Try the demo

Use DMN

Make your limits last

Most of an agent's tokens go to reading code to find the right spot. DMN hands it the spot instead, and a few defaults trim the rest. Here is what each part does, what we measured, and how to check your own numbers.

What saves tokens

Part What it does On by default
Search that returns the code One call returns the sections that answer a question, so the agent doesn't read file after file Once an agent is connected
The skill instead of MCP Agents use DMN through their shell; no tool definitions are sent with every request When you run dmn skill-init
Short answers The briefing asks Claude Code to answer tersely Yes, for Claude Code
The proxy Compresses context that requests keep repeating Yes, for Claude Code agents started in the app

What we measured

Tokens used, lower is better

Without DMN1.00×
DMN search, as a skill0.85×0.71–1.01
Skill + short answers (as shipped)0.65×0.53–0.80
Tokens to find the code behind a real bug, against Claude Code without DMN (1.00×). DMN's own code, 14 tasks run twice, ranges are 95% intervals. Details on the benchmarks page.

On real bug reports in DMN's own code, Claude Code found the right code with fewer tokens with DMN than without it: 0.85× with the search skill alone, and 0.65× with short answers on as well, which is how DMN ships. No run gave a plain agent the short-answer rule without DMN, so part of that 0.65× may come from the rule alone. On a public code base the model couldn't place from memory, agents made 11% fewer requests, but the token saving was too small to measure. On a popular project the model often knew where to look without any tools, and DMN made no clear difference.

These runs measured finding the code, which is only part of a session: agents also write and test code, where DMN changes less. Session-wide, the saving is likely closer to 5–10%. The benchmarks page has the method, every result and where DMN doesn't help.

See your own savings

dmn stats                 # this project, the last 7 days
dmn stats --days 30       # a longer window
dmn stats --all-projects  # every project together

dmn stats compares the code DMN sent your agents with what reading the same files whole would have cost, per agent and per day. It counts bytes, so it needs no model or price. The app shows the same figures in Insights, with the billed token counts of the requests that went through the proxy.

dmn bench builds a search benchmark from your own code and writes a report you can rerun: the same index and seed give the same numbers.

Turn a part off

  • Short answers: add [skill] with compact = "off" to %APPDATA%\dmn\config.toml. A blinded check rated the short answers about as useful as full ones, though they sometimes leave out a secondary detail such as a suggested fix.
  • The proxy: Settings ▸ Context tiering, then restart the app.
  • The skill: dmn skill-init --remove --apply.