Agent VCR

🚀 Now available on PyPI and npm! Install with pip install agent-vcr or npm install @agent-vcr/core.

Record, replay, and diff MCP interactions — like VCR for AI agents.

Test MCP servers and clients without flaky live servers. Mock MCP for testing: record once, replay forever. No more "MCP server was down" or "rate limit" in CI — deterministic, fast, offline.

Agent VCR is a testing framework for the Model Context Protocol (MCP). It transparently records all JSON-RPC 2.0 interactions between an MCP client and server, then replays them deterministically — no real server needed. Easier than rolling your own mocks: one install, one command to record, one to replay. Golden cassettes in seconds.

Python and TypeScript are first-class. The Python implementation has 250+ tests and a full CLI; the TypeScript implementation has 72 unit tests, full CLI, and ships as @agent-vcr/core on npm — ideal for the TypeScript-first MCP ecosystem, where most MCP servers and clients live. Recordings are cross-language: record with Python, replay with TypeScript (or the other way around). See typescript/README.md.

The Problem

If you're building MCP servers or clients, you've hit these walls:

"My tests are flaky because they depend on a live server." External MCP servers go down, rate-limit, or return different results each run. Your CI pipeline fails for reasons that have nothing to do with your code.

"I can't test error handling without breaking the server." How do you verify your client handles a timeout, a malformed response, or a server crash? You'd need to modify the server itself — or just hope for the best.

"I shipped a breaking change and didn't catch it." You updated your MCP server and a downstream client broke. There was no way to detect that tools/call started returning a different schema until a user filed a bug.

"Testing against real APIs is slow and expensive." Each test run hits the real server, waits for real responses, and burns through API quotas. A test suite that should take seconds takes minutes.

How Agent VCR Solves This

Record your MCP interactions once against the real server, save them as .vcr cassettes, and replay them forever:

                Record (once)              Replay (every test run)
                ─────────────              ─────────────────────────
Client ←→ Agent VCR ←→ Real Server    Client ←→ Agent VCR (mock)
                │                                    │
                └──→ session.vcr ────────────────────┘

• Deterministic: Same input, same output, every time
• Fast: No network calls, instant responses
• Offline: Tests run without server access
• Safe: Inject errors without modifying the real server
• Visible: Diff two recordings to catch regressions before they ship

Result: CI that used to be slow and flaky (live MCP server, timeouts, rate limits) becomes fast and deterministic — run in seconds, catch breaking changes before deploy.

Real-World Use Cases

Open source MCP server authors: Ship a .vcr cassette with your server so users can run their client tests without installing or running your server — a distribution story nobody else enables. Enterprises with multi-agent systems: Many AI agents talking to many MCP servers? When server team A pushes a new version, the diff feature catches breaking changes before production. Platform teams use Agent VCR to gate deploys. CI/CD: No more flaky tests because an MCP server was down, rate-limited, or slow. Record golden cassettes; tests run in milliseconds, deterministic every time. Cost control: MCP servers that call paid APIs? Record once — never burn quota again in tests.

Real-World Examples

1. Golden Cassette Testing

Record a "known good" session, commit the .vcr file to your repo, and replay it in CI. If your code changes break the interaction pattern, the test fails immediately.

# Record the golden cassette (once, using the included demo server)
agent-vcr record --transport stdio --server-command "python demo/servers/calculator_v1.py" -o cassettes/golden.vcr

# Every CI run replays it
pytest tests/ --vcr-dir=cassettes

2. MCP Server Compatibility Gates

Before deploying a new server version, record both old and new, then diff:

agent-vcr record --transport stdio --server-command "python demo/servers/calculator_v1.py" -o v1.vcr
agent-vcr record --transport stdio --server-command "python demo/servers/calculator_v2.py" -o v2.vcr
agent-vcr diff v1.vcr v2.vcr --fail-on-breaking

If tools/call changed its response schema, or a method was removed, the diff catches it and exits with code 1 — blocking the deploy.

3. Error Injection for Resilience Testing

Use response overrides to simulate failures without modifying the server:

replayer = MCPReplayer(recording)

# Inject a server error for request id=3
replayer.set_response_override(3, {
    "jsonrpc": "2.0",
    "id": 3,
    "error": {"code": -32603, "message": "Internal server error"}
})

# Your client code should handle this gracefully
response = replayer.handle_request(request)
assert handle_error(response) == expected_fallback

4. Offline Development

Working on a plane? At a coffee shop with bad WiFi? Record your MCP server interactions beforehand and develop against the replay:

# Before going offline
agent-vcr record --transport sse --server-url http://localhost:3000/sse -o dev-session.vcr

# While offline — full mock server on port 3100
agent-vcr replay --file dev-session.vcr --transport sse --port 3100

5. Multi-Agent Regression Testing

When multiple AI agents share MCP infrastructure, one agent's server change can break another. Agent VCR lets each team maintain their own cassettes and run compatibility checks independently.

6. Protocol Evolution Tracking

As the MCP spec evolves, use diffs to track how your server's behavior changes across protocol versions:

result = MCPDiff.compare("mcp-2024-11.vcr", "mcp-2025-03.vcr")
print(f"Added methods: {len(result.added_interactions)}")
print(f"Breaking changes: {len(result.breaking_changes)}")

Quick Start

New to Agent VCR? Follow the hands-on tutorial — 8 hands-on labs covering every use case with real commands.

Installation

Python:

# Recommended
uv pip install agent-vcr

# Or with pip
pip install agent-vcr

TypeScript/Node.js:

npm install @agent-vcr/core
# or
pnpm add @agent-vcr/core

Record a session

# Record stdio-based MCP server (try it now with the included demo server)
agent-vcr record --transport stdio --server-command "python demo/servers/calculator_v1.py" -o session.vcr

# Record SSE-based MCP server (replace URL with your server)
agent-vcr record --transport sse --server-url http://localhost:3000/sse -o session.vcr

Replay as a mock server

# Replay via stdio (pipe to your client)
agent-vcr replay --file session.vcr --transport stdio

# Replay via HTTP+SSE
agent-vcr replay --file session.vcr --transport sse --port 3100

Diff two recordings

agent-vcr diff baseline.vcr current.vcr
agent-vcr diff baseline.vcr current.vcr --format json --fail-on-breaking

Index and search many cassettes

agent-vcr index recordings/ -o index.json
agent-vcr search index.json --method tools/list
agent-vcr search index.json --endpoint-id github

Batch diff

# pairs.json: {"pairs": [{"baseline": "v1.vcr", "current": "v2.vcr"}, ...]}
agent-vcr diff-batch pairs.json --fail-on-breaking

Validate, merge, and analyze

# Validate a recording's schema and structure
agent-vcr validate session.vcr

# Merge multiple recordings into one
agent-vcr merge session1.vcr session2.vcr -o combined.vcr --deduplicate

# Show statistics (method distribution, latency percentiles, error rate)
agent-vcr stats session.vcr

Inspect a recording

agent-vcr inspect session.vcr
agent-vcr inspect session.vcr --format table

Match Strategies

The replayer supports 5 matching strategies for finding recorded responses:

Strategy	Description	Use Case
exact	Full JSON match (excluding jsonrpc and id fields)	Strictest testing
method	Match by method name only	Broad matching
method_and_params	Match method + full params (default)	Standard testing
subset	Match method + partial params (subset)	Flexible testing
sequential	Return interactions in order	Ordered replay

Note: The fuzzy strategy is deprecated; use subset instead. fuzzy is kept as an alias for backward compatibility.

Replayer Features

The replayer supports latency simulation for realistic testing:

# Simulate latency during replay
agent-vcr replay --file session.vcr --simulate-latency

# Scale recorded latencies (1.0 = original, 2.0 = double)
agent-vcr replay --file session.vcr --simulate-latency --latency-multiplier 2.0

Diff Features

Enhanced diff capabilities:

# Compare latency between recordings
agent-vcr diff baseline.vcr current.vcr --compare-latency

Try It Now

The repo ships with sample .vcr cassettes so you can try the CLI immediately:

# Inspect a recording
agent-vcr inspect examples/recordings/calculator-v1.vcr

# Diff two server versions — spot the new tool and schema changes
agent-vcr diff examples/recordings/calculator-v1.vcr examples/recordings/calculator-v2.vcr

# See error handling in action
agent-vcr inspect examples/recordings/calculator-errors.vcr

Sample cassettes included:

File	Description	Interactions
calculator-v1.vcr	Calculator MCP server v1 — add, multiply	3
calculator-v2.vcr	Calculator v2 — adds divide tool + response metadata	4
calculator-errors.vcr	Error scenarios — division by zero, method not found	4

Programmatic Usage

Creating recordings manually

from datetime import datetime
from agent_vcr.core.format import (
    JSONRPCRequest, JSONRPCResponse, VCRInteraction,
    VCRMetadata, VCRRecording, VCRSession,
)

# Build the initialize handshake
init_req = JSONRPCRequest(id=0, method="initialize", params={
    "protocolVersion": "2024-11-05",
    "clientInfo": {"name": "my-client", "version": "1.0.0"},
})
init_resp = JSONRPCResponse(id=0, result={
    "protocolVersion": "2024-11-05",
    "serverInfo": {"name": "my-server", "version": "1.0.0"},
    "capabilities": {"tools": {}},
})

# Build an interaction
interaction = VCRInteraction(
    sequence=0,
    timestamp=datetime.now(),
    direction="client_to_server",
    request=JSONRPCRequest(id=1, method="tools/list", params={}),
    response=JSONRPCResponse(id=1, result={
        "tools": [{"name": "echo", "description": "Echo a message"}]
    }),
    latency_ms=12.5,
)

# Assemble and save
recording = VCRRecording(
    metadata=VCRMetadata(
        version="1.0.0",
        recorded_at=datetime.now(),
        transport="stdio",
    ),
    session=VCRSession(
        initialize_request=init_req,
        initialize_response=init_resp,
        interactions=[interaction],
    ),
)
recording.save("session.vcr")

Replaying in code

from agent_vcr.core.format import VCRRecording
from agent_vcr.replayer import MCPReplayer

recording = VCRRecording.load("session.vcr")
replayer = MCPReplayer(recording, match_strategy="method_and_params")

response = replayer.handle_request({
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/list",
    "params": {}
})
print(response)  # Returns the recorded response

Diffing recordings

from agent_vcr.diff import MCPDiff

result = MCPDiff.compare("baseline.vcr", "current.vcr")

if result.is_identical:
    print("No changes!")
elif result.is_compatible:
    print(f"Compatible changes: {len(result.modified_interactions)} modified")
else:
    print("Breaking changes detected!")
    for change in result.breaking_changes:
        print(f"  - {change}")

Pytest Integration

Agent VCR includes a pytest plugin for seamless test integration.

Using fixtures

import pytest

@pytest.mark.vcr("cassettes/test_tools_list.vcr")
def test_tools_list(vcr_replayer):
    """Test that tools/list returns expected tools."""
    response = vcr_replayer.handle_request({
        "jsonrpc": "2.0",
        "id": 1,
        "method": "tools/list",
        "params": {}
    })
    assert "result" in response
    assert len(response["result"]["tools"]) > 0

Using the async context manager

from agent_vcr.pytest_plugin import vcr_cassette

async def test_with_cassette():
    async with vcr_cassette("my_test.vcr") as cassette:
        response = cassette.replayer.handle_request({
            "jsonrpc": "2.0",
            "id": 1,
            "method": "tools/call",
            "params": {"name": "echo", "arguments": {"message": "hello"}}
        })
        assert response["result"]["content"][0]["text"] == "hello"

CLI options

pytest --vcr-record            # Record new cassettes
pytest --vcr-dir=my_cassettes  # Custom cassette directory

VCR File Format

Recordings use a JSON-based .vcr format:

{
  "format_version": "1.0.0",
  "metadata": {
    "version": "1.0.0",
    "recorded_at": "2026-02-07T10:30:00",
    "transport": "stdio",
    "client_info": {"name": "claude-desktop"},
    "server_info": {"name": "my-mcp-server"},
    "tags": {"env": "staging"}
  },
  "session": {
    "initialize_request": { "jsonrpc": "2.0", "id": 0, "method": "initialize", "params": {} },
    "initialize_response": { "jsonrpc": "2.0", "id": 0, "result": { "capabilities": {} } },
    "capabilities": {},
    "interactions": [
      {
        "sequence": 0,
        "timestamp": "2026-02-07T10:30:05",
        "direction": "client_to_server",
        "request": { "jsonrpc": "2.0", "id": 1, "method": "tools/list", "params": {} },
        "response": { "jsonrpc": "2.0", "id": 1, "result": { "tools": [] } },
        "latency_ms": 12.5
      }
    ]
  }
}

Python vs TypeScript

The Python implementation is complete and tested (250+ tests). The TypeScript/Node.js port mirrors the same architecture and has a full unit test suite (72 tests).

Aspect	Python	TypeScript
Status	Production-ready	72 unit tests, source complete
Tests	250+ tests passing	72 unit tests in tests/unit/
CLI	Fully functional	Fully functional
Test framework	pytest plugin	Jest/Vitest (replay mode)
Recording format	.vcr (JSON)	.vcr (JSON) — same format
Recording mode in integrations	Implemented	Planned for v0.2.0

Cross-language recordings: The .vcr format is plain JSON, so recordings created by Python are loadable by the TypeScript implementation.

Scaling (Multi-MCP, Agent-to-Agent)

We support multi-MCP and agent-to-agent: record multiple sessions (one .vcr per client↔server session), tag each with --session-id, --endpoint-id, and --agent-id, and correlate them in your tests or tooling. Indexing (agent-vcr index, agent-vcr search) and batch diff (agent-vcr diff-batch) let you work across many cassettes. For design and agent-to-agent patterns, see docs/scaling.md.

Example — correlation metadata (record with endpoint/session ids, inspect shows them):

Architecture

See docs/architecture.md for the full system design, data flow diagrams, and design decisions.

Python:

python/src/agent_vcr/
├── core/
│   ├── format.py      # Pydantic models for .vcr format
│   ├── matcher.py     # Request matching strategies
│   └── session.py     # Session lifecycle management
├── transport/
│   ├── base.py        # Abstract transport interface
│   ├── stdio.py       # Subprocess stdio proxy
│   └── sse.py         # HTTP+SSE proxy
├── recorder.py        # Transparent recording proxy
├── replayer.py        # Mock server from recordings
├── diff.py            # Recording comparison engine
├── indexer.py         # Index/search many .vcr files
├── cli.py             # Command-line interface
└── pytest_plugin.py   # Pytest integration

TypeScript:

typescript/src/
├── core/
│   ├── format.ts      # Zod schemas for .vcr format
│   ├── matcher.ts     # Request matching strategies
│   └── session.ts     # Session lifecycle management
├── transport/
│   ├── base.ts        # Abstract transport interface
│   ├── stdio.ts       # Subprocess stdio proxy
│   └── sse.ts         # HTTP+SSE proxy
├── recorder.ts        # Transparent recording proxy
├── replayer.ts        # Mock server from recordings
├── diff.ts            # Recording comparison engine
├── cli.ts             # Command-line interface
└── integrations/
    ├── jest.ts        # Jest integration
    └── vitest.ts      # Vitest integration

Development

Python

# Clone and install
git clone https://github.com/jarvis2021/agent-vcr.git
cd agent-vcr/python

# Setup with uv (recommended)
uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=src/agent_vcr --cov-report=html

# Lint
ruff check src/

# Type check
mypy src/

TypeScript

cd agent-vcr/typescript

npm install
npm run build
npm test

Creating demo GIFs

Use asciinema and agg. See demo/README-GIFS.md for per-lab commands (bash demo/make-lab-gifs.sh <1-12>) and the correlation demo (demo/record-correlation-demo.sh).

• Record (above): assets/lab-1-record.gif — from lab 1 (make-lab-gifs.sh 1), then agg demo/lab-1.cast assets/lab-1-record.gif.
• Replay (above): assets/lab-2-replay.gif — from lab 2 (asciinema rec demo/lab-2.cast -c "bash demo/make-lab-gifs.sh 2"), then agg demo/lab-2.cast assets/lab-2-replay.gif.
• Diff (Real-World Examples): assets/lab-3-diff.gif. Correlation: assets/correlation-demo.gif via demo/record-correlation-demo.sh.

Run all tests (Python + TypeScript)

From the repo root:

cd python && uv run pytest tests/ -v

Contributing

Contributions are welcome! Please see docs/architecture.md for system design context and CONTRIBUTING.md for guidelines.

License

MIT