ColGrep-Agent-2B: hand your coding agent's code search to a local 2B model and save a third of the tokens
What is it?
Before a coding agent can fix a bug, it spends turns working out where to look: grep, open a file, follow a reference, repeat. ColGrep-Agent is a 2-billion-parameter search agent that does that part on your machine. You ask it a question; it runs LightOn's semantic code search ColGrep, reads the likely files, follows references, and hands back file paths, line ranges and the code itself. Install it, then run the one line that matches your coding agent:
brew install lightonai/tap/colgrep
colgrep --install-claude-code # Claude Code
colgrep --install-codex # Codex
colgrep --install-opencode # OpenCodeYou don't need an API key or a separate inference server.
Why does it matter?
LightOn plugged it into Claude Code and ran 52 questions across 13 repos the model hadn't seen. Compared with plain grep, API tokens dropped 34% and cache-aware API cost dropped 40%, with no statistically significant change in accuracyVendor claim. If agent bills hurt, it's a cheap experiment: let the frontier model do the reasoning and the small local model do the digging.
Maturity
BetaIt's an early Apache-2.0 release. The savings numbers come only from LightOn's own 52-question eval. On an M3 Pro each query takes about 20 seconds and 3–4 GB of memory, which can be slower than just running grep.
How to track it
- New versions: watch Releases on GitHub
lightonai/next-plaid(colgrep version bumps). - Does it replicate: look for evaluations or user reports outside LightOn that show similar token savings.
- Integration health: keep an eye on GitHub issues about the Codex and OpenCode integrations.