Environment variables

CooperBench uses environment variables for API keys, configuration paths, and runtime settings.

LLM provider API keys

At least one LLM provider API key is required.

OpenAI

export OPENAI_API_KEY=sk-...

Required for OpenAI models:

gpt-4o
gpt-4o-mini
gpt-4-turbo
o1, o1-mini

Anthropic

export ANTHROPIC_API_KEY=sk-ant-...

Required for Anthropic models:

claude-3-5-sonnet-20241022
anthropic/claude-opus-4-20250514
claude-3-opus-20240229

Google Gemini

export GEMINI_API_KEY=AIza...

Required for Gemini models via API:

gemini/gemini-pro
gemini/gemini-1.5-flash
gemini/gemini-1.5-pro

For Vertex AI models (e.g., vertex_ai/gemini-3-flash-preview), use GCP authentication instead.

Google Cloud Platform

Authentication

Option 1: Application Default Credentials (recommended)

gcloud auth application-default login

This creates ~/.config/gcloud/application_default_credentials.json. Option 2: Service account key

export GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account-key.json

Project configuration

export GOOGLE_CLOUD_PROJECT=my-project-123

Optional. Overrides project configured with cooperbench config gcp.

Redis (for cooperative mode)

export REDIS_URL=redis://localhost:6379

Required for multi-agent communication in cooperative mode. Default: redis://localhost:6379. Override with --redis flag:

cooperbench run --redis redis://my-server:6379

Agent configuration

mini_swe_agent

# Config directory
export MSWEA_CONFIG_DIR=./configs

# Global config directory
export MSWEA_GLOBAL_CONFIG_DIR=~/.config/mini-swe-agent

# Model settings
export MSWEA_MODEL_NAME=gpt-4o
export MSWEA_MODEL_API_KEY=sk-...
export MSWEA_MODEL_API_BASE=https://custom-endpoint.com/v1

# Retry settings
export MSWEA_MODEL_RETRY_STOP_AFTER_ATTEMPT=10

# Cost tracking
export MSWEA_COST_TRACKING=ignore_errors  # or "default"
export MSWEA_GLOBAL_COST_LIMIT=10.0  # USD
export MSWEA_GLOBAL_CALL_LIMIT=1000  # max API calls

# LiteLLM registry
export LITELLM_MODEL_REGISTRY_PATH=/path/to/registry.json

# Startup behavior
export MSWEA_SILENT_STARTUP=1  # Suppress startup messages

# Docker executable
export MSWEA_DOCKER_EXECUTABLE=docker  # or "podman"

swe_agent

# Config directories
export SWE_AGENT_CONFIG_ROOT=/path/to/configs
export SWE_AGENT_CONFIG_DIR=/path/to/configs
export SWE_AGENT_TOOLS_DIR=/path/to/tools
export SWE_AGENT_TRAJECTORY_DIR=/path/to/trajectories

# GitHub token (for fetching issues)
export GITHUB_TOKEN=ghp_...

openhands

# Logging
export LOG_LEVEL=INFO  # DEBUG, INFO, WARNING, ERROR
export LOG_TO_FILE=true
export LOG_DIR=logs
export LOG_ROTATE_WHEN=midnight
export LOG_BACKUP_COUNT=7
export LOG_JSON=false
export LOG_RICH_TRACEBACKS=true

# CI mode
export CI=true  # Suppress interactive prompts

# Debugging
export DEBUG_LLM=true  # Log all LLM calls

# VNC (for browser use)
export OH_ENABLE_VNC=true

External agents

export COOPERBENCH_EXTERNAL_AGENTS=/path/to/agents.py

Path to Python module defining custom agents. See Custom agents for details.

Logging

# Suppress LiteLLM logs
export LITELLM_LOG=ERROR  # or "DEBUG", "INFO", "WARNING"

CooperBench sets LITELLM_LOG=ERROR by default to reduce noise.

HuggingFace (optional)

export HF_TOKEN=hf_...

Optional. Used for uploading results to HuggingFace Hub.

.env file

Create a .env file in your project root:

.env

# LLM Provider API Keys (at least one required)
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GEMINI_API_KEY=AIza...

# Redis (for cooperative mode)
REDIS_URL=redis://localhost:6379

# GCP (optional)
GOOGLE_CLOUD_PROJECT=my-project-123
GOOGLE_APPLICATION_CREDENTIALS=/path/to/key.json

# HuggingFace (optional)
HF_TOKEN=hf_...

# GitHub (optional, for swe_agent)
GITHUB_TOKEN=ghp_...

# Agent settings
MSWEA_GLOBAL_COST_LIMIT=10.0
MSWEA_GLOBAL_CALL_LIMIT=1000

Load with:

source .env

Or use a tool like direnv:

echo 'dotenv' > .envrc
direnv allow

Example .env file

CooperBench provides .env.example:

cp .env.example .env

Then edit .env and fill in your API keys:

.env.example

# CooperBench Environment Variables
# Copy this file to .env and fill in your API keys

# LLM Provider API Keys (at least one required)
OPENAI_API_KEY=
ANTHROPIC_API_KEY=
GEMINI_API_KEY=

# HuggingFace (optional, for saving results)
HF_TOKEN=

GCP-specific variables

When using --backend gcp:

# Project (required)
export GOOGLE_CLOUD_PROJECT=my-project-123

# Authentication (one of these)
export GOOGLE_APPLICATION_CREDENTIALS=/path/to/key.json
# OR use: gcloud auth application-default login

# Region/zone (optional, overrides cooperbench config)
export GCP_REGION=us-central1
export GCP_ZONE=us-central1-a

# Storage bucket (optional)
export GCP_BUCKET=cooperbench-eval-my-project-123

Precedence

Environment variables override configuration files:

Command-line flags (highest priority)
Environment variables
Configuration files (~/.config/cooperbench/config.json)
Built-in defaults (lowest priority)

Example:

# Config file says: redis://localhost:6379
# Environment variable overrides:
export REDIS_URL=redis://production:6379

# Command-line flag overrides both:
cooperbench run --redis redis://staging:6379

Validation

CooperBench validates required variables on startup:

cooperbench run

If API keys are missing:

Error: No LLM API key found. Please set one of:
  - OPENAI_API_KEY
  - ANTHROPIC_API_KEY
  - GEMINI_API_KEY

Security best practices

Never commit .env files:
.gitignore
```
.env
.env.local
```
Use restrictive permissions:
```
chmod 600 .env
```
Rotate API keys regularly: Generate new keys periodically and revoke old ones.
Use service accounts for GCP: Avoid using personal credentials in production.
Limit key scopes: Create API keys with minimal required permissions.

Troubleshooting

”No API key found"

# Check if variable is set
echo $OPENAI_API_KEY

# If empty, set it
export OPENAI_API_KEY=sk-...

# Verify
echo $OPENAI_API_KEY

"GCP authentication failed"

# Re-authenticate
gcloud auth application-default login

# Or check service account key path
echo $GOOGLE_APPLICATION_CREDENTIALS
ls -la $GOOGLE_APPLICATION_CREDENTIALS

"Redis connection refused”

# Check Redis is running
redis-cli ping

# If not, start Redis
redis-server

# Or update Redis URL
export REDIS_URL=redis://correct-host:6379

Variables not loading from .env

# Make sure to source the file
source .env

# Or use direnv
echo 'dotenv' > .envrc
direnv allow

Reference table

Variable	Purpose	Required	Default
`OPENAI_API_KEY`	OpenAI API key	One of 3 LLM keys	-
`ANTHROPIC_API_KEY`	Anthropic API key	One of 3 LLM keys	-
`GEMINI_API_KEY`	Gemini API key	One of 3 LLM keys	-
`REDIS_URL`	Redis connection URL	No	`redis://localhost:6379`
`GOOGLE_CLOUD_PROJECT`	GCP project ID	For GCP backend	-
`GOOGLE_APPLICATION_CREDENTIALS`	GCP service account key path	For GCP backend	-
`HF_TOKEN`	HuggingFace token	No	-
`GITHUB_TOKEN`	GitHub personal access token	No	-
`MSWEA_GLOBAL_COST_LIMIT`	Max cost in USD	No	`0` (unlimited)
`MSWEA_GLOBAL_CALL_LIMIT`	Max LLM API calls	No	`0` (unlimited)
`LITELLM_LOG`	LiteLLM log level	No	`ERROR`
`LOG_LEVEL`	OpenHands log level	No	`INFO`

Commands

Configuration

​LLM provider API keys

​OpenAI

​Anthropic

​Google Gemini

​Google Cloud Platform

​Authentication

​Project configuration

​Redis (for cooperative mode)

​Agent configuration

​mini_swe_agent

​swe_agent

​openhands

​External agents

​Logging

​HuggingFace (optional)

​.env file

​Example .env file

​GCP-specific variables

​Precedence

​Validation

​Security best practices

​Troubleshooting

​”No API key found"

​"GCP authentication failed"

​"Redis connection refused”

​Variables not loading from .env

​Reference table