Get KULVEX — self-hosted AI for your home and lab →
KCode CLIOverview

KCode — AI Coding Assistant

KCode is KULVEX’s standalone AI coding assistant CLI. It runs 100% on your GPU — your code never leaves your machine.

Features

  • 46 built-in tools — bash, read, write, edit, multi-edit, glob, grep, grep-replace, rename, git, agent, web-fetch, web-search, LSP, cron, worktrees, plan mode, tasks, and more
  • Streaming TUI — React/Ink terminal UI with thinking blocks, permission dialogs, spinner, 11 color themes
  • Subagents — Spawn general/explore/plan agents, or define custom agents in ~/.kcode/agents/
  • Git worktree isolation — Agents work on isolated copies of your repo
  • Multi-agent swarm — Orchestrate parallel sub-agents with --agents for divide-and-conquer workflows
  • MCP support — Connect external tool servers with per-server tool allow/block lists
  • Memory system — Persistent YAML-frontmatter memories across sessions
  • Session transcripts — Full conversation logs in JSONL with full-text search
  • 152+ slash commands — /plan, /pin, /memory, /search, /compact, /rewind, /stats, /benchmark, and more
  • Enterprise managed policies — Admin-deployed policy files with locked settings, model restrictions, and audit logging
  • Lifecycle hooks — 28 hook events (PreToolUse, PostToolUse, SubagentStart, etc.) with command, prompt, and HTTP webhook types
  • Security hardening — SSRF protection, protected directories, symlink resolution, permission rules, sensitive file guards

Quick Start

# Interactive REPL
kk
 
# Single-shot query
k "explain this function"
 
# Print mode (no TUI, pipe-friendly)
k --print "list all TODO comments"
 
# With a specific model
k -m mnemo:mark5-max "refactor this module"

How It Works

KCode talks directly to llama-server (port 10091) via OpenAI-compatible SSE streaming. There’s no KULVEX API middleman — it’s a direct connection to Mnemo for minimal latency.

kcode → llama-server:10091 (SSE streaming)
         └── Mnemo model on GPU

It also supports any OpenAI-compatible API (Ollama, vLLM, LM Studio, cloud providers) via the model registry.

Architecture

  • ~20,000+ lines of original TypeScript
  • 70+ source files across core/, tools/, ui/
  • Compiled to a ~103MB standalone Bun binary
  • 488 tests, 29 test files, all passing
Creating AgentsUsage