FauxCall

Build a faithful fake of any API or MCP server, so you can test your agents without touching the real thing.

Give FauxCall an API's Swagger/OpenAPI page, its documentation, or the address of an MCP server. It calls the real service, one endpoint at a time, and builds a local copy that answers the same way: the same URLs, parameters, responses and errors. Then it runs the same tests against both and keeps fixing the fake until they agree at least 98% of the time.

/fauxcall https://petstore3.swagger.io

Why

Testing agents against real services is painful. The data changes between runs, calls cost money and hit rate limits, and you really don't want a test run to place an actual order. Hand-written mocks avoid that but bring their own problems: they slowly drift away from the real API, their data doesn't hang together, and they ignore time completely.

A FauxCall fake is built from real responses, so it behaves like the original:

• It looks the same. Every URL, parameter, status code, header and error message matches, down to how numbers are written (1.50 stays 1.50). MCP fakes have the same tools, resources and prompts.
• Its data adds up. If a stock references an exchange, that exchange exists. If you upload a file, it shows up in the status endpoint and goes through the same steps, at the same pace, as on the real API.
• It knows what time it is. Asking for the current time gets you the actual time. A stock's timestamp moves during trading hours and stops at the last trade overnight, while the price stays where it closed. Dates come back in exactly the real format.
• It behaves like the real thing under load. Real response times, the real rate limits (with proper 429 errors), pagination, background jobs, file uploads and downloads, live streams (Server-Sent Events and WebSockets), and MCP over HTTP, including sign-in with OAuth.
• It's made for testing. You can freeze or move the clock, see every call your agent made, inject failures, give parallel test runs their own separate state, and reset everything between runs.

Quick start (Claude Code)

You need Claude Code and uv. uv takes care of installing Python.

1. Install the plugin:

   claude plugin marketplace add bradnemetski/fauxcall
   claude plugin install fauxcall@fauxcall
   

   (Inside Claude Code, /plugin marketplace add bradnemetski/fauxcall and /plugin install fauxcall@fauxcall do the same.)

2. Build a fake. In Claude Code, point the skill at a spec, a docs page or an MCP server:

   /fauxcall https://petstore3.swagger.io
   /fauxcall https://api.frankfurter.dev/v1/openapi.json
   /fauxcall https://mcp.deepwiki.com/mcp
   

   Before it starts, Claude asks you everything it needs in one go: any credentials, whether you have a Jev key, and permission for calls that would change data on the real service, along with how many such calls it may make. After that it works on its own. If it gets interrupted, run the same command again and it picks up where it stopped.

3. Run the fake and point your agent at it:

   uvx fauxcall up       # starts every fake in ./fauxcall and prints their addresses
   

   Use the printed address instead of the real API's. For an MCP server, FauxCall also prints the snippet to paste into your agent's configuration.

Using it with other agents

FauxCall isn't tied to Claude Code. Its skills are plain Agent Skills (SKILL.md files), and all the real work happens in the fauxcall command, so any agent that reads skills and can run shell commands can use it: Codex, Gemini CLI, Cursor, GitHub Copilot and others.

1. Install the fauxcall command:

   uv tool install fauxcall                                       # once it's on PyPI
   uv tool install git+https://github.com/bradnemetski/fauxcall   # or straight from GitHub
   

2. Copy the skills to where your agent looks for them. Most agents read the shared ~/.agents/skills/ folder (or .agents/skills/ inside a project); some also have their own, such as ~/.cursor/skills/ or .github/skills/. Check your agent's documentation.

   git clone https://github.com/bradnemetski/fauxcall
   mkdir -p ~/.agents/skills
   cp -R fauxcall/plugins/fauxcall/skills/* ~/.agents/skills/
   

   That installs five skills: fauxcall, refresh, verify, extend and serve. If those last four names clash with skills you already have, copy just fauxcall. It covers a whole build on its own.

3. Ask your agent to use it, for example "use the fauxcall skill to fake https://petstore3.swagger.io".

A few things are smoother in Claude Code: it asks the up-front questions as a form, and runs the final independent review as a separate subagent. Other agents do the same steps in the conversation.

What you get

Each fake is a small, ordinary FastAPI project in ./fauxcall/<name>/:

your-project/
├── fauxcall/
│   ├── petstore/
│   │   ├── data/model.json          # everything the fake knows, in one readable file
│   │   ├── data/recordings/         # the real requests and responses it was built from
│   │   ├── data/world/              # the data it serves
│   │   ├── src/fauxcall_petstore/   # its routes, plus any custom code
│   │   ├── tests/                   # its tests
│   │   └── Dockerfile
│   ├── compose.yaml                 # runs all your fakes in containers
│   └── pyproject.toml               # one uv workspace for all of them
└── .github/workflows/fauxcall.yml   # CI that keeps the fakes tested

Run a fake's tests offline, or against the real service to see whether the real API has changed:

cd fauxcall/petstore
uv run pytest                                 # offline tests
uv run pytest tests/e2e --target=parity       # same tests against both; fails below 98%

The examples/ folder has fakes FauxCall built by itself, from a single command each, while we were testing it: Frankfurter, timeapi.io, Swagger Petstore, FakeRESTApi and this repository's own reference API.

Using fakes in tests

Every fake has a small control API under /__fauxcall:

To run tests in parallel, give each run its own name in an X-Fauxcall-Session header. It then gets its own data and its own clock. Set FAUXCALL_FAST=1 to skip the realistic delays and rate limits.

Skills

Privacy and safety

• Your credentials stay private. They live only in each fake's .env file, which you fill in and which git ignores. They're scrubbed from recordings, logs and the call journal, and never copied into Docker images.
• Personal data is replaced with realistic stand-ins as it's recorded, consistently, so a customer's name is replaced the same way in every response.
• Nothing that changes data happens without your OK. Before recording, you see the calls that would create, change or delete things, and approve them once. You also set how many such calls each endpoint may make (100 unless you choose otherwise; set FAUXCALL_APPROVAL_BUDGET to change the default).
• It's polite to the real service. Recording is slowed down to stay within rate limits, and capped.

Jev (optional)

If you have a TypeSafe Jev key, a fake can use it to make judgment calls its recordings don't cover, for example picking which recorded answer best fits a question it hasn't seen before. Its answers are saved with the fake, so they replay later without the key. Without a key, the fake uses simple, predictable rules instead.

The fauxcall command

The skills drive the fauxcall command-line tool, and you can use it directly too:

uvx fauxcall status                 # how far each fake's build has got
uvx fauxcall serve petstore         # run one fake in the foreground
uvx fauxcall verify petstore --target parity
uvx fauxcall --help                 # everything else

Questions

Why does it install httpx2 and httpcore2 instead of httpx and httpcore? They aren't lookalike copies. They're maintained by Pydantic (github.com/pydantic/httpx2), by the author of httpx. FauxCall needs httpx2 for WebSockets and for testing fakes in-process; httpcore2 is the lower-level library httpx2 is built on.

Learn more

• Examples: fakes FauxCall built by itself, ready to run
• How it was tested: the method, the services we tried and the results
• How it works: the pipeline, the model and the runtime
• Configuration: every setting
• Design decisions: what we decided and why
• Contributing and security

License

Apache-2.0