Nace announced the general availability of Drex for Search on October 26, 2026, giving coding agents a tool for locating relevant files, pages and spreadsheet regions before passing their contents to an LLM. The company says the product is available through a CLI, Python and TypeScript SDKs, an MCP server and a REST API. Its homepage dates the announcement, while the release post details the product and benchmarks.
Across six professional-work test sets containing 529 questions, Nace reports 56% fewer input tokens, 62% fewer steps and accuracy rising from 63.1% to 67.1% when coding agents used Drex for Search.
The useful distinction is where Drex intervenes. It aims to decide which parts of a document collection deserve attention before the answering model starts reading. That could reduce expensive browsing loops, but developers should distinguish the search product from Nace’s general-purpose Drex decision API when setting up a trial.
How Drex Narrows a Document Search
Nace describes a three-stage process: classify, squeeze and answer.
- Classify the question. Drex determines whether the task is a quick lookup or needs more involved document reading and table queries.
- Narrow the workspace. It moves through folders, files, sections and pages. For spreadsheets, it can narrow further to tables and columns.
- Pass selected material to the answering LLM. The agent reads the retained pages and tables, then returns an answer with file and page references.
The announcement illustrates this with a question about Q3 journal entries above $50,000 that lack a second approver. Drex identifies a table-oriented task, selects the ledger from neighboring documents, and focuses on relevant rows and the approver column. That is Nace’s example of the intended workflow, not a separately demonstrated result. Nace’s explanation positions Drex as a way to direct BM25 keyword retrieval and semantic search, so those methods search a narrower area.
I’d treat that positioning as more useful than a broad claim about replacing retrieval-augmented generation. Search systems already combine ranking, filtering and document structure. The consequential question is whether Drex’s decisions remove irrelevant material without excluding the evidence needed for the answer.
Early filtering creates that tradeoff. Selecting the right ledger saves reading; selecting the wrong ledger can prevent the answering model from seeing the correct entry. A meaningful evaluation therefore needs to inspect missing evidence, as well as token savings.
Nace also says the hierarchy has a fixed number of levels, allowing a million files to take the same number of narrowing steps as a thousand. That describes decision stages, not proof of constant end-to-end latency. Ingestion, candidate evaluation, storage access and the amount of retained material remain separate considerations.
Vendor Benchmarks Show Fewer Turns, With One Accuracy Drop
Nace’s aggregate evaluation covers OfficeQA, annual and quarterly reports, audit working documents, ledgers and trial balances, invoice and billing files, and accounts receivable and payable files.
All performance figures here are vendor-reported. The announcement notes different run counts and dates between baseline and Drex comparisons. The opened evidence contains no independent replication, so the results are best read as an indication of potential gains.
The aggregate accuracy change from 63.1% to 67.1% is four percentage points, or about a 6.3% relative improvement. That distinction matters because the announcement’s shorthand of 6% more correct answers could otherwise be mistaken for a six-point gain.
Nace also publishes results for OfficeQA Pro, described as 133 questions over 697 scanned PDFs. These entries show how the outcome varies by agent and model:
| Agent and model | Accuracy: default → Drex | Change | Turns per question: default → Drex |
|---|---|---|---|
| Gemini CLI / Gemini 3.1 Pro | 54.9% → 61.7% | +6.8 points | 22.9 → 10.4 |
| Hermes Agent / Qwen3.8-27B | 56.4% → 60.2% | +3.8 points | 20.1 → 6.5 |
| pi / Opus 5.5 | 65.4% → 68.4% | +3.0 points | 7.7 → 3.7 |
| Codex / GPT-6 sol | 66.2% → 70.7% | +4.5 points | 14.3 → 8.6 |
| Hermes Agent / Opus 5.5 | 64.3% → 68.4% | +4.1 points | 8.5 → 3.7 |
| OpenCode / Qwen3.8-27B | 51.1% → 53.4% | +2.3 points | 23.7 → 10.0 |
| pi / Qwen3.8-27B | 43.6% → 45.9% | +2.3 points | 24.3 → 11.1 |
| Claude Code / Opus 5.5 | 66.5% → 69.9% | +3.4 points | 18.5 → 9.1 |
| Codex / GPT-6 luna | 54.9% → 53.4% | −1.5 points | 16.8 → 11.3 |
Source: Nace’s OfficeQA Pro results. Nace lists maximum effort for the Codex and Claude Code configurations, and medium effort for the other configurations shown.
Every row above uses fewer turns with Drex. Accuracy is less uniform: Codex with GPT-6 luna loses 1.5 points despite reducing turns from 16.8 to 11.3.
That exception is important. Fewer tool interactions can indicate a more efficient route to the evidence, but they are not themselves evidence of a better answer. My reading is that reduced search activity is the more consistent result in this table; accuracy needs checking for the particular agent configuration.
Nace’s timing table also reports an average of per-agent median completion times falling from 80.8 seconds to 35.5 seconds across pi, Hermes Agent and Codex. The improvement varies by agent, so the aggregate should not be treated as a latency guarantee.
Nor does 56% fewer input tokens automatically mean a 56% lower bill. Total cost depends on the answering model, Drex usage, document processing and other infrastructure.
Connecting a Coding Agent: Know What the MCP Server Exposes
Nace’s hosted MCP documentation provides a low-friction way to try the underlying decision model. The server uses streamable HTTP at:
https://console.nace.ai/mcp
For Claude Code, the documented command is:
claude mcp add --transport http drex https://console.nace.ai/mcp
This registers the remote server in the current project by default. Adding --scope user makes it available across projects. The documentation says to use /mcp inside Claude Code to inspect the connection.
For Codex, the equivalent registration command is:
codex mcp add drex --url https://console.nace.ai/mcp
Nace says this writes the server configuration to ~/.codex/config.toml, which the Codex CLI, IDE extension and desktop app share. These are documented setup commands, not commands executed for this article.
The important boundary is the tool list. This hosted MCP page documents decide and list_models. It does not describe a workspace-search tool or a document-ingestion procedure. Registering this URL gives an agent access to Drex decisions; it should not be presented as completing a Drex for Search deployment over a folder of files.
That distinction keeps a trial honest. A successful decision call checks model connectivity. A retrieval trial must also establish how documents enter the workspace and how the agent invokes the search workflow.
Free Access Has Per-Network Limits
The hosted decision server accepts calls without an API key under these limits:
| Limit | Allowance |
|---|---|
| Calls per minute | 10 |
| Calls per day | 100 |
| Input per call | 32 KB |
Nace counts the budget by the network address making the request. People sharing an office network may therefore share the allowance. The documentation also permits refusal during heavy demand or when the free tier is disabled.
Authenticated calls run under the account’s limits and billing. Keep API keys out of committed project configuration; Nace documents environment-variable options for supported agents, alongside setup instructions for Cursor, OpenCode, Pi and Hermes Agent.
Frequently Asked Questions
3 questions
1Is Drex for Search Generally Available?
Yes. Nace announced general availability on October 26, 2026, through CLI, SDK, MCP and REST integration paths.
