koda-core

Engine library for the Koda AI coding agent.

Pure logic with zero terminal dependencies — communicates exclusively through EngineEvent (output) and EngineCommand (input) enums over async channels.

What's inside

• LLM providers — 14 providers (Anthropic, OpenAI, Gemini, Groq, Ollama, LM Studio, etc.)
• Tool system — 20+ built-in tools (file ops, shell, search, memory, agents)
• Per-tool approval — three modes (Auto/Strict/Safe) with effect-based safety classification
• Inference loop — streaming tool-use loop with parallel execution
• SQLite persistence — sessions, messages, compaction
• MCP client — connects to external MCP servers for extensibility

Rust edition: 2024

Usage

koda-core is a channel-driven engine. Create async channels, spawn the inference loop, and drive the engine through EngineCommand/EngineEvent pairs:

use koda_core::engine::{EngineCommand, EngineEvent};
use tokio::sync::mpsc;

// The engine communicates exclusively through async channels.
// EngineEvents flow out (streaming text, tool calls, approvals).
// EngineCommands flow in (approval responses, cancellation).
let (cmd_tx, mut cmd_rx) = mpsc::channel::<EngineCommand>(64);
let (evt_tx, mut evt_rx) = mpsc::channel::<EngineEvent>(64);

// Spawn the inference loop, then select over evt_rx for streaming
// output and cmd_tx to send approval decisions back.
// See koda-cli for a complete implementation.

See DESIGN.md for architectural decisions.

License

MIT