ThinCoder for VS Code

Full AI coding agent, inside your editor. Zero bloat.

ThinCoder VS Code gives you the complete ThinCoder agent — multi-turn tool-calling, subagents, plan mode, task tracking — in a side panel. It reads your project, writes code, runs commands, searches the web, and explains its work as it goes.

Like the CLI, it's pure .mjs, zero third-party runtime dependencies, and connects directly to top-tier models via OpenAI-compatible APIs.

Features

• Full agent loop — multi-turn reasoning, tool execution, context compaction
• 20+ built-in tools — read, write, edit, bash, glob, grep, git (diff/status/log), websearch, fetch, checkpoint, question, syntax_check, and more. Tools grouped for parallel execution (consecutive readonly tools run concurrently).
• Subagents — explore (codebase search) / plan (architecture design) / coder (implementation), dispatched with role-specific prompts
• Plan Mode — read-only exploration → design proposal → user approval → implement
• Task tracking — task tool breaks down multi-step work with pending/in_progress/done status
• Verify guard — file changes without verify get pushed back; syntax check must pass before claiming completion
• Context compaction — automatically summarizes old messages when conversation grows too long
• Repo outline — scans project import/export dependencies so the agent understands your codebase structure
• Multi-session — save and switch between conversation sessions with session bar; LLM auto-generates titles
• Image input — paste or drag images into chat, or use read_image tool; supported on vision models (Kimi K3, Qwen3.7, MiniMax M3)
• Reasoning display — collapsible "Thinking..." block shows the model's reasoning process in real-time
• 21 provider presets — DeepSeek, Kimi, Kimi For Coding, GLM, GLM Coding Plan, Qwen, Qwen Token Plan, MiMo, MiMo Token Plan, MiniMax, OpenAI, Claude, Gemini, Grok, Mistral, Volcengine, Hunyuan, Tencent TokenHub, SiliconFlow, OpenRouter, Groq + custom OpenAI-compatible endpoint
• Vector search — semantic code search with BAAI/bge-m3 embeddings via SiliconFlow (configurable in Settings)
• Model selection — choose from all available models per provider, with reasoning effort control
• Permission control — session-level AUTO mode (off by default): every file-modifying tool prompts for approval until you click the AUTO toolbar button or "Approve All"; the flip takes effect immediately, even mid-turn
• Right-click — select code in editor, right-click → "Ask ThinCoder"

Requirements

• VS Code >= 1.104.0
• An API key for at least one supported provider

Quick Start

1. Install the extension from VS Code Marketplace (or vsce package + sideload)
2. Press Ctrl+Alt+T (Mac: Cmd+Alt+T) — the ThinCoder panel opens in the sidebar
3. First run shows a welcome panel: pick a provider, paste your API key, save (the key is verified on save). You can add more providers later via ⚙ (Settings) in the toolbar
4. Start chatting — ThinCoder reads your workspace and responds

Commands

Configuration

All settings under thincoder.* in VS Code settings:

Auto-approve is not a VS Code setting — it is a session-level flag stored in the shared session slot (~/.thincoder/sessions/), same as the CLI. Toggle it with the AUTO toolbar button in the chat panel or via a prompt's "Approve All" button; each session remembers its own state.

Web search (optional Tavily)

The websearch tool defaults to Bing HTML extraction (zero-config, no key). To use Tavily's structured search API — stable JSON results, no HTML scraping, resilient to page-structure changes — add this to ~/.thincoder/config.json (shared with the CLI):

{
  "websearch": {
    "apiKey": "tvly-..."   // https://tavily.com — free monthly tier available
  }
}

No key (or a bad key) → the tool silently falls back to Bing, so agents never lose search.

Supported Providers

Chat panel shortcuts

Keyboard shortcuts inside the chat panel (CLI parity):

Architecture

┌──────────────────────────────────────────┐
│  VS Code Extension Host                  │
│  ┌────────────┐  ┌────────────────────┐  │
│  │ ChatPanel  │  │  Agent Loop        │  │
│  │ (session   │  │  (agent.mjs)       │  │
│  │  mgmt,     │  │  - tool execution  │  │
│  │  settings) │  │  - context compact │  │
│  └─────┬──────┘  │  - subagent spawn  │  │
│        │         └────────┬───────────┘  │
│   postMessage            │               │
│        │         ┌───────┴───────────┐   │
│        │         │  Provider         │   │
│        │         │  (fetch + SSE)    │   │
│        │         └───────────────────┘   │
│  ┌─────┴──────────────────────────────┐  │
│  │  Webview (iframe)                  │  │
│  │  - chat.js (orchestration)         │  │
│  │  - ui.js (DOM rendering)           │  │
│  │  - md.js (Markdown)                │  │
│  └────────────────────────────────────┘  │
└──────────────────────────────────────────┘

• Extension host runs the agent loop, manages sessions, and talks to LLM providers
• Webview handles the chat UI — message display, model selector, session history, settings panel
• Communication is one-way message passing: webview sends userMessage/abort/setup etc., extension sends token/toolCall/complete/error etc.

Development

# No build step — the extension runs directly from source
# Edit extension.mjs, src/*.mjs, webview/*.{js,css,html}

# Package for distribution
npx @vscode/vsce package

Project Structure

thincoder-vscode/
├── extension.mjs         # VS Code entry point
├── package.json          # Extension manifest
├── AGENTS.md             # Developer guide
├── README.md             # This file
├── src/
│   ├── agent/                 # Shell assembly — setup, tool table, run stages, reminders, gates
│   ├── agent.mjs              # Agent loop — tool batching, image injection, compaction, subagents
│   ├── agent-tools/           # Meta-tool wiring — registry re-export + stop/discard face
│   ├── config-mcp.mjs         # MCP config shell — panel edits persist to config
│   ├── embed-config.mjs       # Embedding / vector config shell
│   ├── explore-distill.mjs    # Explore-summary adapter (core summarizer wrapper)
│   ├── extension/             # Chat panel modules — panel-*, session slots/IO, settings, presets
│   ├── i18n.mjs               # UI translation shell (core dictionary + end keys)
│   ├── memory-tool.mjs        # memory tool shell face (core memory single source)
│   ├── prompt-injections.mjs  # Prompt-anchor values for this end — tool face (bash / question)
│   ├── repomap.mjs            # Repository dependency outline (workspace.fs data source)
│   ├── specs.mjs              # Model capability specs (core table + end-side additions)
│   ├── tools/                 # Host tools — code, focus, ide, shell + core tool wiring
│   └── tools.mjs              # Re-export shim → src/tools/index.mjs
├── webview/
│   ├── chat.js           # Frontend orchestration
│   ├── ui.js             # DOM helpers
│   ├── md.js             # Markdown renderer
│   ├── index.html        # Webview shell
│   ├── base.css          # Reset, variables, layout
│   ├── chat.css          # Messages, markdown, tool calls
│   ├── controls.css      # Input area, controls, dropdown
│   ├── session.css       # Session bar
│   └── settings.css      # Settings panel
└── docs/
    └── design/           # Design docs (migration-period archive — authoritative Chinese design docs = repo-root docs/core/design/)

vs. ThinCoder CLI