Kalai
AI automation: safe commands, file edits, and web search.
12 followers
AI automation: safe commands, file edits, and web search.
12 followers
Kalai Agent is a lightweight AI coding assistant for Visual Studio Code. Ask coding questions, get explanations, and chat about your project in a simple sidebar interface.
Free Options
Launch Team / Built With
@culustech Yeah I did, it is super awesome !!!!!
@jnrkalculus Thank you Bossman
how does it handle large codebases, does it index the whole project or just the file you have open? curious about context limits.
Hi@mtoreli24044 - great question, this is exactly the right thing to ask before trusting an agent with a big repo.
Kalai indexes the whole workspace, not just the open file. When you first open a project, it scans and indexes your files locally in VS Code, then keeps that index updated. For each query, it searches the index and pulls in the most relevant files as context.
How it handles large codebases in practice:
Respects .gitignore and skips the usual suspects (node_modules, dist, build, .git, etc.) plus binary files.
Index guardrails: by default it indexes up to 2,000 files, skips files larger than 512 KB, and reads up to 50 KB of content per file. If your repo is bigger, indexing is truncated and you'll see a note.
Relevance-first context: for every prompt it picks the top 10 most relevant files based on keyword/path matching, not the whole project tree.
Context payload budget: the combined content of those relevant files is capped at 200,000 characters per request. It also prioritizes an AIREF.md file if you have one, so project-level guidance always fits in first.
Model context window: the default coding model (kimi-k2.7-code) has a 262,144-token context window, and there's a context-budget manager that trims older tool-call turns and truncates the largest message if history grows too large.
So the short version: it does project-wide indexing, but only the most relevant slices of the codebase are sent with each request, with hard caps to keep latency and cost reasonable.
Let me know if you'd like more detail on any part!
The sidebar setup feels really natural, just asks a question and it pulls in the right context from my file without me copying anything over. Honestly surprised how little it gets in the way compared to heavier assistants.
@abdurrahmaaz1o Hey Abdurrahman, thank you so much for this - really means a lot! So glad to hear Kalai fits smoothly into your workflow and that the context awareness is working well for you. Keeping it lightweight and helpful is the goal, so feedback like this keeps us going. If you ever run into anything or have ideas, feel free to reach out anytime!
does it use my own api key or is it a hosted model on your end, and how does that change the pricing?
Hi@cemreelibacebu - thanks for asking.
Kalai is usage-based, no subscription. You top up a wallet from kalai portal and pay only for the tokens you use. No need to bring your own API key - the AI provider key is managed on our backend.
Pricing depends on the model (e.g., our default coding model starts around $0.00095 / 1K input tokens and $0.004 / 1K output tokens). Credits don’t expire, and you can recharge anytime.
Hope that helps!
how does it compare to the built-in copilot stuff already in vscode, especially on price?
Hi@tuncay2tva - great comparison to ask about.
The biggest difference is the pricing model:
VS Code Copilot is a subscription - you pay a fixed monthly/annual fee regardless of how much you actually code.
Kalai is usage-based, no subscription. You top up a wallet and pay only for the tokens you consume. If you code lightly one month, you spend less; if you have a heavy month, you spend more.
Model-wise, Copilot is built around OpenAI/Codex, while Kalai runs on Kimi/Moonshot models (with kimi-k2.7-code as our default for coding tasks). We also keep your code visible alongside the chat rather than the inline-only or side-panel flow you may be used to.
For occasional users or anyone who dislikes another monthly subscription, Kalai is usually the cheaper and more flexible option. For someone already deep in the GitHub ecosystem and coding many hours a day, Copilot can still make sense.
Hope that helps!
How does Kalai handle the context window when you're chatting about a large project with lots of files open?
Hi@aye9748910302 - great question.
Kalai handles this in a few layers:
Project context
It indexes your whole workspace (respecting .gitignore), so it knows about all your files, not just the open ones.
When you ask something, it searches the index and pulls in the most relevant files rather than dumping everything into the prompt.
Limits & guardrails
Up to 10 relevant files are included per request.
File content is capped at ~50 KB per file, with a total context payload limit of 200,000 characters.
The default model, kimi-k2.7-code, has a 262,144-token context window, so there's plenty of room even after project context is added.
If your chat history grows long, Kalai trims older tool turns and truncates the largest message to stay within budget.
So even with a big project and many open files, you're not feeding the whole repo into every request - just the smartest slice, within hard limits.
Hope that helps!
Slim and tidy setup, no extra config needed. The sidebar chat feels natural while keeping my code in view, which I appreciate more than I expected.
Hi@toprakcwek - thank you, really glad the sidebar layout clicked for you! Keeping your code in view while chatting was exactly the goal, so it's great to hear it feels natural. If you ever have feedback or feature ideas as you use it more, we'd love to hear them.