Try the demo

Agents · Checked on 2026-10-03

DMN for Claude Code

A new Claude Code session knows your CLAUDE.md and nothing else: what the last session found is gone. DMN keeps it. Run Claude Code in a DMN pane, or keep your terminal and connect it with two commands.

What Claude Code gets

  • Notes that outlive the session. Each new session, and each one after a compact, starts with your rules, then the newest notes, each with its age.
  • Fixes it already found. When a command fails with an error earlier sessions fixed, Claude Code is told the fix in one line.
  • Code search in one call. The matching code comes back, not a list of files to read.
  • Claims on files. With several sessions on one checkout, each sees which files the others are editing.
  • A fresh index. Every file it edits is re-indexed straight away.

What stays the same

  • Your Claude plan. DMN doesn't sell AI and needs no account.
  • Your CLAUDE.md. DMN's notes sit beside it.
  • Claude Code's own tools. It can still grep and read files; the engine is one more option.
  • Your code stays on your machine. The engine only listens on 127.0.0.1.

Set it up

In DMN, start Claude Code from + agent in the Fleet panel, in its terminal or in DMN's chat pane. The app can set up the rest. For Claude Code in your own terminal:

  1. Install DMN

    From the download page. No account.

  2. Add the skill

    Writes ~/.claude/skills/dmn/SKILL.md, which teaches Claude Code the dmn command line. It costs one line of context until it's used. 

    dmn skill-init --apply --agent claude

  3. Add the hooks

    Adds DMN's hooks to your project's .claude/settings.json (--global for every project): the briefing at session start, the re-index after each edit, and the one-line fix when a command fails. 

    dmn hook-init --apply

Restart Claude Code afterwards. Leave out --apply to see what a command would change first. Prefer MCP tools to a skill? dmn mcp-init --apply --agent claude adds DMN's MCP server instead. To check or undo any of it, see Connect your agents.

Several sessions on one project

Before a session edits, it claims the files it's about to change. Another session checks before touching a path, sees who holds it, and works around it. Claims are advisory: they never block a write, and they expire on their own. For full isolation, DMN's spawn sheet can start a session in its own git worktree instead. Without DMN, see a worktree and a tab for each agent.

Questions

Does this replace CLAUDE.md?

No. CLAUDE.md is what you write down for Claude. DMN's notes are what agents learn as they work, saved in .dmn/memories, and Codex, Gemini CLI and Cursor can read them too. How both compare with auto memory: Claude Code's memory between sessions.

Will it use fewer tokens?

It depends on the code. Tracking down real bugs in DMN's own code, Claude Code needed fewer requests and fewer tokens with DMN. On public code it needed fewer requests, with no measurable token saving. The benchmarks have the method and every result, including where DMN doesn't help.

How does it compare with Claude Code's desktop app?

See DMN and Claude Code. The two fit together.

See it with sample data

▶ Try the demo Other agents