Deep32: retrieval-augmented question answering for network and security engineering

Deep32 answers network and security engineering questions from a local corpus of 365 RFCs and 489 Cisco documents. Claude Code or Codex does the research through eleven MCP tools: it searches the corpus, reads whole sections, looks up RFC status and checks each quotation against the stored text. Two more collections come with it and are searched only when you ask for them: 241 PowerShell documents and 85 about Deep32 itself.

The corpus, the index and the search engine run on your own machines. The language model is hosted, so the question, the queries the agent writes and the passages it gets back go to the model provider.

This release is an alpha. It assumes you're comfortable installing an operating system in a VM and running Docker.

0.1.0-alpha.2 · 20 November 2026 Code: Apache-2.0 DOI 10.5281/zenodo.24272647 OpenSSF Best Practices OpenSSF Scorecard

At a glance

At a glance: hit@1 0.955; hit@8 0.991; 854 documents (365 RFCs, 489 Cisco); 11 MCP tools for Claude Code and Codex; median search 542 ms with hybrid search and reranking.
Retrieval on 696 labeled questions, measured on my development setup through the gateway, 2 November 2026. The evaluation has the details.

The three parts

A deployment has three parts. They can share one machine or sit on separate ones, at home or across sites. They talk over your own Tailscale network, your tailnet, and each pair of parts is introduced once with a short code.

Workstation: search and archive. OpenBSD with the gateway and the Library, from disc 1.

Workstation

An OpenBSD VM or machine that holds the corpus and the index, with the search engine, the MCP gateway, the Ask window and the Deep32 Library. It installs from disc 1 and imports its content from disc 2.

Forge: the GPU powerhouse. Embed, rerank and convert, from disc 3.

Forge

The machine with the NVIDIA card. It runs the embedding and reranking models in Docker, behind one authenticated door. Without a Forge, search works on keywords only.

Headend: your agent, Claude Code or Codex, on your PC.

Agent PC

The PC where Claude Code or Codex runs. Connect pairs it with the Workstation and installs the research skill. The rest of the documentation calls it the headend.

Deployment topology: the headend, the Workstation and the Forge on your tailnet. The headend's agent reaches the Workstation's MCP gateway over HTTPS with an agent token; the Workstation calls the Forge's models over HTTPS with a bearer token; the parts are paired once with a short code. Only the hosted language model is outside, and it receives the question and the retrieved passages.
Figure 1. The three parts of an install. They can all run on one machine or on separate ones.

The Workstation

Your agent can use Deep32 over MCP while it works, or you can type a question in the Workstation's Ask window and have the agent on your PC answer it there.

The Deep32 Workstation desktop. The Ask window shows a cited answer about BGP keepalive and hold timers; a network calculator, a text editor and the desktop style menu are open beside it.
Figure 2. The Workstation on my laptop, in the HP-UX CDE look the release installs. Ask is answering the BGP timer question, with the network calculator, a text editor and the desktop style menu open. Panels that exist only in my private build are blanked.

Read more

Support

Questions go to Discussions and bugs to Issues. If you like Deep32 and want to support it, everything donated goes to costs for this project.