Bellwether
Open-source CLI for testing MCP servers and detecting schema drift
Bellwether
The open-source MCP testing tool. Catch drift before your users do.
What is MCP? MCP (Model Context Protocol) is how AI assistants like Claude connect to external tools—read files, query databases, call APIs. When those tool schemas change, AI workflows break silently.
Why Bellwether?
MCP servers expose tools with JSON schemas. When those schemas change—a parameter renamed, a type modified, a tool removed—AI agents break silently. Bellwether catches these changes before they reach production.
| Problem | Solution |
|---|---|
| Breaking changes slip into production | Drift detection catches schema changes in CI |
| No standard for MCP testing | Native MCP support understands tools, prompts, resources |
| Manual testing misses edge cases | Automated exploration covers what humans miss |
| Documentation gets stale | CONTRACT.md generated from actual behavior |
Quick Start
npm install -g @dotsetlabs/bellwether
bellwether init npx @mcp/your-server
bellwether check
That's it. No API keys. No LLM costs. Runs in seconds.
Product Focus
Bellwether is intentionally opinionated:
- Core workflow (default):
init->check->baseline - Advanced workflow (opt-in):
explore,watch,discover,golden,contract,registry
If you only need CI-safe drift detection, you can stay entirely in the core workflow.
Two Modes
| Mode | Purpose | Cost | When to Use |
|---|---|---|---|
check | Schema drift detection | Free | CI/CD, every PR |
explore | LLM-powered behavioral testing | LLM API costs | Local dev, deep analysis |
Most users only need check. It's deterministic, fast, and catches the changes that break AI agents.
CI/CD Workflow
Store your baseline in git. Run checks in CI. No account needed.
# 1. Initialize and save baseline (one-time setup)
bellwether init npx @mcp/your-server
bellwether check
bellwether baseline save
git add bellwether.yaml bellwether-baseline.json
git commit -m "Add Bellwether baseline"
# 2. Add to CI (.github/workflows/bellwether.yml)
name: MCP Drift Detection
on: [pull_request]
jobs:
check:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- run: npx @dotsetlabs/bellwether check --fail-on-drift
What It Detects
| Change | Example | Severity |
|---|---|---|
| Tool added/removed | delete_file appears or disappears | Breaking |
| Schema changed | Parameter path becomes required | Breaking |
| Parameter renamed | path to file_path | Breaking |
| Description changed | Tool help text updated | Warning |
| Performance regression | Latency increased >10% | Warning |
| Tool annotations changed | readOnlyHint flipped to false | Warning |
| Output schema changed | Return type structure modified | Warning |
| Entity title changed | Tool/prompt/resource title updated | Info |
| Task support changed | Execution mode switched to async | Warning |
| Server instructions changed | Server-level instructions updated | Info |
| Prompt added/removed | Prompt template appears or disappears | Breaking |
| Resource changed | Resource URI or MIME type modified | Warning |
Comparisons are protocol-version-aware — version-specific fields (annotations, titles, output schemas, etc.) are only compared when both baselines support the relevant MCP protocol version.
Command Tiers
Core Commands (Recommended)
| Command | Purpose |
|---|---|
init | Create bellwether.yaml |
check | Deterministic schema drift detection |
baseline save | Save snapshot for future comparisons |
baseline compare | Compare latest check output to saved baseline |
Advanced Commands (Optional)
| Command | Purpose |
|---|---|
explore | LLM behavioral testing and AGENTS.md generation |
watch | Continuous checking on file changes |
discover | Capability inspection without tests |
registry | Search MCP Registry |
golden | Golden output regression testing |
contract | Contract validation and generation |
auth | Manage LLM provider API keys |
validate-config | Validate bellwether.yaml without running tests |
CI/CD Exit Codes
| Code | Meaning | Suggested Action |
|---|---|---|
0 | No changes | Pass |
1 | Info-level changes | Pass or warn |
2 | Warning-level changes | Warn |
3 | Breaking changes | Fail |
4 | Runtime error | Fail |
5 | Low confidence metrics | Warn or fail |
GitHub Action
- uses: dotsetlabs/[email protected]
with:
version: '2.1.3'
server-command: 'npx @mcp/your-server'
baseline-path: './bellwether-baseline.json'
fail-on-severity: 'warning'
Configuration
All settings live in bellwether.yaml. Create one with presets:
bellwether init npx @mcp/your-server # Default (free, fast)
bellwether init --preset ci npx @mcp/server # Optimized for CI/CD
bellwether init --preset local npx @mcp/server # Local Ollama (free)
For remote MCP servers that require auth headers, configure:
server:
transport: sse
url: "https://api.example.com/mcp"
headers:
Authorization: "Bearer ${MCP_SERVER_TOKEN}"
Or use one-off CLI overrides:
bellwether check -H "Authorization: Bearer $MCP_SERVER_TOKEN"
Environment Variables
| Variable | Description |
|---|---|
OPENAI_API_KEY | OpenAI API key (explore only) |
ANTHROPIC_API_KEY | Anthropic API key (explore only) |
OLLAMA_BASE_URL | Ollama URL (default: http://localhost:11434) |
Documentation
docs.bellwether.sh — Full reference for configuration and commands.
Project Governance
Community
- GitHub Discussions - Questions and ideas
- GitHub Issues - Bug reports
- Contributing - How to contribute
Development
git clone https://github.com/dotsetlabs/bellwether
cd bellwether
npm install
npm run build
npm test
License
MIT License - see LICENSE for details.
Built by Dotset Labs
Related Servers
Scout Monitoring MCP
sponsorPut performance and error data directly in the hands of your AI assistant.
Alpha Vantage MCP Server
sponsorAccess financial market data: realtime & historical stock, ETF, options, forex, crypto, commodities, fundamentals, technical indicators, & more
Nexus MCP Bridge for VSCode
A bridge that connects Claude Desktop to your VSCode workspace, enabling local file system access.
Berry MCP Server
A universal framework for easily creating and deploying Model Context Protocol servers with any tools.
maximumsats-mcp
Bitcoin Lightning + Nostr Web-of-Trust tools for agents (L402 pay-per-call endpoints)
Dive AI Agent
An open-source desktop application for hosting MCP servers that integrates with function-calling LLMs.
Remote MCP Server (Authless)
A remote MCP server deployable on Cloudflare Workers, without authentication.
GitHub Workflow Debugger MCP
Diagnose and fix GitHub Actions workflow failures using the GitHub API.
cxpak
Spends CPU cycles so you don't spend tokens. The LLM gets a briefing packet instead of a flashlight in a dark room.
Replicate FLUX.1 Kontext [Max]
Image generation and editing using the FLUX.1 Kontext [Max] model via the Replicate API, featuring advanced text rendering and contextual understanding.
Honeybadger
Interact with the Honeybadger API for error and uptime monitoring.
Model Context Protocol servers
A collection of reference implementations for the Model Context Protocol (MCP), showcasing various MCP servers implemented with TypeScript and Python SDKs.