Muse Code 一周额度用尽实测反馈

Vebjørn @VJNCapital原推发布于 2026.09.14

在 X 上打开原推

I burned through my weekly usage on $META Muse Code My biggest takeaway was not what I expected: The model is not the problem The product around it is Here is my full review + feedback @alexandr_wang -> The model Muse Spark 1.3 Max is really good It is fast, follows instructions extremely well and handled pretty much everything I threw at it. Great agentic model UI work, API integrations, database migrations, creative work, refactoring for scalability and efficiency, and more complex logic It rarely just completely fails at a task My biggest issue is that it does not adapt to the existing patterns in the repo as well as Claude/Codex It will solve the task correctly, but sometimes in a way that creates unnecessary follow-up work because it introduced a new pattern instead of understanding how the rest of the codebase already works And I am honestly not sure how much of this is even the model I think the harness is limiting it -> The usage limits This is probably the best part The 5-hour limit is extremely generous I could run multiple parallel agents for hours before hitting it And after using the full weekly limit, I still think this is one of the biggest advantages Muse has over Claude/Codex With Claude/Codex I constantly think about usage Should I use the expensive model here? Should I switch models? Am I going to kill my 5-hour limit? With Muse I basically just used the best model at Max for everything and worked That is how these subscriptions should feel -> The harness / CLI This is where Muse Code falls behind badly The CLI feels clumsy and honestly feels like it is limiting how useful the model can be Session management is bad Running multiple agents means having multiple terminals open with no good naming or grouping Switching between parallel tasks is annoying There is weird autoscrolling I had multiple random freezes while the model was thinking where I had to refresh the session and resume it I also got random rate-limit retries There is no proper compacted project/chat memory that lets Muse understand the codebase across sessions That last one is probably the biggest problem Claude/Codex increasingly feel like they understand the project Muse feels more like a very good engineer that wakes up with amnesia every new session lol The voice experience also needs work I want proper record -> transcribe -> prompt. Now the only option is 3rd party + copy/paste Codex has basically nailed this imo And I really miss having proper usage stats visible: context usage, 5-hour usage and weekly usage Just put it in the HUD -> Value The value is ridiculous $15 for this amount of frontier model usage is probably the best value coding subscription I have used But after using the full weekly limit I also realized something Price is not the main reason I pay for Claude/Codex I am paying for the complete package The model, harness, project context, session management, voice and reliability And right now I would still rather pay more for Claude/Codex because of that -> Would I replace Claude/Codex with Muse? Not yet But I don’t think the model is what needs to improve much (it will on next iteration for sure though) Muse Spark is already good enough to compete Muse Code just isn’t Give me a proper desktop app, projects with grouped/named sessions, easy shortcuts between parallel agents, persistent/compacted project memory, better session reliability, proper voice recording/transcription, usage stats, output preferences… That would probably be what makes this an actual hit imo. What i am left with is that the potential feels limited by the CLI and I just can't bother trying to work around it Until then I will probably end my Muse Code subscription Muse subscription in Codex would be the true meta