Blume Review (2026): Pricing, Features & Honest Verdict

Blume is a desktop sidecar that watches your coding agents and maintains the rules, skills and hooks that steer them. Two named founders, a privacy page that names its subprocessors, and free with no plans published. Best for: developers running Claude Code, Codex and Cursor side by side. Price: Free (free plan: yes). Rating: 7/10.

What is Blume?

Blume launched on Product Hunt on 3 September 2026 and finished eleventh with 105 upvotes. It is a desktop app that sits beside your coding agents rather than replacing them. It reads the conversation transcripts that Claude Code, Codex and Cursor already write to your machine, tracks the hidden files that shape those agents, and proposes changes when it spots you correcting the same thing over and over. The pitch is that rules, skills, hooks and corrections pile up across every harness you use, and nobody is maintaining that pile.

The company is two engineers who met at NTNU in Trondheim. Peder Aaby works from Oslo, holds a master’s degree from NTNU and is an AWS-certified solutions architect and certified application security engineer. Olav Ljosland works from Toronto and most recently helped build Wordware, the YC S24 AI workflow platform. Their careers site lists 2025 as the founding year and two open roles, Head of Growth and Founding Engineer, which is more corporate infrastructure than most launches at this stage carry. We found the team page one of the more checkable ones we have read this year: both founders link their LinkedIn profiles and both claims resolve.

What are Blume’s key features?

Agents: one window on every harness

The Agents panel shows what each coding agent is doing right now and which rules, skills and instructions are guiding that work. The sample view lists a Codex session aligning archived ticket handling, a Claude Code session separating catalog sync detection, and a Cursor session refreshing a billing settings flow, each with its branch and state. The specific value here is knowing the moment an agent finishes or stops to ask for approval, which is the thing you currently discover by tabbing through terminals.

Setup: the hidden files, finally in one list

Every skill, rule, agent file and hook across Claude Code, Cursor and Codex, collected into one place. If you have ever tried to remember whether a rule lives in the project AGENTS.md, the global CLAUDE.md, a Cursor rules file or a hook script, this is the panel that pays for the install. It is also the least glamorous feature on the site and the one most clearly shipped.

Usage: how much of each plan you have left

Blume tracks your remaining Claude, Codex and Cursor capacity before you hit a limit mid-task. The privacy detail matters more than the feature: usage and spending information is fetched directly from Anthropic, ChatGPT and Cursor using credentials you control, displayed locally in the app, and explicitly not transmitted to Blume servers for analytics. The company states it does not use your AI spending data to build profiles, ad systems or training data.

Improve: corrections become durable rules

This is the headline. Blume reviews your rules, skills and hooks and proposes a fix, with the evidence attached. The sample suggestions are specific in a way that reads like real output: “In 4 conversations, you asked agents to run the relevant package tests and typecheck before saying a change was ready”, proposing a verification section in the project agent guidance. Another spots the same versioning, release-note, signing and smoke-test sequence re-explained across three conversations and proposes turning it into a Claude Code skill, then exposing the same workflow to Codex. You preview the exact diff, then apply, dismiss, or save it for later.

What does Blume actually ship today?

Read the capability grid carefully, because it is doing more work than the hero copy. Three features are listed as shipped: Agent Overview, Hidden Files and Rules, and Usage Tracking. Three carry a “Soon” label: Auto-Fixes, Analytics, and Local Domain Model. Three more are marked “Next”: Central Domain Model, Team Conflict Resolution, and Auto-Improve Mode. So six of the nine capabilities on the page are roadmap.

That would be fine, and honestly labelled, except for one overlap. The Auto-Fixes entry marked “Soon” is described as detecting mismatches between your agent setups and your chat instructions, proposing the fix, and asking you to approve. Two sections up the same page, the Improve panel is described in the present tense as reviewing your rules, skills and hooks on-device and proposing the fix for you to preview or dismiss. Those are close to the same sentence with different tense and a different label. A reader arriving from the tagline, which promises that Blume turns agent sessions into better rules and skills, cannot tell from the page which half of that is live.

The page source settles part of it. Blume ships its feature-flag state into the HTML, and cloud-improve-processing is set to false. The privacy page says the same thing in words: “Cloud processing will remain unavailable until the deployed provider route and these privacy controls are verified.” So the cloud-assisted path for Improve is switched off, and the local path that runs through your own connected provider plans is what you get today. Several other flags are off too, including artifact-sync, delivery-lifecycle and desktop-mcp-insights-agent-workflow. The same payload shows the homepage headline and the privacy copy are both running as A/B variants, which is a normal growth practice and a reminder that the wording you read may not be the wording someone else reads.

How much does Blume cost?

Blume is free. There is no pricing page: the /pricing route returns a 404 and the homepage carries no price at all. The only committed statement about money is in the terms of service, clause 10, which reads: “The Service is currently provided free of charge; we may introduce paid features in the future on notice.”

That is a real free plan rather than a beta with a countdown, and putting it in the contract is better than putting it on a landing page. But it is also the whole story. There is no published ladder, no indication of which features would sit behind a future paid tier, and no free-forever commitment for the current feature set. If you are choosing a tool for a team, price it as unknown rather than as zero. The terms are governed by Delaware law with jurisdiction in New Castle County, both parties waive a jury trial, and total liability is capped at the greater of $100 or what you have paid, which for now is nothing.

Compare that to the neighbours. Conductor charges $50 per month for Pro. Warp charges $20 per month for its paid individual tier and $50 per user per month for Business. Blume at zero is a strong entry point, and it is the only one of the three that does not run your agents for you, so it is additive rather than a replacement.

Does Blume keep your data local?

Mostly, and the company is more specific about the exceptions than most. Conversation transcripts are read from your machine and stored locally in a SQLite database. Your primary archive stays on your disk. Usage and spend data never reaches Blume’s servers. Telemetry is tied to a pseudonymous identifier. You can delete local data, reset your device identifier, remove connected accounts and uninstall.

The exception is written down clearly. Some features need cloud AI processing, and when they run, portions of conversation data may be sent to models through Blume’s infrastructure to generate titles and summaries, extract metadata and insights, analyse usage patterns, and to “power future intelligence features”. That last purpose is the vague one, and it is the only item on the list a privacy reviewer would object to. For the explicit Cloud option, Blume names its subprocessors, which almost nobody does: excerpts route through Vercel AI Gateway to Google Gemini, under a stated requirement of zero data retention and no training, with web search, grounding, file uploads, prompt caching and provider fallback all disabled. Cloud is never selected automatically or used as a fallback, consent lives on your device, and a change to the disclosure triggers a fresh review prompt.

The one criticism worth making is the register. The page is headlined “Privacy first, not privacy theater”, then describes its systems as “designed to be stateless whenever possible”, says the company “aims to make that behavior clear”, and calls telemetry “designed to avoid collecting conversation content”. Those are statements of intent, not commitments, and a page that opens by calling out theatre invites the comparison. The substance underneath is better than the hedging suggests.

Who is Blume best for?

Use Blume if you run two or more coding agents in parallel and have lost track of which rules apply where. The Setup panel alone solves a real filing problem, and the Usage panel removes a category of mid-task surprise. It suits developers who already write AGENTS.md and CLAUDE.md files by hand and want the repeated corrections turned into durable rules without doing the archaeology themselves. Since it is free and local-first, the cost of trying it is your download time.

Skip Blume if you use exactly one agent and one rules file, because the observation layer has little to observe. Skip it if you were sold by the analytics screenshots, which show correction, steering and frustration percentiles: that panel is marked “Soon” and is not what you get on install. And skip it if you need a documented support surface, because the docs section currently holds a single four-minute guide about making Claude usage work on Mac. For a product whose value is understanding a complicated multi-agent setup, one guide is thin.

Best Blume alternatives

ccusage covers the Usage panel and nothing else, for free. It is an MIT-licensed CLI that analyses coding agent token usage and cost from your local data, it has shipped 132 versions since May 2025, and it pulled 84,583 npm downloads in a single recent week. If plan limits are the only thing you wanted to see, install this instead and skip the desktop app entirely. It does nothing about rules, skills or hooks.

Conductor attacks the other half of the problem by running the agents rather than watching them. The free plan covers local use, Pro is $50 per month and adds cloud workspaces on isolated microVMs, and Teams is $60. It is the better buy if your bottleneck is running many agents at once. It will not audit your rulebook or tell you that you have explained the same release process three times.

Warp is the incumbent AI terminal and overlaps on agent visibility and spend tracking. Free to start, $20 per month for the paid individual tier, $200 per month for the high-usage tier, and $50 per user per month for Business. It is a replacement for your terminal rather than a companion to your agents, so adopting it is a bigger change than installing Blume, and its usage tracking is scoped to Warp’s own credits rather than your Claude and Codex plans.

Final verdict: is Blume worth it?

Blume is free, local-first, and built by two people whose backgrounds check out. The Setup and Usage panels solve real problems today, the privacy page names actual subprocessors and disables the specific model behaviours that would leak context, and the terms put the free plan in the contract rather than in marketing copy. For anyone running Claude Code, Codex and Cursor together, the download is easy to recommend on those three things alone.

The reservation is that the feature the tagline sells is the feature closest to the edge of what ships. Six of nine capabilities on the page are roadmap, the flag for cloud-assisted improvements is off in the shipped page source, and the same capability appears once in the present tense and once labelled “Soon”. Add a documentation section with one article in it and you have a product that is clearly early, honestly labelled, and worth watching. Install it now for the rulebook inventory and the usage tracking, treat the automatic improvement story as a promise rather than a purchase, and check back when the analytics panel loses its badge.