xcode-copilot-server

A proxy API server that lets you use GitHub Copilot in Xcode, either as a custom model provider or as the backend for Claude Agent and Codex Agent.

• Why
• Quick start
• Configuration
  • Config file
  • Settings
• Agent skills
• Launchd agent
• CLI reference
• Security
• Testing
• License

Why

[!WARNING] Most other Copilot proxy servers (whether remote or local) work by reverse-engineering private Copilot APIs and spoofing the user-agent to impersonate VS Code. GitHub's abuse detection flags this kind of usage and can restrict or suspend your Copilot access. The proxy projects themselves warn about this risk. You should avoid using any of those servers. This server uses the official Copilot SDK, which is the sanctioned way to build on top of Copilot.

Xcode 26 added support for third-party LLM providers, but it only supports ChatGPT and Claude out of the box. If you have a GitHub Copilot subscription, there's no built-in way to use it.

This server wraps the GitHub Copilot SDK and exposes it as an API that Xcode can talk to. It's built on copilot-sdk-proxy, which handles the SDK integration and protocol translation. Integration tests run against llm-mock-server. It supports three providers:

• OpenAI (default): Exposes an OpenAI-compatible completions API so Xcode can use Copilot as a custom model provider. Xcode handles tool execution directly.
• Claude: Exposes an Anthropic-compatible API so Xcode can use Copilot as the backend for Claude Agent. A built-in tool bridge intercepts tool calls and routes them back to Xcode for execution.
• Codex: Exposes an OpenAI Responses-compatible API so Xcode can use Copilot as the backend for Codex Agent. Same tool bridge as Claude.

In OpenAI mode, the server also connects to Xcode's built-in MCP tools (via xcrun mcpbridge), giving Copilot access to your project's build logs, indexes and other context. This requires Xcode 26.3 or later. Claude and Codex handle MCP internally through their own agents.

Quick start

You need Node.js 22 or later and a GitHub Copilot subscription.

1. Authenticate with one of these (you only need one):

copilot login # Copilot CLI
gh auth login # GitHub CLI

Or set a GITHUB_TOKEN environment variable with a valid fine-grained Copilot access token.

2. Install:

npm install -g xcode-copilot-server

3. Start the server:

xcode-copilot-server

The server runs in auto mode by default, which registers all three providers and auto-patches Claude and Codex settings. Pick the provider you want to use in Xcode and follow the setup below.

[!TIP] If you only need one provider, use --proxy to run a single provider instead (e.g. xcode-copilot-server --proxy claude). See CLI reference for details.

[!TIP] If you want to run the server in the background with automatic start/stop, see Launchd agent below.

Configuration

Config file

The server reads its configuration from a config.json5 file. By default, it uses the bundled one, but you can point to your own with --config:

xcode-copilot-server --config ./my-config.json5

The config file uses JSON5 format, which supports comments and trailing commas. In auto mode, all three provider sections are active. In single-provider mode (--proxy), only the specified section is used:

{
  openai: {
    // No tool bridge needed, as Xcode drives tool execution directly.
    toolBridge: false,
    // Tool bridge timeout in minutes. 0: disabled.
    toolBridgeTimeout: 0,

    mcpServers: {
      xcode: {
        type: "local",
        command: "xcrun",
        args: ["mcpbridge"],
        allowedTools: ["*"],
      },
    },
  },

  claude: {
    // Intercepts tool calls and forwards them to Xcode so Claude Agent
    // drives tool execution through the IDE instead of the Copilot CLI.
    toolBridge: true,
    // Tool bridge timeout in minutes. 0: disabled.
    toolBridgeTimeout: 0,

    // No MCP servers needed, as Claude Agent handles tools natively.
    mcpServers: {},
  },

  codex: {
    // Same as Claude: intercepts tool calls and forwards them to Xcode
    // so Codex drives tool execution through the IDE.
    toolBridge: true,
    // Tool bridge timeout in minutes. 0: disabled.
    toolBridgeTimeout: 0,

    mcpServers: {},
  },

  // Built-in CLI tools allowlist.
  // ["*"] to allow all, [] to deny all, or a list of specific tool names.
  //
  // Empty by default so Xcode can handle all operations (search, read, edit)
  // through its UI. Enabling CLI tools lets the Copilot session perform
  // those operations directly, bypassing Xcode.
  allowedCliTools: [],

  // Maximum request body size in MiB.
  bodyLimit: 10,

  // Server-level request timeout in minutes. 0: disabled.
  requestTimeout: 0,

  // Filename patterns to filter out from search results in the prompt.
  //
  // Xcode can include full file contents for every search match, so add patterns
  // here to strip files that bloat the prompt (e.g. ["mock", "generated"]).
  excludedFilePatterns: [],

  // Reasoning effort for models that support it: "low", "medium", "high", "xhigh"
  reasoningEffort: "xhigh",

  // Auto-approve permission requests.
  // true to approve all, false to deny all,
  // or an array of kinds: "read", "write", "shell", "mcp", "url"
  autoApprovePermissions: ["read", "mcp"],
}

Settings

In auto mode, Claude and Codex settings are patched on startup and restored on shutdown. For Claude, the server creates (or updates) settings.json at ~/Library/Developer/Xcode/CodingAssistant/ClaudeAgentConfig/ to point to the server. For Codex, OPENAI_BASE_URL and OPENAI_API_KEY are set via launchctl setenv so Xcode (and any Codex process it spawns) can reach the server.

OpenAI doesn't need any settings patching since it uses Xcode's built-in locally hosted provider support.

If you'd rather configure settings yourself instead of using auto-patch, see the manual steps below.

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:8080",
    "ANTHROPIC_AUTH_TOKEN": "12345"
  }
}

xcode-copilot-server patch-settings --proxy claude --port 8080
xcode-copilot-server restore-settings --proxy claude

launchctl setenv OPENAI_BASE_URL http://localhost:8080/v1
launchctl setenv OPENAI_API_KEY 12345

launchctl unsetenv OPENAI_BASE_URL
launchctl unsetenv OPENAI_API_KEY

xcode-copilot-server patch-settings --proxy codex --port 8080
xcode-copilot-server restore-settings --proxy codex

Agent skills

Agent skills are an open standard for adding custom instructions and file context to AI coding agents. All three providers support them through the Copilot SDK session, and each agent has its own skill paths:

Launchd agent

Instead of starting the server manually every time, you can install it as a launchd agent. This uses macOS socket activation, so the server starts automatically when something connects to the port (e.g. when Xcode sends its first request) and you don't need to keep a terminal open.

Installing

xcode-copilot-server install-agent

This writes a plist to ~/Library/LaunchAgents/ and loads it with launchctl. The agent is set up with socket activation on the specified port, so launchd owns the socket and starts the server on demand. Settings are patched at install time.

The install-agent subcommand accepts the same options as the main command (--port, --proxy, --log-level, --config, --cwd), plus --idle-timeout which defaults to 60 minutes for the agent. After 60 minutes with no requests, the server shuts itself down. The next incoming connection will start it again automatically.

Server logs go to ~/Library/Logs/xcode-copilot-server.out.log and ~/Library/Logs/xcode-copilot-server.err.log.

Uninstalling

xcode-copilot-server uninstall-agent

This unloads the agent, deletes the plist, and restores any patched settings if the agent was installed with --auto-patch.

How it works

Launchd creates a socket on the configured port and waits. When a connection comes in (e.g. Xcode sends a request), launchd starts the server process and hands over the socket. The server handles requests as normal.

If the server crashes, launchd doesn't restart it immediately, but the next incoming connection will start a fresh process. If --idle-timeout is set (defaults to 60 minutes for the agent), the server exits after that many minutes of inactivity, and launchd will start it again on the next connection.

To check if the agent is loaded:

launchctl list | grep xcode-copilot-server

CLI reference

xcode-copilot-server [options]

Options:
  -p, --port <number>            Port to listen on (default: 8080)
  --proxy <provider>             API format: openai, claude, codex (default: auto)
  -l, --log-level <level>        Log verbosity (default: info)
  -c, --config <path>            Path to config file
  --cwd <path>                   Working directory for Copilot sessions
  --auto-patch                   Auto-patch settings on start, restore on exit (implicit in auto mode)
  --idle-timeout <minutes>       Shut down after N minutes of inactivity (default: 0, disabled)
  -v, --version                  Output the version number
  -h, --help                     Show help

Commands:
  patch-settings                 Patch provider settings and exit (all providers unless --proxy is set)
  restore-settings               Restore provider settings from backup and exit
  install-agent                  Install a launchd agent with socket activation
  uninstall-agent                Uninstall the launchd agent and restore settings

By default, the server runs in auto mode and registers all providers. Use --proxy to run a single provider instead:

Security

This server acts as a local proxy between Xcode and GitHub Copilot. It runs on your machine and isn't meant for the internet or shared networks.

• The server binds to 127.0.0.1, so it's only reachable from your machine. Incoming requests are checked for expected user-agent headers (Xcode/ for OpenAI and Codex, claude-cli/ for Claude), which means casual or accidental connections from other tools will be rejected. This isn't a strong security boundary since user-agent headers can be trivially spoofed, but it means only the expected client talks to the server under normal use.

• The bundled config sets autoApprovePermissions to ["read", "mcp"], which lets the Copilot session read files and call MCP tools without prompting. Writes, shell commands, and URL fetches are denied by default. You can set it to true to approve everything, false to deny everything, or pick specific kinds from "read", "write", "shell", "mcp", and "url".

• MCP servers defined in the config are spawned as child processes. The bundled config uses xcrun mcpbridge, which is an Apple-signed binary. If you add your own MCP servers, make sure you trust the commands you're configuring.

• When you use install-agent, the generated plist file includes your PATH in cleartext so the agent can find Node.js. The file is written to ~/Library/LaunchAgents/ which is only readable by your user account by default.

[!NOTE] If you authenticated with a GITHUB_TOKEN environment variable, that token is also embedded in the plist. If you'd rather not have a token in the plist, use gh auth login or copilot login instead.

License

MIT License

Copyright (c) 2026 Suyash Srijan

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.