KlaatAI/klaatcode

Open-source AI coding agent for the terminal. Claude Code-grade accuracy with smart model routing — uses the right AI model for each task, cutting costs 10x. Supports Claude, GPT, Gemini, DeepSeek & more.

What it solves

Klaat Code is a terminal-native AI coding agent designed to provide high-accuracy coding assistance while drastically reducing the cost and token usage associated with large language models. It addresses the common issues of "lost in the middle" context loss, runaway agent costs (doom-loops), and the inefficiency of reading entire files when only specific symbols are needed.

How it works

Klaat Code acts as a thin client to the Klaatu routing service. Instead of using a single model, it uses a smart router (Klaatu-o1) that classifies every request and dispatches it to one of six cost tiers (from nano to titan) based on the task's complexity. It utilizes a server-side code knowledge graph to perform semantic search and impact analysis, allowing the agent to query symbols and callers rather than reading full files. To maintain efficiency, it employs mechanical compaction and noise-filtering of tool outputs to keep the context window clean and affordable.

Who it’s for

It is built for developers who want a terminal-first AI coding agent that can read, edit, and verify its own work with strict cost controls and high reproducibility.

Highlights

  • Smart Model Routing: Automatically switches between model tiers per-request to balance cost and performance.
  • Code Knowledge Graph: Uses a call graph for semantic search and planning, reducing token usage by 5–15×.
  • Cost Guards: Includes a real-time burn-rate monitor, per-task cost attribution, and a "doom-loop" breaker to prevent runaway billing.
  • Context Management: Features automatic summarization with a self-check mechanism to verify that critical state wasn't lost during compaction.
  • Agentic Toolset: Includes tools for file editing, shell command execution, web search, and multi-agent delegation for scoped work.
  • Coded Diagnostics: Automatically runs linters and typecheckers after edits to fix mistakes before returning control to the user.
  • MCP Support: Full Model Context Protocol client supporting both stdio and streamable HTTP transports.

Related

  • Project
    colbymchenry/codegraph

    CodeGraph is a Rust‑based local tool that builds a full‑code knowledge graph for any project (supports dozens of languages) and integrates it with AI coding assistants (Claude Code, Cursor, Copilot, Gemini, etc.). The graph lets agents retrieve exact symbols and call‑paths in a single request, cutting tool calls, file reads, token usage and cost dramatically. Install via a one‑line script, run `codegraph init` per repo, and the CLI auto‑updates the graph as you edit. A browser UI (`codegraph ui`) visualises the graph, and the CLI can uninstall itself cleanly.

  • Project
    0xku/kon

    A minimal coding agent harness with a tiny core prompt and a small toolset, designed for efficiency and transparency in terminal-based AI coding.

  • Project
    aovestdipaperino/tokensave

    TokenSave is a local MCP server that builds a semantic knowledge graph of a codebase, letting AI coding agents query symbols, callers, and impact instead of scanning files with grep and read, cutting token usage and tool calls.

  • Project
    LiuMengxuan04/MiniCode

    A lightweight terminal coding assistant that implements a Claude Code-like agentic workflow for local development, focusing on simplicity and extensibility.

  • Project
    tigicion/dao-code

    A terminal-native AI coding agent optimized for DeepSeek V4 that uses advanced cache engineering to provide high-capability software engineering at a fraction of the cost of competitors.