Environment Variables

LLM Backend

The active backend is resolved in this order:

1. Explicit preference saved by cxp install --setup (stored in keychain + local file)

2. Environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, NVIDIA_API_KEY)

3. Keychain / file cache from a previous cxp install --setup run

4. Claude Code CLI fallback — if none of the above, and claude is in PATH

The recommended way to configure a backend is cxp install --setup, which presents the options as distinct choices with clear trade-offs. See Claude Code setup and Cursor setup.

Claude Code (subscription, via claude -p) and Anthropic API (direct HTTP, billed per token) are separate options — the former is free but slower; the latter is faster and parallelizes better.

Backend environment variables

Variable	Backend
ANTHROPIC_API_KEY	Anthropic Messages API
OPENAI_API_KEY	OpenAI chat completions
NVIDIA_API_KEY	NVIDIA NIM (OpenAI-compatible)

If set, env vars override the saved preference for that process.

Reset a saved NVIDIA key:

cxp --reset-nvidia-api-key

Reset Anthropic or OpenAI keys:

cxp install --setup    # re-run the wizard and pick a new backend

---

Team Sync

Variable	Default	Description
CXP_API_KEY	—	Team API key. Alternative to cxp auth — useful in CI or containers.
CXP_API_URL	https://contextpool-server-nj1f.onrender.com	Override the API endpoint. Use for self-hosted servers or local dev.

---

Performance Tuning

Variable	Default	Description
CXP_CONCURRENCY	8	Max concurrent LLM calls during fetch_project_context and cxp init. Increase for faster bulk processing; decrease if you hit rate limits.
CXP_MAX_CHARS	20000	Transcript characters sent to the LLM per session. Lowering this (e.g. 8000) speeds up each call significantly with minimal quality loss.

Example — tune for speed on a large project:

CXP_CONCURRENCY=12 CXP_MAX_CHARS=8000 cxp init claude-code --local

---

LLM Tuning

Variable	Default	Description
MODEL	per-backend default¹	Override the model used for summarization.
REPAIR_MODEL	same as MODEL	Model used for JSON repair if the first response is malformed.
TEMPERATURE	0.0	Sampling temperature.
TOP_P	0.95	Top-p nucleus sampling.
MAX_COMPLETION_TOKENS	4096	Max tokens per LLM response.

¹ Default models:

• Claude Code CLI / Anthropic: claude-haiku-4-5-20251001

• OpenAI: gpt-4o-mini

• NVIDIA NIM: qwen/qwen3.5-122b-a10b

---

Extraction Behavior

Variable	Default	Description
SANITIZE_CHAT	1	Strip tool calls, thinking blocks, file snapshots, and injected context before sending to the LLM. Set to 0 to disable.
EXTRACT_USER_QUERIES_ONLY	0	Send only user messages to the LLM, not assistant replies. Set to 1 to enable.

---

Debugging

Variable	Default	Description
DEBUG_LLM_OUTPUT	0	Print the raw LLM response to stderr before parsing. Set to 1 to enable. Useful for diagnosing extraction failures.

---

In MCP Config (Cursor)

If you prefer not to use the keychain, pass env vars directly in ~/.cursor/mcp.json. This is also useful in headless or container environments:

{
  "mcpServers": {
    "contextpool": {
      "command": "cxp",
      "args": ["mcp"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "CXP_API_KEY": "cxp_team_...",
        "CXP_CONCURRENCY": "10"
      }
    }
  }
}