Kincode — AI skill for Claude Code

kincode

A lightweight AI coding assistant for your terminal. Written in Go. Single binary. Zero dependencies.

Like Claude Code, but open-source and 10x lighter.

Features

• Single binary (~10MB), zero runtime dependencies
• Multi-provider: Anthropic, OpenAI, Ollama (any OpenAI-compatible endpoint)
• 14 built-in tools: bash, bash_output, file_read / file_write / file_edit / multi_edit, glob, grep, git, web_fetch, web_search, memory, agent_spawn, todo_write
• Approval gate — permissions: {mode: ask, ask: [...], allow: [...]} in the soul. A matching call stops for a human: a terminal prompt, or an approval card in KinClaw Mac. Same rule grammar as kinclaw, and a bash(...) allow rule is checked against every simple command the shell would run — bash(go test*) does not wave through go test ./... > ~/.zshrc
• Builds after it edits — a round that changed files is followed by go build ./... / cargo check / tsc --noEmit / swift build, and the compiler's answer goes into the tool result the model reads. The edit → verify → fix loop is mechanical rather than something the prompt asks for
• Undo — POST /api/undo takes back the last turn's file changes, and only those: your own uncommitted work is out of scope by construction, which git checkout cannot say
• Reads the repo — a read-only git tool (status / diff / log / show / blame, allowed in plan mode) plus branch, dirty files and recent commits in the system prompt, so a session does not open with four rounds of orientation
• Background commands — bash(..., background: true) for a dev server or a slow suite; bash_output follows its log. And cd carries into the next call, the way a terminal does it
• Image input — vision blocks for Anthropic claude-3+ and OpenAI gpt-4o. Send screenshots, mockups, diagrams via POST /api/chat {message, images:[{media_type, data}]}
• Plan mode — POST /api/plan_mode {enabled} toggl