Question 1

Are context savings exact?

Accepted Answer

No. Savings are estimated per response. They depend on the query, the repository, and which tools are invoked. We report estimates so you can reason about them, not as guarantees.

Question 2

Will CRG always use fewer tokens than reading files directly?

Accepted Answer

No. For small single-file changes in tiny packages, graph metadata can exceed the raw file size. CRG pays off on multi-file changes, larger repositories, and architecture-level questions.

Question 3

How accurate is impact analysis?

Accepted Answer

Impact analysis is conservative by design — it favours recall, so it can return false positives. In the current evaluation sample we see 100% recall and an average F1 of 0.54.

Question 4

Does flow detection work for every language?

Accepted Answer

Flow detection is strongest in Python and framework-heavy cases today and continues to improve across other languages. Structural queries (call graphs, blast radius, communities) work broadly.

Question 5

Do I have to use cloud embeddings?

Accepted Answer

No. Cloud embeddings (OpenAI, Google, MiniMax) are opt-in. Hybrid search works with local BM25 alone, and you can enable local sentence-transformers embeddings without sending code off your machine.

Question 6

Is CRG only for Claude Code?

Accepted Answer

No. CRG is an MCP and CLI tool that works with any MCP-compatible AI coding tool, including Codex, Claude Code, Cursor, Windsurf, Zed, Continue, OpenCode, Antigravity, Gemini CLI, Qwen, Kiro, Qoder, GitHub Copilot, and GitHub Copilot CLI.