Cursor / Windsurf

Cursor and Windsurf are AI-native code editors that support custom API endpoints. Both can route LLM requests through Candela for full observability.

Cursor

OpenAI-Compatible Models (Gemini, Ollama)

1. Open Settings → Models
2. Add a new OpenAI API model:

Setting	Value
API Base	http://localhost:1234/v1
API Key	candela
Model	gemini-3.5-pro (or any model from /v1/models)

Claude Models

Cursor has native Anthropic support. To route Claude through Candela:

1. Open Settings → Models → Anthropic
2. Set API Base URL to:

   http://localhost:8181/proxy/anthropic-direct

3. Enter your Anthropic API key

Or use the Vertex AI route (no API key needed):

http://localhost:8181/proxy/anthropic-vertex

Environment Variable (Alternative)

# Route all Cursor Anthropic requests through Candela
export ANTHROPIC_BASE_URL="http://localhost:8181/proxy/anthropic-vertex"

---

Windsurf

Windsurf supports custom OpenAI-compatible endpoints:

1. Open Settings → AI → Custom Model
2. Configure:

Setting	Value
Base URL	http://localhost:1234/v1
API Key	candela
Model	gemini-3.5-pro

---

What Gets Traced

Both editors generate many LLM calls during a single coding session — inline completions, chat, code actions, and multi-file edits. Candela captures all of them:

Feature	Traced
Chat / inline completions	✅
Multi-file edits	✅
Code explanations	✅
Terminal commands	✅

Check http://localhost:8181/_local/ to see per-request token counts, cost, and latency across your entire coding session.

Troubleshooting

Symptom	Cause	Fix
Model dropdown empty	Cursor can’t reach /v1/models	Verify candela start is running
”Invalid API key”	Cursor validates keys before use	Use candela as the key — it passes validation
High cost for completions	Many small requests add up	Check traces to identify chatty features
Traces not appearing	Wrong base URL	Use :1234 for OpenAI-compat, :8181/proxy/* for provider-specific