KodaCode

What makes KodaCode truly yours

Your AI, your rules. Prompts, agents, themes, models, and permissions are all files you control.

Engineer: Explore → Plan → Execute

The default agent researches your codebase, creates a structured plan, and waits for your approval before making changes. You review the approach before any code is touched.

Coder: Direct Implementation

For straightforward tasks, the coder agent skips planning and makes changes immediately. Switch with Tab — use the right workflow for the task at hand.

Sandboxed by Default

Every tool call is confined to your project directory. Path escapes, delete commands, and external access require explicit permission. Glob-based allow/ask/deny rules per tool and file pattern.

Multi-Provider, Your Choice

Anthropic, OpenAI, Google, Groq, Ollama, LM Studio, or any OpenAI-compatible endpoint. Switch mid-session. Mix cloud and local models.

Configurable at Every Layer

Utility models, budget caps, reasoning effort, permissions, LSP servers, diagnostics, agents, system prompts, themes — all controlled via YAML config and markdown files.

Project Memory

Layer global, project, and personal instructions via KODA.md files. Use /remember to save learnings on the fly. Conventions carry across sessions.

Built In

No plugins required for the core workflow.

Cost Control

Utility model routing sends titles, compaction, and research tasks to a cheap model. Adaptive reasoning budgets scale down during tool loops. Per-session budget caps stop runaway spending.

Context Management

Pruning, compaction, and safety limits keep long sessions coherent. Old tool outputs are replaced with summaries. The model doesn’t lose track of what you’re doing.

LSP Code Intelligence

Go-to-definition, find references, hover info, and diagnostics via language servers. Auto-discovers servers from your project’s tooling. Auto-lints after every edit.

Live Diff Preview

See file changes as the model writes them — before they hit disk. Cancel mid-write if something looks wrong.

Background Tasks & Subagents

Long-running commands execute in the background while the model keeps working. Subagents handle parallel research. Results feed back automatically.

MCP Integration

Extend with any Model Context Protocol server — stdio or SSE. Tools merge with built-ins and respect the same permission system.

Sessions & Snapshots

Persistent sessions with branching and per-turn git snapshots. Resume where you left off or replay the workspace state at any turn.

Loop Detection

Detects when the model gets stuck repeating tool calls. Progressive intervention: warn, nudge, then stop. Prevents wasted tokens on infinite loops.

Auto-Diagnostics

After every file edit, linters and LSP diagnostics run automatically. The model sees errors immediately and fixes them in the same turn — no manual re-run needed.

Quick Install

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/sageil/koda/main/install.sh | sh

# Homebrew
brew tap sageil/tap && brew install koda

# Go
go install github.com/sageil/kodacode/v1/cmd/koda@latest

Then run koda to start.