CodeRun AI Agent

A multi-provider AI coding agent for VS Code. Supports Ollama, OpenAI, Anthropic, Google Gemini, OpenRouter, and any OpenAI-compatible API.

Features

• Multi-provider support — Switch between Ollama, OpenAI, Anthropic, Gemini, OpenRouter, and generic OpenAI-compatible endpoints
• Copilot-style tool permissions — Allow/Deny dangerous actions with "Always Allow" / "Always Deny" options
• Real-time streaming — Live thinking tokens, content, tool calls, and terminal output
• VS Code terminal integration — Real terminal execution, not just child_process
• File operations — Read, write, edit, delete, search with workspace safety checks
• Modular architecture — Clean separation: providers, tools, prompts, loop, storage

Supported Providers

Compatible with: LM Studio, vLLM, LocalAI, and any OpenAI-compatible endpoint.

Setup

1. Install dependencies:

   npm install
   

2. Launch in VS Code Extension Host: Press F5 or run the "Run Extension" debug configuration.

3. Configure your provider:

   • Open the CodeRun sidebar (activity bar icon)
   • Go to Settings (⚙️)
   • Select Provider, enter Base URL and API Key if needed
   • Click "Refresh Models" then select a model
   • Click "Save Settings"

4. Start coding! Select a chat, choose a model, and ask anything.

Settings

All settings under the coderun.* namespace:

API keys are stored securely using VS Code's SecretStorage — never in plain text.

Architecture

src/
  extension.js          — VS Code extension entry point
  agent.js              — Public agent API (thin wrapper)
  agentLoop.js          — Core reasoning loop (Think → Plan → Act → Verify)
  promptBuilder.js      — Assembles prompts from system + history + tools + memory
  providerManager.js    — Provider factory (creates the right provider)
  providerOllama.js     — Ollama API provider
  providerOpenAI.js    — OpenAI API provider
  providerAnthropic.js  — Anthropic Claude provider
  providerGemini.js     — Google Gemini provider
  providerOpenRouter.js — OpenRouter provider
  providerCompatible.js — Generic OpenAI-compatible provider
  toolRegistry.js       — Maps tool names to implementations
  toolDefinitions.js    — Auto-registering tool schemas for LLM
  toolExecutor.js       — Formats tool results for LLM context
  tools.js              — All tool implementations (file, dir, terminal, etc.)
  conversationStore.js  — Conversation persistence
  settingsStore.js      — Settings persistence
  workspaceContext.js   — Workspace info provider
  memoryManager.js      — Memory management (placeholder)
  skillsManager.js      — Skills management (placeholder)
  mcpManager.js         — MCP support (placeholder)
  terminalManager.js    — VS Code terminal integration
  terminalRenderer.js   — Terminal output in chat
  permissions.js        — Permission system for dangerous actions
  events.js             — Event bus for loose coupling
  constants.js          — All constants, defaults, system prompt
  utils.js              — Small utility functions
  config.js             — VS Code configuration reader
  Dashboard.js          — Frontend dashboard shell
  ChatSpace.js          — Frontend chat UI
  MarkdownRenderer.js   — Markdown to HTML with syntax highlighting
  Dashboard.css         — Dashboard styles
  ChatSpace.css         — Chat styles

Agent Loop

User Request
    ↓
Prompt Builder (system + history + workspace + tools + memory)
    ↓
Provider.chat() — model decides
    ↓
Tool Calls? → YES → Permission Check → Execute Tool → Append Result
    ↓                                          ↓
    NO ←———————————————————————————————————————
    ↓
Final Response

Max 20 iterations. If reached, asks: "Maximum agent iterations reached (20). The task may not be complete. Do you want me to continue?"

Keyboard Shortcuts

• Ctrl+L — New chat
• Enter — Send message
• Shift+Enter — New line in input

Safety

• All file paths are resolved relative to workspace root
• Path traversal attacks are blocked
• Dangerous actions require explicit permission
• API keys stored in VS Code SecretStorage (encrypted)
• No telemetry, no cloud backends, complete privacy with local providers

License

MIT