Cutad.AI — AI Coding, Debugging & Planning for VS Code

An AI assistant that codes, debugs, and plans inside VS Code — powered by the CutadAI model gateway, with agent orchestration, sub-agent specialists, 25 built-in skills, and MCP in both directions.

@cutad what does this repo do?              ← chat with reasoning display
@cutad /agent fix the failing tests         ← autonomous coding with approvals
@cutad /plan add OAuth login                ← durable plans with step approval
@cutad /debug why does auth.spec.ts fail?   ← real debugger + test state

Quick start

1. Install the extension (VSIX or Marketplace).
2. Set your API key — command palette → Cutad.AI: Set API Key (or edit ~/.cutad/config.json).
3. Chat — open the Chat view (Ctrl+Alt+I) and type @cutad hello.

The key is stored in ~/.cutad/config.json (home directory, outside any repo). Get a key at ai.cutad.web.id.

Commands

Palette commands: Set API Key · Select Model · Test Connection · Open Config · Show Diagnostics · Start MCP Server · Reconnect MCP Servers · Open Plan

Features

Chat with reasoning

Streams answers with the model's thinking shown in a collapsible block (when the model supports it). Context comes from your active file/selection, #-referenced files, and workspace diagnostics — with an explicit token budget and omission notes.

Coding agent with approvals

Every mutating action (file writes, terminal commands, git operations) requires confirmation. Read-only tools run freely. Destructive commands (rm -rf, git reset --hard, force push, curl | sh) can never be auto-approved. In untrusted workspaces, all mutations are disabled.

Planning mode

Plans are durable artifacts stored in SQLite with step-by-step execution, per-step approval, skip/abort, and resume across window reloads (LangGraph checkpointing).

Debug mode

Reads real runtime state via the debug adapter: threads, stack frames, scopes, and variables at a breakpoint. Runs your test suite (auto-detects npm/pytest/cargo/go) and parses failures from vitest/jest/mocha/pytest/cargo output. Evidence-first: reproduce → read state → one hypothesis → fix root cause → verify.

Sub-agent specialists

Delegate focused work to isolated specialists: explorer (map unfamiliar code), researcher (investigate dependencies), reviewer (independent diff review), tester (run verification). All read-only; the parent agent owns every mutation.

25 built-in skills

Markdown playbooks in ~/.cutad/skills/ — each with an objective, goals, and triggers. Matched automatically by your prompt and injected into the agent's context.

Debugging · Code Review · Refactoring · Test Writing · Planning · API Integration · Error Handling · Security Review · Performance · TypeScript Strict · Git Workflow · Code Exploration · Dependency Management · Async & Concurrency · Data Modeling · CLI Development · Frontend & UI · Documentation · Release Management · Incident Response · Log & Output Analysis · Context Management · Tool Orchestration · Prompt Engineering · Skill Authoring

Skills are plain markdown — edit them, delete them, or ask the agent to create new ones (create_skill tool). The agent also learns reusable procedures from completed sessions automatically.

MCP — both directions

Consume external MCP servers — add them to .vscode/mcp.json, .mcp.json, or ~/.cutad/mcp.json:

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
    }
  }
}

Expose Cutad.AI's tools as an MCP server — Cutad.AI: Start MCP Server, or point external clients at dist/mcp-server.js:

{
  "mcpServers": {
    "cutad": {
      "command": "node",
      "args": ["<extension-path>/dist/mcp-server.js", "--workspace", "<your-repo>"]
    }
  }
}

In-editor mode routes mutations through the VS Code approval gate. Standalone mode has no UI and auto-approves — use it only in trusted workspaces.

Built-in tools (20)

Read: read_file · list_dir · search_text · find_files · get_workspace_tree · get_diagnostics · get_open_editors · get_selection · git_status · git_diff · get_debug_state · list_breakpoints · fetch_url · list_skills · read_skill Mutating (approval required): apply_patch · write_file · run_command · run_tests · set_breakpoint · remove_breakpoint · launch_debug · create_skill · delete_skill · delegate_to_specialist

Configuration

All settings live in ~/.cutad/config.json:

{
  "apiKey": "cag_...",
  "baseUrl": "https://ai.cutad.web.id/v1",
  "model": "workbuddy/deepseek-v4.1-flash",
  "maxOutputTokens": 8192,
  "reasoning": { "show": true, "effort": "high" },
  "agent": { "maxIterations": 25, "autoApprove": [] },
  "tools": { "truncateKB": 64 },
  "telemetry": { "enabled": false }
}

Storage

~/.cutad/
├── config.json          ← settings + API key
├── cutad.db             ← SQLite: sessions, messages, memory, knowledge, tool usage, plans
├── sessions/<id>/
│   └── state.json       ← per-session JSON state (counters, active plan, files changed)
├── skills/*.md          ← skill playbooks (edit freely)
└── mcp.json             ← optional global MCP server config

Session scratch files go to your OS temp dir (%TEMP%/cutad/<id>/), never into the workspace.

Requirements

• VS Code 1.140+
• A CutadAI API key (cag_...) with an active plan

Troubleshooting

Diagnostics: Cutad.AI: Show Diagnostics opens the output channel with version, model, config path, and recent errors.

License

MIT