Experience-Driven Black-Box Cost Control for Coding Agents
Python agentic-aiagentic-codingai-codingai-coding-assistantanthropicartificial-intelligenceclaude-codecodexcoding-agentcoding-agentscost-optimizationcost-reductiondeveloper-toolsgithub-copilotllmopen-sourceopenaiproductivitytoken-optimization
# Copy to experiments/.env (git-ignored) and fill in. One KEY=value per line: no quotes, no
# $VAR expansion, and a `#` starts a comment only at the start of a line. A variable already
# exported in your shell wins over this file.
# --- API keys -------------------------------------------------------------------------------
# Claude Code bills the Anthropic API, Codex the OpenAI API.
ANTHROPIC_API_KEY=
OPENAI_API_KEY=
# Only for an OpenAI key issued against a regional endpoint:
# OPENAI_BASE_URL=https://us.api.openai.com/v1
# The Copilot CLI bills the Anthropic API too, through its own provider settings. It does not
# read ANTHROPIC_API_KEY, so give it the same key again here.
COPILOT_PROVIDER_TYPE=anthropic
COPILOT_PROVIDER_BASE_URL=https://api.anthropic.com
COPILOT_PROVIDER_API_KEY=
# --- Data (absolute paths; see README.md, Setup) ---------------------------------------------
# DATA holds the benchmark data, the staged rtk binary ($DATA/scaffolds) and, unless
# TE_SWEBENCH_GRADING is set, the grading file ($DATA/swebench/swebench_grading.json).
DATA=
TOKENECONOMIZER_RESULTS=
TOKENECONOMIZER_SWEBENCH_TASKS=
TE_SWEBENCH_REPO=
TOKENECONOMIZER_WEBBENCH_PROJECTS=
TOKENECONOMIZER_WEBBENCH_TASKS=