- Nace AI released Drex for Search, a retrieval layer that lets AI coding agents like Claude Code, Codex and Cursor search documents via CLI, SDK, MCP or REST.
- The company reports 92.05% accuracy on the BrowseComp-Plus benchmark using GPT-5, at an average of 12.6 search calls per question.
- On the OfficeQA Pro benchmark, Nace says Drex for Search raised accuracy by up to 6 percentage points while cutting agent turns by 62% and tokens by 56%.
- Drex narrows searches step by step — folder, file, section, page — rather than returning top-k similarity matches, and every answer cites its source file and page.
What Happened
Nace AI released Drex for Search, a retrieval tool built to let AI agents search over more than a million files, MarkTechPost reported (Oct 26, 2026). The company describes it as a hosted system that pairs with answering models such as GPT-5, Gemini 3.5 Flash or Qwen 27B.
Why It Matters
Reliable document retrieval is a known bottleneck for coding and research agents, a problem also targeted by tools like ProofShot’s visual verification for coding agents and tested by benchmarks such as SentinelBench. Nace frames retrieval as a sequence of decisions rather than a single similarity search, narrowing from folder to page before an answer reaches the model.
Technical Details
Nace reports 92.05% accuracy on BrowseComp-Plus using GPT-5 at 12.6 search calls per question, and 77.5% recall-at-1 on VisDocAgentBench, which the company says is 10 points above Claude Opus 5. On its OfficeQA Pro test across nine agent-model pairs, Nace reports accuracy gains of up to 6 percentage points alongside 62% fewer agent turns and 56% fewer tokens — figures the company reports itself and that have not been independently verified.
Who’s Affected
Teams running coding agents like Claude Code, Codex, Cursor and OpenCode over large document sets — contracts, SEC filings, spreadsheets — are the direct target, since Drex integrates through existing CLI, SDK, MCP and REST interfaces. Nace offers the service on its own cloud, a dedicated cloud, customer AWS or Azure accounts, or air-gapped on-prem deployments.
What’s Next
Nace has not disclosed pricing for Drex for Search beyond its existing managed and dedicated-cloud tiers. The company says it will keep publishing results against public leaderboards including BrowseComp-Plus and VisDocAgentBench as adoption grows.
