Devexanalytics for AI-readable docs
Log in
Live · capturing agent traffic on docs.tenstorrent.com

Know what coding agents are looking for
in your docs.

Devex records the questions Claude Code, Cursor, and Copilot send to your documentation — and which chunks they actually read in response.

01
force intent

Agents have to say what they want
before they can read.

Your docs URL stops handing out pages. Agents have to send a query at /q/{question} first. Their question, in plain English, is the signal you never had before.

ttnn.matmul supported memory_config combinations on wormhole
Claude Code · US · 2h ago
vllm-server rootless docker cgroupv2 mount fix
Cursor · DE · 47m ago
tt-forge AOT compile cache reuse for Mistral-7B
Copilot · IN · 12m ago
02
re-index for trace

Every read leaves a specific trace.

Each page becomes a shallow graph of small chunks — we call them atoms. An atom is one signature, one example, one warning, one paragraph. Page-level analytics says a doc got 200 hits. This tells you which specific atom 173 of them actually read, in what order, and where they stopped.

session.log · agent: Claude Code · 28s
11:42:03 query "ttnn.matmul memory_config rules"
11:42:09 fetch ttnn.matmul.signature
11:42:17 fetch ttnn.matmul.memory-support-table
11:42:24 fetch ttnn.matmul.example
11:42:31 end 28-second session
03
analytics surface

Pulled together into one place.

Themed clusters of agent intent. Per-query drill. Per-atom demand. Sessions traced end to end. Spike detection. Dead-end atoms flagged. Everything an analytics tool would do, if it knew what an "agent session" was.

queries
200
across 8 themes
themes
8
matmulinstallvllm+5
atoms
4,135
382 reached · 4 dead-end

No SDK. No plugin. No change
for developers.

The only thing we touch is your docs server. Agents already hitting your docs become the data — they don't even know they're being instrumented.

See it on real data.

Three-minute walkthrough on the live Tenstorrent docset.

View the demo
invitation-only beyond the demo · log in