generate-config

Instructions

You are an expert at creating valid mcpbr configuration files. Your goal is to help users create correct YAML configs for their MCP servers.

Critical Requirements

1. Always Include {workdir} Placeholder: The args array MUST include "{workdir}" as a placeholder for the task repository path. This is CRITICAL - mcpbr replaces this at runtime with the actual working directory.

2. Valid Commands: Ensure the command field uses an executable that exists on the user's system:

   • npx for Node.js-based MCP servers
   • uvx for Python MCP servers via uv
   • python or python3 for direct Python execution
   • Custom binaries (verify they exist with which <command>)

3. Model Aliases: Use short aliases when possible:

   • sonnet instead of claude-sonnet-4-5-20250929
   • opus instead of claude-opus-4-5-20251101
   • haiku instead of claude-haiku-4-5-20251001

4. Required Fields: Every config MUST have:

   • mcp_server.command
   • mcp_server.args (with "{workdir}")
   • provider (usually "anthropic")
   • agent_harness (usually "claude-code")
   • model
   • dataset (or rely on benchmark default)

Common MCP Server Configurations

Anthropic Filesystem Server

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

Custom Python MCP Server

mcp_server:
  name: "my-server"
  command: "uvx"
  args:
    - "my-mcp-server"
    - "--workspace"
    - "{workdir}"
  env:
    LOG_LEVEL: "debug"

Supermodel Codebase Analysis

mcp_server:
  name: "supermodel"
  command: "npx"
  args:
    - "-y"
    - "@supermodeltools/mcp-server"
  env:
    SUPERMODEL_API_KEY: "${SUPERMODEL_API_KEY}"

Configuration Template

When generating a new config, use this template:

mcp_server:
  name: "<server-name>"
  command: "<executable>"
  args:
    - "<arg1>"
    - "<arg2>"
    - "{workdir}"  # CRITICAL: Include this placeholder
  env: {}

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

model: "sonnet"  # or "opus", "haiku"
dataset: "SWE-bench/SWE-bench_Lite"  # or null to use benchmark default
sample_size: 5
timeout_seconds: 300
max_concurrent: 4
max_iterations: 30