KlaatAI — AI Coding Assistant for VS Code

Your AI pair-programmer that lives inside the editor. Agentic chat with 30+ tools, inline edits, tab completions, and a code graph + memory system that learns how you work — all powered by automatic multi-model routing so you never manage API keys or pick models.

_{Same 33-task suite, one harness, identical prompts. Full methodology: klaatai.com/benchmarks}

Why KlaatAI?

• No API keys. Sign in once — KlaatAI routes every request to the best model for the task.
• Multi-model routing. Fast models for quick edits, frontier models for hard reasoning — automatic or manual override.
• Full agent, not just chat. Reads, writes, edits files, runs shell commands, searches your codebase, browses the web, reviews PRs — all with inline approvals.
• Learns you. Remembers your stack, conventions, and style across sessions. Fully transparent and deletable.
• Cross-file intelligence. Completions and Cmd+K edits understand your imports, types, and APIs — not just the open buffer.

Features

Agentic Chat

Open the KlaatAI sidebar (Cmd/Ctrl+Shift+K) and describe what you want. The agent executes multi-step tasks autonomously with inline approvals — nothing destructive happens without your OK.

30+ built-in tools:

Four agent modes for different workflows:

Enrich prompts with @-mentions:

Drag & drop files from the OS or VS Code explorer, or paste screenshots directly into chat.

Inline Edit — Cmd/Ctrl+K

Select code, press Cmd+K, describe the change. Review the diff inline and apply with a single undo step.

• Cross-file context — knows your imports so it uses the right types and APIs
• 9-pass fuzzy edit engine for precise, reliable patches
• Streams the diff as it generates
• Single Ctrl+Z reverts the entire change

Tab Completions

Multi-line, fill-in-the-middle ghost-text suggestions as you type.

• Cross-file aware — pulls exported symbols from your imports
• In-memory cache + in-flight dedup for low latency
• Toggle with klaatai.inlineCompletions

Memory — Learns Your Style

KlaatAI remembers your stack, conventions, and response preferences across sessions — no manual setup.

• Automatic with salience and decay (forgets what's no longer relevant)
• Full transparency: KlaatAI: View Memory shows exactly what it knows
• Delete anything or turn it off entirely with klaatai.memory

Code Graph — Knows Your Codebase

Indexes your repo into a code knowledge graph — symbols, call-graph edges, and relationships — so it finds the right code without reading every file.

• Tree-sitter symbol extraction (TypeScript, Python, Go, Rust, Java, and more)
• Background indexing with progress UI
• Shared with the KlaatCode CLI — index once, use everywhere

Auto-Debug

Catches errors your edits introduce and fixes them automatically.

• Watches LSP diagnostics on KlaatAI-edited files
• One-click or fully automatic fix rounds
• Fix with KlaatAI lightbulb on error squiggles
• Toggle with klaatai.autoDebug

Multi-Model Routing

Automatic tier selection routes each request to the best model — override anytime via the status bar.

The ladder is Nano → Fast → Code → Reason → Heavy → Titan. Titan is the top rung: Auto reaches it only by escalating above Heavy when a turn keeps failing or outgrows the context — never from a prompt's wording alone — and you can always pick it yourself. It is the priciest tier by a wide margin, so per-plan daily caps apply and exhausted requests fall back to Heavy.

Extensibility

• MCP (Model Context Protocol) — connect external tool servers via klaatai.json
• Custom plugins — add your own tools in .klaatai/tools/*.js
• Workspace rules — auto-loaded from .klaatai/rules.md, AGENTS.md, or .cursorrules

Agent Depth

• Background agents in isolated worktrees — fire a task that runs in its own git worktree; your working tree stays untouched until you review and merge the result (PR-style merge-back).
• Regression gate — after a run of edits, KlaatAI auto-detects and runs your test suite, surfacing failures before you do.
• Adversarial verify — cross-review a diff with a different, stronger model than the one that wrote it.
• Blast-radius awareness — when the agent edits a symbol, it sees who calls it (from the code graph) plus a churn heat-map flagging fragile call sites.

Getting Started

1. Install the extension from the VS Code Marketplace.
2. Click the KlaatAI icon in the activity bar.
3. Sign in — opens your browser, no API keys needed.
   • On SSH/remote/headless: use KlaatAI: Sign In (Device Code) instead.
4. Open a folder or repo and start chatting. The code graph indexes in the background.

Commands & Keybindings

Settings

Pricing

Install, sign in, start coding — no credit card, no API keys. A free tier gets you going; paid tiers unlock higher limits and priority routing. The differentiator is automatic multi-model routing: you get frontier-model quality on the hard tasks and cheap-model speed on the easy ones, without paying frontier prices on every request.

Why it's cheaper without being worse

The same efficiency engine powers our CLI, which solved 33/33 tasks at $0.89 — 1.7× cheaper than the nearest rival and 5.4× cheaper than Claude Code (Sonnet 5) on one harness with identical prompts (see the chart above). This extension ships that engine: output compression, per-tier tool dialects, attention-ordered context, and budget guards — all surfaced transparently (see Show Context Breakdown).

Privacy & Security

• Memory can be turned off (klaatai.memory: false). View Memory lets you inspect and delete anything.
• Code indexing can be turned off (klaatai.indexProject: false).
• Telemetry is opt-in and follows your VS Code telemetry setting — off means nothing is sent (klaatai.telemetry).
• Sign-in tokens are stored in the OS keychain via VS Code SecretStorage.
• No code is sent to third parties beyond the model providers used for inference. Full details: PRIVACY.md.

Links

• Website
• Documentation
• Changelog
• Report an Issue

Made with care by the KlaatAI team.