Blume.codes

What is Blume.codes?

Blume is a local-first desktop sidecar that watches coding-agent sessions and turns repeated corrections into reviewable diffs on rules and skills. It is for people running Cursor, Claude Code, Codex, Amp or Pi who keep retyping the same instruction and want that instruction to land in the repo's rulebook — after they approve it.

Key features

• Desktop sidecar that sits next to Cursor, Claude Code, Codex, Amp and Pi and shows which agents are running, waiting on you, or finished
• Tracks the hidden files that shape agents: skills, hooks, rules and related project guidance, in one graph
• Reads local session files, clusters repeated corrections / steering / frustration, and proposes a rule or skill as a reviewable diff
• Approve, dismiss, or save a suggestion for later — the site's wording is that you see every change before it lands
• Local from the start: conversation history is stored on the device; the blog states analysis runs locally using your Claude Code or Codex harness and sessions/code are never sent to Blume
• Usage tracking for Claude Code and Codex plan limits and token burn (listed as a current capability on the homepage)
• Downloads for macOS, Windows and Linux; no account required
• Auto-Fixes, setup-health over time, and Auto-Improve Mode are marked Soon or Next on the homepage and are not shipped as automatic writers

Who it's for

• A developer who has corrected the same coding agent twice and wants that correction to become a project rule without hand-editing AGENTS.md
• Someone running more than one harness (Claude Code plus Codex plus Cursor) who cannot see which agent is waiting and which skills it is using
• A repo whose skills and hooks have piled up and need a reviewable maintenance pass rather than another undocumented prompt
• A privacy-sensitive workflow that cannot upload session logs to a hosted 'agent memory' product

When not to use it

Not an auto-rewriter of your rulebook: Auto-Fixes and Auto-Improve are marked Soon/Next, and the shipped loop is propose-then-approve. Not a hosted team memory layer (Central Domain Model is listed as Next). Not useful if your coding agent does not write local session files Blume can read.

FAQ