Blume Codes Review: Stop Repeating AI Rules

AI coding agents move fast, but repeated corrections get annoying even faster. That problem is the reason I focused on Blume.codes. Instead of replacing Claude Code, Codex, Cursor, or the other tools already doing the coding, Blume sits beside them and keeps track of the instructions, rules, skills, and corrections surrounding the work.

What I like most is how focused the idea is. Blume is a desktop sidecar for coding agents. Its current interface works around Claude Code, Codex, Cursor, omp, and Pi. The desktop window brings agent activity, project setup, hidden instruction files, and provider usage into one place. That immediately gives Blume a stronger purpose than a dashboard filled with decorative AI statistics.

The feature that matters most is Improve. Repeated corrections and similar friction are grouped into patterns. From there, Blume prepares a concrete change to the agent setup. The proposed rule or skill appears alongside the evidence that triggered it and an exact diff.

I remain the final decision-maker. I review the change and then apply, dismiss, or leave it for later. The setup is not silently rewritten behind my back. That review step is essential because agent instructions directly influence future coding sessions.

This solves a real weakness in AI-assisted coding. A correction buried inside one conversation does not automatically become permanent project knowledge. Files such as CLAUDE.md, agent rules, skills, hooks, and other setup files carry those instructions into future sessions.

Blume gives repeated feedback a route into that persistent setup. I see strong value here for developers who repeatedly correct naming rules, testing requirements, release steps, architecture decisions, import styles, or repository-specific habits.

The value rises when several agents touch the same repository. One forgotten instruction quickly turns into repeated cleanup work. Keeping those decisions visible and reusable gives the project a steadier operating memory across separate sessions.

I also like the visibility side. The Agents area tracks whether an agent is working, finished, or waiting for approval. Setup brings skills, rules, hooks, and agent files together. Usage tracks Claude Code and Codex plan limits and token use before the account hits its limit.

That makes Blume.codes more than a correction tracker. I see it as a compact control center for developers moving between several coding agents.

Privacy deserves special attention because the app reads coding conversations. Conversation transcripts start on the device and are stored locally in a SQLite database. The core design avoids building a permanent cloud archive of those chats. Usage and spending information also stays local rather than feeding an advertising or profiling system.

There is also a separate Cloud Improvements path. It requires an explicit choice and uses limited conversation excerpts and project context for transient processing. The privacy page currently states that this cloud route remains unavailable until its provider route and privacy controls are verified. I prefer that separation over pretending every feature follows exactly the same privacy path.

Blume is still clearly an early product. The current homepage lists Auto-Fixes, Analytics, and Local Domain Model as upcoming features. Central Domain Model, Team Conflict Resolution, and Auto-Improve Mode sit even further down the roadmap. I judge Blume by what exists today rather than awarding points for future promises.

Pricing is refreshingly simple right now. As of September 4, 2026, Blume is free. Its Terms of Service, last updated August 10, 2026, explicitly confirm that the service currently has no Blume usage fee. Costs from connected AI providers remain separate. The terms also leave the door open to paid Blume features later.

My main reservation is maturity. Several ambitious features remain unfinished, while AI-generated configuration changes still demand careful human review. A proposed hook deserves even more attention because hooks involve automatically executed commands. I would never approve that type of modification blindly.

Blume.codes gets the core idea right: stop wasting useful corrections inside forgotten AI chats. It keeps coding agents visible, organizes the instructions controlling them, and turns repeated friction into reusable context. For developers already switching between Claude Code, Codex, and Cursor, I see a genuinely useful companion rather than another AI tool fighting to become the center of the workflow.