Configuration

Radar reads configuration from environment variables. The only required value is a worker model API key.

Required

export RADAR_API_KEY=sk-...

If you use Moonshot, Radar also accepts MOONSHOT_API_KEY as a fallback.

Provider Examples

Moonshot / Kimi:

export RADAR_BASE_URL=https://api.moonshot.ai/v1
export RADAR_MODEL=kimi-k2.5

DeepSeek:

export RADAR_BASE_URL=https://api.deepseek.com/v1
export RADAR_MODEL=deepseek-chat

Together:

export RADAR_BASE_URL=https://api.together.xyz/v1
export RADAR_MODEL=Qwen/Qwen2.5-72B-Instruct-Turbo

Local Ollama:

export RADAR_BASE_URL=http://localhost:11434/v1
export RADAR_MODEL=qwen2.5-coder:32b

Limits

Radar enforces input limits before calling the worker model.

Variable	Default	Purpose
RADAR_MAX_READ	8192	Max completion tokens for ask, search, and summarize
RADAR_MAX_WRITE	16384	Max completion tokens for write
RADAR_MAX_FILES	30	Max files for ask
RADAR_MAX_SEARCH_MATCHES	20	Max search results Radar reads
RADAR_MAX_BYTES_PER_FILE	500000	Max bytes read from one file
RADAR_MAX_CORPUS_BYTES	2000000	Max bytes sent to one worker call
RADAR_TIMEOUT_MS	30000	OpenAI-compatible request timeout

Verbose Diagnostics

Set RADAR_VERBOSE=1 to print delegation metrics:

export RADAR_VERBOSE=1

Verbose output includes file counts, input bytes, corpus-build time, API time, total time, and token usage when the provider returns it.