pip install mockworld-mcp
mockworld run mock:payments      # a stateful fake Stripe as an MCP server, one command

Status: released — mockworld-mcp on PyPI. Deterministic engine, 6 built-in mocks, MCP stdio + HTTP, fault injection, control plane, registry, world composition, record-mode, snapshots, and a stampede Target. Companion to stampede.

Contents

• Why
• Install
• Quickstart
• Use it in your tests
• Built-in mocks
• Author & share mocks
• Compose, record, and simulate
• Project structure
• Documentation
• Contributing
• License
• Part of the Swarm Proof toolkit

Why

Agents need to do things — charge a card, send an email, place a trade, update a record — but you can't point a half-finished, non-deterministic agent at real Stripe/Gmail/an exchange during development. So teams hand-build throwaway mocks for every project, or test against nothing and find failures in production.

The 2026 crop of agent sandboxes (Veris Sandbox, AWS ToolSimulator) fills that gap by putting an LLM in the response path — convenient, but it means your mock never behaves the same way twice. A stochastic mock can't give you a CI run that's green for the same reason twice, a byte-identical bug repro, or an offline/air-gapped test.

mockworld takes the opposite bet — deterministic, MCP-native, open:

• Deterministic & LLM-free. A seed fully determines state, IDs, timing, and every injected fault. reset --seed 42 produces the same decline, every time, across 50 parallel CI workers. No LLM in the hot path, ever — that's the moat, not a footnote.
• MCP-native & agent-realistic. Services are real MCP servers with agent-grade tool descriptions, stateful behavior, and business-logic fault semantics (declines, insufficient funds, rate limits, disputes) — the things agents actually stress. Postman/WireMock are built for human-driven HTTP testing; they don't speak MCP and can't model business state.
• Open & self-hostable. pip install, runs on your laptop, offline, free, Apache-2.0 — not a hosted SaaS.

Install

pip install mockworld-mcp

Requires Python 3.11+. The distribution is named mockworld-mcp; the import package and CLI are simply mockworld.

Quickstart

mockworld list                        # the built-in mocks
mockworld run mock:payments           # a stateful fake Stripe over stdio (MCP)
mockworld run mock:payments --transport http --port 8931   # Streamable HTTP + control plane
mockworld run mock:payments --seed 42 --faults hostile     # deterministic + adversarial
mockworld inspect mock:crm            # tools, faults, and state shape without running
mockworld demo mock:payments          # prove determinism: same seed → identical transcript

Point any MCP client (or a stampede swarm) at it. mockworld reset --seed 42 returns a running server to a byte-identical world, every time.

Use it in your tests

Installing mockworld gives every pytest run a mockworld fixture — a deterministic fake Stripe in two lines:

def test_agent_handles_a_decline(mockworld):
    pay = mockworld.start("mock:payments", seed=7, faults="hostile")
    cust = pay.call("create_customer", {"name": "Ada", "balance": 10_000}).data
    result = my_agent.charge(pay, cust["id"], 2_500)   # your agent, against a fake Stripe
    assert result.retried_sanely

Seeded and in-memory, so each test is deterministic and isolated — 50 parallel workers never collide.

Built-in mocks

Each enforces real stateful invariants and injects seeded, business-shaped faults. A declarative mock.yaml (plus an optional Python handler) defines a mock in minutes.

Author & share mocks

mockworld new mystripe                # scaffold a runnable, clean-linting mock to grow from
mockworld validate ./mystripe         # schema, handler ABI, determinism smells, description quality
mockworld pack ./mystripe             # print a registry entry (checksum + metadata) to publish

mockworld search weather              # the public registry
mockworld add mock:weather            # install a community mock — checksum-verified + safety-gated

The registry (swarmproof/mockworld-registry) is an index-as-repo: contribute a mock by opening a PR with a folder and an index entry. See docs/AUTHORING.md and mock:hello.

Compose, record, and simulate

# Compose several mocks into one world with a shared customer namespace:
mockworld run world:examples/worlds/ecommerce.yaml --seed 42
#   → payments + crm + email share the same 50 customers: charge → update CRM → email, consistently.

# Scaffold a runnable mock from an OpenAPI spec — or from captured traffic (HAR):
mockworld record --openapi ./petstore.yaml --out ./petstore_mock
mockworld record --har ./session.har --name orders --out ./orders_mock

# Run a scripted-persona swarm → an Agent Readiness Report (the misuse map):
mockworld swarm mock:crm --agents 200 --goal hide --seed 42
#   ⚠ 32.5% of agents destroyed data they meant to hide (delete vs archive) — reproducible.
mockworld swarm mock:crm --agents 200 --seed 42 --descriptions ambiguous
#   ⚠ 45.5% — vaguer tool descriptions, more destroyed data. Legibility is a measurable property.

# Save a dirtied world as a portable artifact; reload it anywhere to reproduce a bug:
mockworld snapshot save mock:payments bug123.mw.json --seed 7

# Govern fidelity drift against a real provider's OpenAPI contract:
mockworld verify mock:payments --against ./stripe-openapi.yaml

# Export target-side traces (OTel GenAI profile) to any OTLP collector:
mockworld run mock:payments --otlp http://localhost:4318

The joint chaos demo — a transport interruption and a business decline at once, with the side-effect firing exactly once — runs standalone:

python examples/demos/exactly_once_under_chaos.py

Project structure

mockworld/
├── src/mockworld/
│   ├── determinism.py       # the seeded entropy funnel (clock/ids/rng/fault-dice)
│   ├── state.py             # copy-on-write state store (memory / sqlite)
│   ├── session.py           # per-session isolation
│   ├── schema.py            # the mock.yaml pydantic models
│   ├── faults.py            # business-logic fault injector
│   ├── dispatch.py          # declarative CRUD + Python handler ABI
│   ├── engine.py            # the transport-free call path (start here)
│   ├── server.py            # MCP exposure: stdio + Streamable HTTP + resources
│   ├── control.py           # control plane + stampede Target protocol
│   ├── trace.py             # OTel-GenAI-profile spans + NDJSON + OTLP export
│   ├── registry.py          # add / search / pack (index-as-repo)
│   ├── world.py             # compose mocks with a shared identity namespace
│   ├── record.py            # scaffold a mock from OpenAPI / HAR
│   ├── snapshot.py          # portable scenario snapshots (+ migration)
│   ├── swarm.py             # persona swarm → Agent Readiness Report
│   ├── verify.py            # contract-drift check vs OpenAPI
│   ├── cli.py               # the mockworld command
│   └── mocks/               # payments · crm · exchange · email · files · hello
├── tests/                   # 86 tests mapping to the TEST-PLAN gates
├── docs/                    # ARCHITECTURE · PRD · AUTHORING · RELEASING · TEST-PLAN · …
├── examples/                # worlds/ · demos/ · registry/
└── .github/workflows/       # ci.yml · release.yml

The engine is deliberately free of any MCP dependency — server.py, control.py, and cli.py are thin adapters over it. That keeps the determinism and isolation tests fast and pure.

Documentation

Contributing

Contributions welcome — bug reports, new mocks, and features. Please read CONTRIBUTING.md. The core principles: determinism is non-negotiable (all entropy comes from the seeded ctx; the validator enforces it), faults are business-logic only, and every mock ships a fidelity.md.

git clone https://github.com/swarmproof/mockworld && cd mockworld
uv venv && uv pip install -e ".[dev]"
python -m pytest -q

License

Apache-2.0. Mocks are LLM-free — deterministic services by design. Citable via CITATION.cff.

Part of the Swarm Proof toolkit

Trust infrastructure for the agent economy — seven projects, one thesis.