APIs work?
Measured on 2026-10-29 with scripts/coverage.ts, which builds a twin from each spec, creates objects, then calls every operation (up to 40 per API):
The calls that don't succeed get proper errors, mostly "not found" for an id that was never created, or a missing parent object. Starting and answering is not the same as matching the real API; verify measures that.
Popular APIs whose providers publish a current spec (Square, PayPal, Adyen, Discord, Jira, Cloudflare, Vercel, Supabase, Clerk, Hugging Face and more) are fetched from the provider's own published source rather than the directory, which is often out of date. Run mockanything list for all of them.
Twins built straight from the public Stripe, Twilio, GitHub and Resend specs, with no hand-tuning, pass the official SDKs' basic create/read/update/list/delete flows (scripts/conformance.ts).
From a recording
Record your app talking to the real API (browser DevTools → Network → "Save all as HAR", or any proxy that writes HAR) and start a twin from the file:
Mock Anything rebuilds the endpoints, id formats, list shapes and status lifecycles it saw. It keeps field names and shapes only: header values, cookies, query values and free text from the recording are never copied into the twin, and emails, URLs and dates are replaced with made-up values.
No spec at all? Let your agent build one: the MCP tool twin_from_docs reads a documentation page and gives the agent a short format guide; the agent writes a twin definition and starts it with twin_start.
For agents
As a skill (no setup beyond the file)
Copy skills/mock-anything into ~/.claude/skills/ (or your agent's skills folder). Then tell the agent "mock Stripe", "test this against a fake Notion", or point it at a spec, a HAR file or a docs page. The skill uses only the command line: it starts a twin, points your app at it, and drives it:
For an API with only a docs page, npx @shipatmidnight/mockanything docs <url> gives the agent the page text and the twin format; the agent writes twin.json, checks it with npx @shipatmidnight/mockanything validate twin.json, and starts it.
MCP
Cursor, Windsurf and other MCP clients:
Tools: twin_start, twin_from_docs, twin_env, twin_events, twin_state, twin_scenario, twin_chaos, twin_webhook_add, twin_emit, twin_advance_clock, twin_reset, twin_stop, twin_list, packs_list.
An agent can say "start a Stripe twin, build the checkout, then force a declined card and check my code handles it" without asking you for keys.
Keep data between restarts
Objects, id counters, the test clock, the webhook secret and registered webhook endpoints are saved shortly after each change and when the twin stops. The file is readable by you only. Delete it (or press Reset in the dashboard) to start fresh. Agents get the same with twin_start and persist: true.
Dashboard
Every twin has a local dashboard (the link with its access token is printed at start): live requests and responses, webhooks sent and their delivery status, the twin's stored state, scenarios to force, and chaos controls.
Command line
Safe by default
- Listens on
localhost only. Delivers webhooks to localhost only unless you allow a host.
- Refuses live keys (
sk_live_… and similar) and real-looking card numbers. Use the provider's documented test values.
- Test checkout pages are plain, carry a "not a real payment" banner, and only exist on your machine.
- No telemetry. Nothing is sent anywhere except the webhooks you configure.
How faithful is a twin?
A twin is only as good as the facts it was built from. Packs cover the main flow of each API, not every endpoint, and inferred twins make guesses that you can inspect and override.
verify measures it. It runs the same scripted calls against the twin and against the provider's real sandbox, using your own test-mode key, and compares status codes, field names, field types, id formats and error codes:
It only runs with test-mode keys (live keys are refused), makes about one call a second with a hard cap per run, and creates a few test objects in your sandbox (the Stripe suite deletes its customer afterwards). The report lists field paths, types and code values only, never response data or your key. Built-in suites: stripe, paystack; write your own for any twin (see packs/verify/). Webhooks are not compared yet.
mockanything list shows each pack's published score and the date it was measured. Use twins to build and test quickly; run your final checks against the provider's sandbox before you ship.
Twin definitions
Every twin, built-in or inferred, is a plain YAML/JSON definition run by one engine. Definitions are data: they cannot run code. See packs/ for examples and src/mcp/guide.ts for the format.
Contributing
Packs for more APIs are very welcome, especially regional ones. Read CONTRIBUTING.md first: packs contain facts only (paths, fields, error codes, documented test values), never copied documentation text or logos.
Community
Questions, ideas and show-and-tell: the Ship at Midnight Discord, one server for every tool, with a channel for Mock Anything.
Licence and notices
Apache-2.0. See LICENSE and NOTICE.
Mock Anything is a free, non-commercial passion project by an independent developer. It is not affiliated with, endorsed by, or sponsored by any company whose API it simulates. Stripe, Paystack, Twilio, Resend, GitHub, OpenAI, Anthropic, Claude and other names are trademarks of their respective owners and are used only to describe compatibility.
