Azure AI Agent for VS Code

An agentic coding assistant that runs against your Azure AI Foundry / Azure OpenAI deployment (for example gpt-5.4). It can inspect and modify your workspace, run commands, browse the web, use MCP tools, accept image input, and stream results live inside VS Code.

The extension builds with a single tsc run and ships as a small .vsix with no runtime dependencies.

Features

• Agent loop with tool calling
  • Workspace tools: read_file, write_file, edit_file, apply_patch, delete_file, list_directory, glob_files, search_text, find_symbols, get_diagnostics, get_workspace_info, get_definition, get_references
  • Execution tools: run_command, run_task, spawn_agent
  • Web tools: web_search, web_fetch
  • Memory tools: read_memory, write_memory
  • MCP tools: discovered dynamically from configured MCP servers
• Streaming responses — assistant text streams token-by-token.
• Diff-preview approval — write_file, edit_file, and apply_patch can open VS Code's diff editor before changes are applied.
• Multimodal input — paste, drag-drop, or pick image files and send them with prompts.
• Workspace instructions — automatically loads .azure-ai-agent/instructions.md, AGENTS.md, or CLAUDE.md into the system prompt.
• Run history — stores previous runs per workspace and lets you replay them.
• Token usage and cost tracking — shown in the status bar and available via /usage.
• @mention files — fuzzy-search workspace files from the chat input.
• Slash commands — /help, /clear, /cancel, /usage, /settings, /key, /setup, /searchkey, /mcp.
• SecretStorage for keys — Azure API key and web-search API key can be stored securely in VS Code SecretStorage.
• Approval controls — global approval mode plus MCP-specific approval mode.
• Compact mode — denser chat layout that persists in workspace settings.
• Modern settings UI — guided setup, health checks, diagnostics, presets, profile import/export, MCP config editing.
• MCP support
  • Launch stdio MCP servers from settings
  • Discover MCP tools dynamically
  • Invoke MCP tools from the agent
  • Inspect MCP status and discovered tools from commands/UI

Quick start

1. npm install
2. Press F5 to launch an Extension Development Host, or run npm run package to build a .vsix.
3. In VS Code, open Azure AI: Open Settings UI.
4. Set:
   • Azure endpoint
   • deployment name
   • API version
   • Azure API key
5. Open chat with Ctrl+Shift+A.

Commands

Settings highlights

• azure-ai-agent.endpoint — Azure OpenAI / AI Foundry endpoint URL
• azure-ai-agent.deployment — deployment name
• azure-ai-agent.apiVersion — REST API version
• azure-ai-agent.approvalMode — auto / destructive / always
• azure-ai-agent.mcp.approvalMode — MCP-specific approval mode
• azure-ai-agent.diffPreview — show diff editor before writes
• azure-ai-agent.streamResponses — token streaming
• azure-ai-agent.maxIterations — max tool-loop steps
• azure-ai-agent.instructionsFiles — files searched for workspace instructions
• azure-ai-agent.webBrowsing.* — browsing controls
• azure-ai-agent.webSearch.* — search provider and key
• azure-ai-agent.safety.* — secret scrubbing and command auto-approval patterns
• azure-ai-agent.ui.compactMode — compact chat layout
• azure-ai-agent.mcp.servers — configured MCP stdio servers

MCP server configuration

Example:

[
  {
    "id": "my-server",
    "command": "node",
    "args": ["./path/to/server.js"],
    "cwd": ".",
    "env": {
      "EXAMPLE_TOKEN": "value"
    },
    "enabled": true
  }
]

Configured MCP tools are exposed to the agent dynamically.

Build & package

npm install
npm run compile
npm run package
npm test

This produces a .vsix such as:

azure-ai-agent-1.1.0.vsix

Publishing to the VS Code Marketplace

Publisher: CRJTECH

One-time setup:

npm install -g @vscode/vsce
vsce login CRJTECH

Then publish:

npm run compile
vsce package
vsce publish

Or publish the built package directly:

vsce publish --packagePath azure-ai-agent-1.1.0.vsix

Repository

Current remote:

• https://github.com/crjtech3-spec/agent-vs.git

If you want to publish this as a brand-new GitHub repository instead, create the new empty repo first, then update the remote and push.