mcpbr

# One-liner install (installs + runs quick test)
curl -sSL https://raw.githubusercontent.com/greynewell/mcpbr/main/install.sh | bash

# Or install and run manually
pip install mcpbr && mcpbr run -n 1

Benchmark your MCP server against real GitHub issues. One command, hard numbers.

Model Context Protocol Benchmark Runner

Stop guessing if your MCP server actually helps. Get hard numbers comparing tool-assisted vs. baseline agent performance on real GitHub issues.

What You Get

Real metrics showing whether your MCP server improves agent performance on SWE-bench tasks. No vibes, just data.

Why mcpbr?

MCP servers promise to make LLMs better at coding tasks. But how do you prove it?

mcpbr runs controlled experiments: same model, same tasks, same environment - the only variable is your MCP server. You get:

• Apples-to-apples comparison against a baseline agent
• Real GitHub issues from SWE-bench (not toy examples)
• Reproducible results via Docker containers with pinned dependencies

Read the full origin story: Why I Built mcpbr — the problem, the approach, and where the project is headed.

Supported Benchmarks

mcpbr supports 30+ benchmarks across 10 categories through a flexible abstraction layer:

Featured Benchmarks

SWE-bench (Default) - Real GitHub issues requiring bug fixes. Three variants: Verified (500 manually validated), Lite (300 curated), and Full (2,294 complete). Pre-built Docker images available.

CyberGym - Security vulnerabilities requiring PoC exploits. 4 difficulty levels controlling context. Uses AddressSanitizer for crash detection.

MCPToolBench++ - Large-scale MCP tool use evaluation across 45+ categories. Tests tool discovery, selection, invocation, and result interpretation.

GSM8K - Grade-school math word problems testing chain-of-thought reasoning with numeric answer matching.

# Run SWE-bench Verified (default)
mcpbr run -c config.yaml

# Run any benchmark
mcpbr run -c config.yaml --benchmark humaneval -n 20
mcpbr run -c config.yaml --benchmark gsm8k -n 50
mcpbr run -c config.yaml --benchmark cybergym --level 2

# List all available benchmarks
mcpbr benchmarks

See the benchmarks guide for details on each benchmark and how to configure them.

Overview

This harness runs two parallel evaluations for each task:

1. MCP Agent: LLM with access to tools from your MCP server
2. Baseline Agent: LLM without tools (single-shot generation)

By comparing these, you can measure the effectiveness of your MCP server for different software engineering tasks. See the MCP integration guide for tips on testing your server.

Regression Detection

mcpbr includes built-in regression detection to catch performance degradations between MCP server versions:

Key Features

• Automatic Detection: Compare current results against a baseline to identify regressions
• Detailed Reports: See exactly which tasks regressed and which improved
• Threshold-Based Exit Codes: Fail CI/CD pipelines when regression rate exceeds acceptable limits
• Multi-Channel Alerts: Send notifications via Slack, Discord, or email

How It Works

A regression is detected when a task that passed in the baseline now fails in the current run. This helps you catch issues before deploying new versions of your MCP server.

# First, run a baseline evaluation and save results
mcpbr run -c config.yaml -o baseline.json

# Later, compare a new version against the baseline
mcpbr run -c config.yaml --baseline-results baseline.json --regression-threshold 0.1

# With notifications
mcpbr run -c config.yaml --baseline-results baseline.json \
  --regression-threshold 0.1 \
  --slack-webhook https://hooks.slack.com/services/YOUR/WEBHOOK/URL

Use Cases

• CI/CD Integration: Automatically detect regressions in pull requests
• Version Comparison: Compare different versions of your MCP server
• Performance Monitoring: Track MCP server performance over time
• Team Notifications: Alert your team when regressions are detected

Example Output

======================================================================
REGRESSION DETECTION REPORT
======================================================================

Total tasks compared: 25
Regressions detected: 2
Improvements detected: 5
Regression rate: 8.0%

REGRESSIONS (previously passed, now failed):
----------------------------------------------------------------------
  - django__django-11099
    Error: Timeout
  - sympy__sympy-18087
    Error: Test suite failed

IMPROVEMENTS (previously failed, now passed):
----------------------------------------------------------------------
  - astropy__astropy-12907
  - pytest-dev__pytest-7373
  - scikit-learn__scikit-learn-25570
  - matplotlib__matplotlib-23913
  - requests__requests-3362

======================================================================

For CI/CD integration, use --regression-threshold to fail the build when regressions exceed an acceptable rate:

# .github/workflows/test-mcp.yml
- name: Run mcpbr with regression detection
  run: |
    mcpbr run -c config.yaml \
      --baseline-results baseline.json \
      --regression-threshold 0.1 \
      -o current.json

This will exit with code 1 if the regression rate exceeds 10%, failing the CI job.

Installation

Full installation guide with detailed setup instructions.

via npm

# Run with npx (no installation)
npx mcpbr-cli run -c config.yaml

# Or install globally
npm install -g mcpbr-cli
mcpbr run -c config.yaml

Package: mcpbr-cli on npm

Note: The npm package requires Python 3.11+ and the mcpbr Python package (pip install mcpbr)

via pip

# Install from PyPI
pip install mcpbr

# Or install from source
git clone https://github.com/greynewell/mcpbr.git
cd mcpbr
pip install -e .

# Or with uv
uv pip install -e .

Note for Apple Silicon users: The harness automatically uses x86_64 Docker images via emulation. This may be slower than native ARM64 images but ensures compatibility with all SWE-bench tasks.

Quick Start

Option 1: Use Example Configurations (Recommended)

Get started in seconds with our example configurations:

# Set your API key
export ANTHROPIC_API_KEY="your-api-key"

# Run your first evaluation using an example config
mcpbr run -c examples/quick-start/getting-started.yaml -v

This runs 5 SWE-bench tasks with the filesystem server. Expected runtime: 15-30 minutes, cost: $2-5.

Explore 25+ example configurations in the examples/ directory:

• Quick Start: Getting started, testing servers, comparing models
• Benchmarks: SWE-bench Lite/Full, CyberGym basic/advanced
• MCP Servers: Filesystem, GitHub, Brave Search, databases, custom servers
• Scenarios: Cost-optimized, performance-optimized, CI/CD, regression detection

See the Examples README for the complete guide.

Option 2: Generate Custom Configuration

1. Set your API key:

export ANTHROPIC_API_KEY="your-api-key"

2. Run mcpbr (config auto-created if missing):

# Config is auto-created on first run
mcpbr run -n 1

# Or explicitly generate a config file first
mcpbr init

3. Edit the configuration to point to your MCP server:

mcp_server:
  command: "npx"
  args:
    - "-y"
    - "@modelcontextprotocol/server-filesystem"
    - "{workdir}"
  env: {}

provider: "anthropic"
agent_harness: "claude-code"

model: "sonnet"  # or full name: "claude-sonnet-4-5-20250929"
dataset: "SWE-bench/SWE-bench_Lite"
sample_size: 10
timeout_seconds: 300
max_concurrent: 4

# Optional: disable default logging (logs are saved to output_dir/logs/ by default)
# disable_logs: true

4. Run the evaluation:

mcpbr run --config config.yaml

Infrastructure Modes

mcpbr supports running evaluations on different infrastructure platforms, allowing you to scale evaluations or offload compute-intensive tasks to cloud VMs.

Local (Default)

Run evaluations on your local machine:

infrastructure:
  mode: local  # default

This is the default mode - evaluations run directly on your machine using local Docker containers.

Azure VM

Run evaluations on Azure Virtual Machines with automatic provisioning and cleanup:

infrastructure:
  mode: azure
  azure:
    resource_group: mcpbr-benchmarks
    location: eastus
    cpu_cores: 10
    memory_gb: 40

Key features:

• Zero manual VM setup - provisioned automatically from config
• Automatic Docker, Python, and mcpbr installation
• Test task validation before full evaluation
• Auto-cleanup after completion (configurable)
• Cost-optimized with automatic VM deletion

Example usage:

# Run evaluation on Azure VM
mcpbr run -c azure-config.yaml

# VM is automatically created, evaluation runs, results are downloaded, VM is deleted

See docs/infrastructure/azure.md for full documentation including:

• Prerequisites and authentication
• VM sizing and cost estimation
• Debugging with preserve_on_error
• Troubleshooting guide

Side-by-Side Server Comparison

Compare two MCP servers head-to-head in a single evaluation run to see which implementation performs better.

Quick Example

# comparison-config.yaml
comparison_mode: true

mcp_server_a:
  name: "Task Queries"
  command: node
  args: [build/index.js]
  cwd: /path/to/task-queries

mcp_server_b:
  name: "Edge Identity"
  command: node
  args: [build/index.js]
  cwd: /path/to/edge-identity

benchmark: swe-bench-lite
sample_size: 10

mcpbr run -c comparison-config.yaml -o results.json

Results Output

Side-by-Side MCP Server Comparison

┏━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━┳━━━━━━━━━━┓
┃ Metric            ┃ Task Queries ┃ Edge Identity┃ Δ (A - B)┃
┡━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━╇━━━━━━━━━━┩
│ Resolved Tasks    │ 4/10         │ 2/10         │ +2       │
│ Resolution Rate   │ 40.0%        │ 20.0%        │ +100.0%  │
└───────────────────┴──────────────┴──────────────┴──────────┘

✓ Task Queries unique wins: 2 tasks
  - django__django-12286
  - astropy__astropy-7606

Use cases:

• A/B testing: Compare optimized vs. baseline implementations
• Tool evaluation: Test different MCP tool sets
• Version comparison: Benchmark v2.0 vs. v1.5

See docs/comparison-mode.md for complete documentation.

Claude Code Integration

mcpbr includes a built-in Claude Code plugin that makes Claude an expert at running benchmarks correctly. The plugin provides specialized skills and knowledge about mcpbr configuration, execution, and troubleshooting.

Installation Options

You have three ways to enable the mcpbr plugin in Claude Code:

Option 1: Clone Repository (Automatic Detection)

When you clone this repository, Claude Code automatically detects and loads the plugin:

git clone https://github.com/greynewell/mcpbr.git
cd mcpbr

# Plugin is now active - try asking Claude:
# "Run the SWE-bench Lite eval with 5 tasks"

Best for: Contributors, developers testing changes, or users who want the latest unreleased features.

Option 2: npm Global Install (Planned for v0.4.0)

Install the plugin globally via npm for use across any project:

# Planned for v0.4.0 (not yet released)
npm install -g @mcpbr/claude-code-plugin

Note: The npm package is not yet published. This installation method will be available in a future release. Track progress in issue #265.

Best for: Users who want plugin features available in any directory.

Option 3: Claude Code Plugin Manager (Planned for v0.4.0)

Install via Claude Code's built-in plugin manager:

1. Open Claude Code settings
2. Navigate to Plugins > Browse
3. Search for "mcpbr"
4. Click Install

Note: Plugin manager installation is not yet available. This installation method will be available after plugin marketplace submission. Track progress in issue #267.

Best for: Users who prefer a GUI and want automatic updates.

Installation Comparison

What You Get

The plugin includes three specialized skills that enhance Claude's ability to work with mcpbr:

1. run-benchmark

Expert at running evaluations with proper validation and error handling.

Capabilities:

• Validates prerequisites (Docker running, API keys set, config files exist)
• Constructs correct mcpbr run commands with appropriate flags
• Handles errors gracefully with actionable troubleshooting steps
• Monitors progress and provides meaningful status updates

Example interactions:

• "Run the SWE-bench Lite benchmark with 10 tasks"
• "Evaluate my MCP server using CyberGym level 2"
• "Test my config with a single task"

2. generate-config

Generates valid mcpbr configuration files with benchmark-specific templates.

Capabilities:

• Ensures required {workdir} placeholder is included in MCP server args
• Validates MCP server command syntax
• Provides templates for different benchmarks (SWE-bench, CyberGym, MCPToolBench++)
• Suggests appropriate timeouts and concurrency settings

Example interactions:

• "Generate a config for the filesystem MCP server"
• "Create a config for testing my custom MCP server"
• "Set up a CyberGym evaluation config"

3. swe-bench-lite

Quick-start command for running SWE-bench Lite evaluations.

Capabilities:

• Pre-configured for 5-task evaluation (fast testing)
• Includes sensible defaults for output files and logging
• Perfect for demonstrations and initial testing
• Automatically sets up verbose output for debugging

Example interactions:

• "Run a quick SWE-bench Lite test"
• "Show me how mcpbr works"
• "Test the filesystem server"

Benefits

When using Claude Code with the mcpbr plugin active, Claude will automatically:

• Verify Docker is running before starting evaluations
• Check for required API keys (ANTHROPIC_API_KEY)
• Generate valid configurations with proper {workdir} placeholders
• Use correct CLI flags and avoid deprecated options
• Provide contextual troubleshooting when issues occur
• Follow mcpbr best practices for optimal results

Troubleshooting

Plugin not detected in cloned repository:

• Ensure you're in the repository root directory
• Verify the claude-code.json file exists in the repo
• Try restarting Claude Code

Skills not appearing:

• Check Claude Code version (requires v2.0+)
• Verify plugin is listed in Settings > Plugins
• Try running /reload-plugins in Claude Code

Commands failing:

• Ensure mcpbr is installed: pip install mcpbr
• Verify Docker is running: docker info
• Check API key is set: echo $ANTHROPIC_API_KEY

For more help, see the troubleshooting guide or open an issue.

Configuration

Full configuration reference with all options and examples.

MCP Server Configuration

The mcp_server section defines how to start your MCP server:

Example Configurations

Anthropic Filesystem Server:

mcp_server:
  command: "npx"
  args: ["-y", "@modelcontextprotocol/server-filesystem", "{workdir}"]

Custom Python MCP Server:

mcp_server:
  command: "python"
  args: ["-m", "my_mcp_server", "--workspace", "{workdir}"]
  env:
    LOG_LEVEL: "debug"

Supermodel Codebase Analysis Server:

mcp_server:
  command: "npx"
  args: ["-y", "@supermodeltools/mcp-server"]
  env:
    SUPERMODEL_API_KEY: "${SUPERMODEL_API_KEY}"

MCP Timeout Configuration

mcpbr supports configurable timeouts for MCP server operations to handle different server types and workloads.

Configuration Fields

These fields map to the MCP_TIMEOUT and MCP_TOOL_TIMEOUT environment variables used by Claude Code. See the Claude Code settings documentation for more details.

Example Configuration

mcp_server:
  command: "npx"
  args: ["-y", "@modelcontextprotocol/server-filesystem", "{workdir}"]
  startup_timeout_ms: 60000      # 60 seconds for server to start
  tool_timeout_ms: 900000        # 15 minutes for long-running tools

Common Timeout Values

Different server types require different timeout settings based on their operational characteristics:

When to adjust timeouts:

• Increase startup_timeout_ms if your server takes longer to initialize (e.g., loading large models, establishing database connections)
• Increase tool_timeout_ms if your tools perform long-running operations (e.g., codebase analysis, file processing, AI inference)
• Decrease timeouts for fast servers to fail quickly on connection issues

Custom Agent Prompt

You can customize the prompt sent to the agent using the agent_prompt field:

agent_prompt: |
  Fix the following bug in this repository:

  {problem_statement}

  Make the minimal changes necessary to fix the issue.
  Focus on the root cause, not symptoms.

Use {problem_statement} as a placeholder for the SWE-bench issue text. You can also override the prompt via CLI with --prompt.

Evaluation Parameters

CLI Reference

Full CLI documentation with all commands and options.

Get help for any command with --help or -h:

mcpbr --help
mcpbr run --help
mcpbr init --help

Commands Overview

mcpbr run

Run SWE-bench evaluation with the configured MCP server.

# Full evaluation (MCP + baseline)
mcpbr run -c config.yaml

# Run only MCP evaluation
mcpbr run -c config.yaml -M

# Run only baseline evaluation
mcpbr run -c config.yaml -B

# Override model
mcpbr run -c config.yaml -m claude-3-5-sonnet-20241022

# Override sample size
mcpbr run -c config.yaml -n 50

# Save results and report
mcpbr run -c config.yaml -o results.json -r report.md

# Save JUnit XML for CI/CD
mcpbr run -c config.yaml --output-junit junit.xml

# Run specific tasks
mcpbr run -c config.yaml -t astropy__astropy-12907 -t django__django-11099

# Verbose output with per-instance logs
mcpbr run -c config.yaml -v --log-dir logs/

# Very verbose output
mcpbr run -c config.yaml -vv

# Run CyberGym benchmark
mcpbr run -c config.yaml --benchmark cybergym --level 2

# Run CyberGym with specific tasks
mcpbr run -c config.yaml --benchmark cybergym --level 3 -n 5

# Regression detection - compare against baseline
mcpbr run -c config.yaml --baseline-results baseline.json

# Regression detection with threshold (exit 1 if exceeded)
mcpbr run -c config.yaml --baseline-results baseline.json --regression-threshold 0.1

# Regression detection with Slack notifications
mcpbr run -c config.yaml --baseline-results baseline.json --slack-webhook https://hooks.slack.com/...

# Regression detection with Discord notifications
mcpbr run -c config.yaml --baseline-results baseline.json --discord-webhook https://discord.com/api/webhooks/...

# Regression detection with email notifications
mcpbr run -c config.yaml --baseline-results baseline.json \
  --email-to team@example.com --email-from mcpbr@example.com \
  --smtp-host smtp.gmail.com --smtp-port 587 \
  --smtp-user user@gmail.com --smtp-password "app-password"

mcpbr init

Generate an example configuration file.

mcpbr init
mcpbr init -o my-config.yaml

mcpbr models

List supported Anthropic models for evaluation.

mcpbr cleanup

Remove orphaned mcpbr Docker containers that were not properly cleaned up.

# Preview containers to remove
mcpbr cleanup --dry-run

# Remove containers with confirmation
mcpbr cleanup

# Remove containers without confirmation
mcpbr cleanup -f

Performance Profiling

mcpbr includes comprehensive performance profiling to understand MCP server overhead and identify optimization opportunities.

Enable Profiling

# Via CLI flag
mcpbr run -c config.yaml --profile

# Or in config.yaml
enable_profiling: true

What Gets Measured

• Tool call latencies with percentiles (p50, p95, p99)
• Memory usage (peak and average RSS/VMS)
• Infrastructure overhead (Docker and MCP server startup times)
• Tool discovery speed (time to first tool use)
• Tool switching overhead (time between tool calls)
• Automated insights from profiling data

Example Profiling Output

{
  "profiling": {
    "task_duration_seconds": 140.5,
    "tool_call_latencies": {
      "Read": {"count": 15, "avg_seconds": 0.8, "p95_seconds": 1.5},
      "Bash": {"avg_seconds": 2.3, "p95_seconds": 5.1}
    },
    "memory_profile": {"peak_rss_mb": 512.3, "avg_rss_mb": 387.5},
    "docker_startup_seconds": 2.1,
    "mcp_server_startup_seconds": 1.8
  }
}

Automated Insights

The profiler automatically identifies performance issues:

- Bash is the slowest tool (avg: 2.3s, p95: 5.1s)
- Docker startup adds 2.1s overhead per task
- Fast tool discovery: first tool use in 8.3s

See docs/profiling.md for complete profiling documentation.

Example Run

Here's what a typical evaluation looks like:

$ mcpbr run -c config.yaml -v -o results.json --log-dir my-logs

mcpbr Evaluation
  Config: config.yaml
  Provider: anthropic
  Model: sonnet
  Agent Harness: claude-code
  Dataset: SWE-bench/SWE-bench_Lite
  Sample size: 10
  Run MCP: True, Run Baseline: True
  Pre-built images: True
  Log dir: my-logs

Loading dataset: SWE-bench/SWE-bench_Lite
Evaluating 10 tasks
Provider: anthropic, Harness: claude-code
14:23:15 [MCP] Starting mcp run for astropy-12907:mcp
14:23:22 astropy-12907:mcp    > TodoWrite
14:23:22 astropy-12907:mcp    < Todos have been modified successfully...
14:23:26 astropy-12907:mcp    > Glob
14:23:26 astropy-12907:mcp    > Grep
14:23:27 astropy-12907:mcp    < $WORKDIR/astropy/modeling/separable.py
14:23:27 astropy-12907:mcp    < Found 5 files: astropy/modeling/tests/test_separable.py...
...
14:27:43 astropy-12907:mcp    * done turns=31 tokens=115/6,542
14:28:30 [BASELINE] Starting baseline run for astropy-12907:baseline
...

Output

Understanding evaluation results - detailed guide to interpreting output.

Console Output

The harness displays real-time progress with verbose mode (-v) and a final summary table:

Evaluation Results

                 Summary
+-----------------+-----------+----------+
| Metric          | MCP Agent | Baseline |
+-----------------+-----------+----------+
| Resolved        | 8/25      | 5/25     |
| Resolution Rate | 32.0%     | 20.0%    |
+-----------------+-----------+----------+

Improvement: +60.0%

Per-Task Results
+------------------------+------+----------+-------+
| Instance ID            | MCP  | Baseline | Error |
+------------------------+------+----------+-------+
| astropy__astropy-12907 | PASS |   PASS   |       |
| django__django-11099   | PASS |   FAIL   |       |
| sympy__sympy-18087     | FAIL |   FAIL   |       |
+------------------------+------+----------+-------+

Results saved to results.json

JSON Output (--output)

{
  "metadata": {
    "timestamp": "2026-01-17T07:23:39.871437+00:00",
    "config": {
      "model": "sonnet",
      "provider": "anthropic",
      "agent_harness": "claude-code",
      "dataset": "SWE-bench/SWE-bench_Lite",
      "sample_size": 25,
      "timeout_seconds": 600,
      "max_iterations": 30
    },
    "mcp_server": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "{workdir}"]
    }
  },
  "summary": {
    "mcp": {"resolved": 8, "total": 25, "rate": 0.32},
    "baseline": {"resolved": 5, "total": 25, "rate": 0.20},
    "improvement": "+60.0%"
  },
  "tasks": [
    {
      "instance_id": "astropy__astropy-12907",
      "mcp": {
        "patch_generated": true,
        "tokens": {"input": 115, "output": 6542},
        "iterations": 30,
        "tool_calls": 72,
        "tool_usage": {
          "TodoWrite": 4, "Task": 1, "Glob": 4,
          "Grep": 11, "Bash": 27, "Read": 22,
          "Write": 2, "Edit": 1
        },
        "resolved": true,
        "patch_applied": true,
        "fail_to_pass": {"passed": 2, "total": 2},
        "pass_to_pass": {"passed": 10, "total": 10}
      },
      "baseline": {
        "patch_generated": true,
        "tokens": {"input": 63, "output": 7615},
        "iterations": 30,
        "tool_calls": 57,
        "tool_usage": {
          "TodoWrite": 4, "Glob": 3, "Grep": 4,
          "Read": 14, "Bash": 26, "Write": 4, "Edit": 1
        },
        "resolved": true,
        "patch_applied": true
      }
    }
  ]
}

Output Directory Structure

By default, mcpbr consolidates all outputs into a single timestamped directory:

.mcpbr_run_20260126_133000/
├── config.yaml                # Copy of configuration used
├── evaluation_state.json      # Task results and state
├── logs/                      # Detailed MCP server logs
│   ├── task_1_mcp.log
│   ├── task_2_mcp.log
│   └── ...
└── README.txt                 # Auto-generated explanation

This makes it easy to:

• Archive results: tar -czf results.tar.gz .mcpbr_run_*
• Clean up: rm -rf .mcpbr_run_*
• Share: Just zip one directory

You can customize the output directory:

# Custom output directory
mcpbr run -c config.yaml --output-dir ./my-results

# Or in config.yaml
output_dir: "./my-results"

Note: The --output-dir CLI flag takes precedence over the output_dir config setting. This ensures that the README.txt file in the output directory reflects the final effective configuration values after all CLI overrides are applied.

Markdown Report (--report)

Generates a human-readable report with:

• Summary statistics
• Per-task results table
• Analysis of which tasks each agent solved

Per-Instance Logs (--log-dir)

Logging is enabled by default to prevent data loss. Detailed execution traces are automatically saved to output_dir/logs/ unless disabled.

To disable logging:

# Via CLI flag
mcpbr run -c config.yaml --disable-logs

# Or in config file
disable_logs: true

Creates a directory with detailed JSON log files for each task run. Filenames include timestamps to prevent overwrites:

my-logs/
  astropy__astropy-12907_mcp_20260117_143052.json
  astropy__astropy-12907_baseline_20260117_143156.json
  django__django-11099_mcp_20260117_144023.json
  django__django-11099_baseline_20260117_144512.json

Each log file contains the full stream of events from the agent CLI:

{
  "instance_id": "astropy__astropy-12907",
  "run_type": "mcp",
  "events": [
    {
      "type": "system",
      "subtype": "init",
      "cwd": "/workspace",
      "tools": ["Task", "Bash", "Glob", "Grep", "Read", "Edit", "Write", "TodoWrite"],
      "model": "claude-sonnet-4-5-20250929",
      "claude_code_version": "2.1.12"
    },
    {
      "type": "assistant",
      "message": {
        "content": [{"type": "text", "text": "I'll help you fix this bug..."}]
      }
    },
    {
      "type": "assistant",
      "message": {
        "content": [{"type": "tool_use", "name": "Grep", "input": {"pattern": "separability"}}]
      }
    },
    {
      "type": "result",
      "num_turns": 31,
      "usage": {"input_tokens": 115, "output_tokens": 6542}
    }
  ]
}

This is useful for debugging failed runs or analyzing agent behavior in detail.

JUnit XML Output (--output-junit)

The harness can generate JUnit XML reports for integration with CI/CD systems like GitHub Actions, GitLab CI, and Jenkins. Each task is represented as a test case, with resolved/unresolved tasks mapped to pass/fail states.

mcpbr run -c config.yaml --output-junit junit.xml

The JUnit XML report includes:

• Test Suites: Separate suites for MCP and baseline evaluations
• Test Cases: Each task is a test case with timing information
• Failures: Unresolved tasks with detailed error messages
• Properties: Metadata about model, provider, benchmark configuration
• System Output: Token usage, tool calls, and test results per task

CI/CD Integration Examples

GitHub Actions:

name: MCP Benchmark

on: [push, pull_request]

jobs:
  benchmark:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3

      - name: Set up Python
        uses: actions/setup-python@v4
        with:
          python-version: '3.11'

      - name: Install mcpbr
        run: pip install mcpbr

      - name: Run benchmark
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          mcpbr run -c config.yaml --output-junit junit.xml

      - name: Publish Test Results
        uses: EnricoMi/publish-unit-test-result-action@v2
        if: always()
        with:
          files: junit.xml

GitLab CI:

benchmark:
  image: python:3.11
  services:
    - docker:dind
  script:
    - pip install mcpbr
    - mcpbr run -c config.yaml --output-junit junit.xml
  artifacts:
    reports:
      junit: junit.xml

Jenkins:

pipeline {
    agent any
    stages {
        stage('Benchmark') {
            steps {
                sh 'pip install mcpbr'
                sh 'mcpbr run -c config.yaml --output-junit junit.xml'
            }
        }
    }
    post {
        always {
            junit 'junit.xml'
        }
    }
}

The JUnit XML format enables native test result visualization in your CI/CD dashboard, making it easy to track benchmark performance over time and identify regressions.

How It Works

Architecture deep dive - learn how mcpbr works internally.

1. Load Tasks: Fetches tasks from the selected benchmark (SWE-bench, CyberGym, or MCPToolBench++) via HuggingFace
2. Create Environment: For each task, creates an isolated Docker environment with the repository and dependencies
3. Run MCP Agent: Invokes Claude Code CLI inside the Docker container, letting it explore and generate a solution (patch or PoC)
4. Run Baseline: Same as MCP agent but without the MCP server
5. Evaluate: Runs benchmark-specific evaluation (test suites for SWE-bench, crash detection for CyberGym, tool use accuracy for MCPToolBench++)
6. Report: Aggregates results and calculates improvement

Pre-built Docker Images

The harness uses pre-built SWE-bench Docker images from Epoch AI's registry when available. These images come with:

• The repository checked out at the correct commit
• All project dependencies pre-installed and validated
• A consistent environment for reproducible evaluations

The agent (Claude Code CLI) runs inside the container, which means:

• Python imports work correctly (e.g., from astropy import ...)
• The agent can run tests and verify fixes
• No dependency conflicts with the host machine

If a pre-built image is not available for a task, the harness falls back to cloning the repository and attempting to install dependencies (less reliable).

Architecture

mcpbr/
├── src/mcpbr/
│   ├── cli.py           # Command-line interface
│   ├── config.py        # Configuration models
│   ├── models.py        # Supported model registry
│   ├── providers.py     # LLM provider abstractions (extensible)
│   ├── harnesses.py     # Agent harness implementations (extensible)
│   ├── benchmarks/      # Benchmark abstraction layer (25+ benchmarks)
│   │   ├── __init__.py      # Registry and factory
│   │   ├── base.py          # Benchmark protocol
│   │   ├── swebench.py      # SWE-bench (Verified/Lite/Full)
│   │   ├── cybergym.py      # CyberGym security
│   │   ├── humaneval.py     # HumanEval code generation
│   │   ├── gsm8k.py         # GSM8K math reasoning
│   │   ├── mcptoolbench.py  # MCPToolBench++ tool use
│   │   ├── apps.py          # APPS coding problems
│   │   ├── mbpp.py          # MBPP Python problems
│   │   ├── math_benchmark.py # MATH competition math
│   │   └── ...              # 15+ more benchmarks
│   ├── harness.py       # Main orchestrator
│   ├── agent.py         # Baseline agent implementation
│   ├── docker_env.py    # Docker environment management + in-container execution
│   ├── evaluation.py    # Patch application and testing
│   ├── log_formatter.py # Log formatting and per-instance logging
│   └── reporting.py     # Output formatting
├── tests/
│   ├── test_*.py        # Unit tests
│   ├── test_benchmarks.py # Benchmark tests
│   └── test_integration.py  # Integration tests
├── Dockerfile           # Fallback image for task environments
└── config/
    └── example.yaml     # Example configuration

The architecture uses Protocol-based abstractions for providers, harnesses, and benchmarks, making it easy to add support for additional LLM providers, agent backends, or software engineering benchmarks in the future. See the API reference and benchmarks guide for more details.

Execution Flow

┌─────────────────────────────────────────────────────────────────┐
│                         Host Machine                            │
│  ┌───────────────────────────────────────────────────────────┐  │
│  │                    mcpbr Harness (Python)                 │  │
│  │  - Loads SWE-bench tasks from HuggingFace                 │  │
│  │  - Pulls pre-built Docker images                          │  │
│  │  - Orchestrates agent runs                                │  │
│  │  - Collects results and generates reports                 │  │
│  └─────────────────────────┬─────────────────────────────────┘  │
│                            │ docker exec                        │
│  ┌─────────────────────────▼─────────────────────────────────┐  │
│  │              Docker Container (per task)                  │  │
│  │  ┌─────────────────────────────────────────────────────┐  │  │
│  │  │  Pre-built SWE-bench Image                          │  │  │
│  │  │  - Repository at correct commit                     │  │  │
│  │  │  - All dependencies installed (astropy, django...)  │  │  │
│  │  │  - Node.js + Claude CLI (installed at startup)      │  │  │
│  │  └─────────────────────────────────────────────────────┘  │  │
│  │                                                           │  │
│  │  Agent (Claude Code CLI) runs HERE:                       │  │
│  │  - Makes API calls to Anthropic                           │  │
│  │  - Executes Bash commands (with working imports!)         │  │
│  │  - Reads/writes files                                     │  │
│  │  - Generates patches                                      │  │
│  │                                                           │  │
│  │  Evaluation runs HERE:                                    │  │
│  │  - Applies patch via git                                  │  │
│  │  - Runs pytest with task's test suite                     │  │
│  └───────────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘

Troubleshooting

FAQ - Quick answers to common questions

Full troubleshooting guide - Detailed solutions to common issues

Docker Issues

Ensure Docker is running:

docker info

Pre-built Image Not Found

If the harness can't pull a pre-built image for a task, it will fall back to building from scratch. You can also manually pull images:

docker pull ghcr.io/epoch-research/swe-bench.eval.x86_64.astropy__astropy-12907

Slow on Apple Silicon

On ARM64 Macs, x86_64 Docker images run via emulation which is slower. This is normal. If you're experiencing issues, ensure you have Rosetta 2 installed:

softwareupdate --install-rosetta

MCP Server Not Starting

Test your MCP server independently:

npx -y @modelcontextprotocol/server-filesystem /tmp/test

API Key Issues

Ensure your Anthropic API key is set:

export ANTHROPIC_API_KEY="sk-ant-..."

Timeout Issues

Increase the timeout in your config:

timeout_seconds: 600

Claude CLI Not Found

Ensure the Claude Code CLI is installed and in your PATH:

which claude  # Should return the path to the CLI

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run unit tests
pytest -m "not integration"

# Run integration tests (requires API keys and Docker)
pytest -m integration

# Run all tests
pytest

# Lint
ruff check src/

Creating Releases

We use an automated workflow for releases. See the Release Guide for full details.

Quick start for maintainers:

# Patch release (bug fixes) - most common
gh workflow run release.yml -f version_bump=patch

# Minor release (new features)
gh workflow run release.yml -f version_bump=minor

# Major release (breaking changes)
gh workflow run release.yml -f version_bump=major

For AI agents: See the AI Agent Guide for a quick reference.

The workflow automatically:

• Bumps version in pyproject.toml
• Syncs version to all package files
• Creates git tag and GitHub release
• Triggers PyPI and npm publication

Roadmap

We're building the defacto standard for MCP server benchmarking! Our v1.0 Roadmap includes 200+ features across 11 strategic categories:

🎯 Good First Issues | 🙋 Help Wanted | 📋 View Roadmap

Roadmap Highlights

Phase 1: Foundation (v0.3.0)

• ✅ JUnit XML output format for CI/CD integration
• CSV, YAML, XML output formats
• Config validation and templates
• Results persistence and recovery
• Cost analysis in reports

Phase 2: Benchmarks (v0.4.0)

• ✅ 30+ benchmarks across 10 categories
• ✅ Custom benchmark YAML support
• ✅ Custom metrics, failure analysis, sampling strategies
• ✅ Dataset versioning, latency metrics, GPU support, few-shot learning

Phase 3: Developer Experience (v0.5.0)

• Real-time dashboard
• Interactive config wizard
• Shell completion
• Pre-flight checks

Phase 4: Platform Expansion (v0.6.0)

• NPM package
• GitHub Action for CI/CD
• Homebrew formula
• Official Docker image

Phase 5: MCP Testing Suite (v1.0.0)

• Tool coverage analysis
• Performance profiling
• Error rate monitoring
• Security scanning

Get Involved

We welcome contributions! Check out our 30+ good first issues perfect for newcomers:

• Output Formats: CSV/YAML/XML export
• Configuration: Validation, templates, shell completion
• Platform: Homebrew formula, Conda package
• Documentation: Best practices, examples, guides

See the contributing guide to get started!

Best Practices

New to mcpbr or want to optimize your workflow? Check out the Best Practices Guide for:

• Benchmark selection guidelines
• MCP server configuration tips
• Performance optimization strategies
• Cost management techniques
• CI/CD integration patterns
• Debugging workflows
• Common pitfalls to avoid

Contributing

Please see CONTRIBUTING.md or the contributing guide for guidelines on how to contribute.

All contributors are expected to follow our Community Guidelines.

License

MIT - see LICENSE for details.

Built by Grey Newell | Why I Built mcpbr | About