Does anyone have a recommendation of a similar high quality, super fast tab auto-complete LLM for VSCode without switching over to Cursor? I’m fairly happy with Claude pro + VSCode for my agentic tinkering (mostly just using plan mode), but Supermaven filled a niche of a drop-in high speed autocomplete that was nice.
At Kilo (VScode plugin, BYOK so works with Claude Pro) we just rolled out autocomplete to 10% of our user base on Friday and we're happy with what we're seeing. The autocomplete is based a Mistal model (Codestral) that is fast and accurate https://mistral.ai/news/codestral We copied a lot of code from Continue.dev (best open source autocomplete in our view). We plan to roll it out to 100% of our users no later than Tuesday. The autocomplete is consumption priced (no markup) but the cost per month should be low single digit dollars.
Kilo’s been on my list of tools to try anyway. Good to hear you’ve got a strategy for useful autocomplete. Any idea how it compares with Supermaven both in terms of speed and quality? Supermaven really had found a sweet spot for both. I found LLM autocomplete was generally more annoying than it was worth prior to Supermaven.
We are working on a fully local coding assistant with auto complete and agentic modes. We created a novel post training pipeline to optimize an 80b param model to run on a standard laptop (16gb RAM) so we can offer truly unlimited and private AI coding.
I have been meaning to check Roo Code out ever since I saw Claude Code using paths with "roo" in it when executing sub agent workflows (while it was slightly confused). I assumed they borrowed the feature and did not bother changing the paths in some places.
Supermaven still is the smartest & best tab autocomplete I've ever used. And it's a shame there is ZERO competitors who come close to how "smart" it felt with it's suggestions. All other tab autocomplete AI's feel annoying and almost always get in my way.
I stopped using Supermaven a while back when their JetBrains plugins broke and the only fix was a manual edit to a java file every time the IDE updated :(
I'm building a supermaven competitor for jetbrains. We use the Jetbrains PSI (basically Jetbrain's version of the LSP) to pull definitions into context to make the autocomplete smarter. My colleague wrote a blog on this here: https://blog.sweep.dev/posts/autocomplete-context.
Copilot isn’t bad. Roo Code is an open source cursor alternative that works with VS Code. I like it more than Cursor. Much heavier than only tab completion though
Was inevitable but it is cool to see that the Cursor team are really making good use of the tech. Tab complete + the new Composer model are very good at targeted, narrow tasks.
Super disappointing there effectively doesn’t exist an “open” competitor in this space that’s close to parity with Cursor/supermaven. Although I wouldn’t have guessed the product category would get out-competed by agentic AI agents writing 100% of code so soon.
Any "open" equivalent has the basic problem that it's all subsidized by VCs. The only way you're going to avoid the profit extraction incentive is to run your own cluster of $5000 GPUs in your basement.
Does anyone have a recommendation of a similar high quality, super fast tab auto-complete LLM for VSCode without switching over to Cursor? I’m fairly happy with Claude pro + VSCode for my agentic tinkering (mostly just using plan mode), but Supermaven filled a niche of a drop-in high speed autocomplete that was nice.
At Kilo (VScode plugin, BYOK so works with Claude Pro) we just rolled out autocomplete to 10% of our user base on Friday and we're happy with what we're seeing. The autocomplete is based a Mistal model (Codestral) that is fast and accurate https://mistral.ai/news/codestral We copied a lot of code from Continue.dev (best open source autocomplete in our view). We plan to roll it out to 100% of our users no later than Tuesday. The autocomplete is consumption priced (no markup) but the cost per month should be low single digit dollars.
Kilo’s been on my list of tools to try anyway. Good to hear you’ve got a strategy for useful autocomplete. Any idea how it compares with Supermaven both in terms of speed and quality? Supermaven really had found a sweet spot for both. I found LLM autocomplete was generally more annoying than it was worth prior to Supermaven.
Sorry but I'm don't know how Kilo autocomplete compares with Supermaven. Maybe someone else does?
BTW For completeness, while we’re still in rollout for existing users autocomplete is default on for all new users already.
It doesn't look near instant like Supermaven because the default delay before suggestions is 3 s.
https://kilo.ai/docs/basic-usage/autocomplete
Hitching my wagon to this train to ask: same question, but for nvim!
We are working on a fully local coding assistant with auto complete and agentic modes. We created a novel post training pipeline to optimize an 80b param model to run on a standard laptop (16gb RAM) so we can offer truly unlimited and private AI coding.
Sign up for our beta https://cortex.build
I was interested but looks like it's only available for Macos.
Roo Code is awesome and free and open source.
I have been meaning to check Roo Code out ever since I saw Claude Code using paths with "roo" in it when executing sub agent workflows (while it was slightly confused). I assumed they borrowed the feature and did not bother changing the paths in some places.
I haven't used it recently but perhaps TabNine
RIP. Supermaven had great support for neovim for tab completion, will definitely miss that.
Supermaven still is the smartest & best tab autocomplete I've ever used. And it's a shame there is ZERO competitors who come close to how "smart" it felt with it's suggestions. All other tab autocomplete AI's feel annoying and almost always get in my way.
I stopped using Supermaven a while back when their JetBrains plugins broke and the only fix was a manual edit to a java file every time the IDE updated :(
I'm building a supermaven competitor for jetbrains. We use the Jetbrains PSI (basically Jetbrain's version of the LSP) to pull definitions into context to make the autocomplete smarter. My colleague wrote a blog on this here: https://blog.sweep.dev/posts/autocomplete-context.
All I wanted was fast on-demand code completion. Guess 2025 is a lost year on that front.
Copilot isn’t bad. Roo Code is an open source cursor alternative that works with VS Code. I like it more than Cursor. Much heavier than only tab completion though
* in an editor other than vscode, for me anyway was important. In my testing Supermaven was the best in nvim. Copilot is... fine I guess.
Was inevitable but it is cool to see that the Cursor team are really making good use of the tech. Tab complete + the new Composer model are very good at targeted, narrow tasks.
it's crazy that the big link at the bottom is still there..
Hmmm they did charge my CC two days ago on 22nd though…
Supermaven was awesome RIP
Super disappointing there effectively doesn’t exist an “open” competitor in this space that’s close to parity with Cursor/supermaven. Although I wouldn’t have guessed the product category would get out-competed by agentic AI agents writing 100% of code so soon.
Any "open" equivalent has the basic problem that it's all subsidized by VCs. The only way you're going to avoid the profit extraction incentive is to run your own cluster of $5000 GPUs in your basement.
Based on, of all people, Pewdipie's videos, seems around 25k in GPUs (4090s mostly) will get you pretty usable local results.
Roo Code is excellent
You're all over this thread recommending Roo Code, but my understanding is that it's a coding agent, rather than super-fast tab complete