Add other agent support

[Norus6]

Hey, I see there are several pull requests for implementing other agents. As a codex user myself, it would be nice to implement codex, or just being agent agnostic.

[pablodelucca]

Hi @Norus6! This is one of the project's top priorities. I've seen the PRs and indeed many are trying to add support for other agents, but most aren't truly agnostic. To avoid constant rewrites and allow for a futureproof solution, we're looking for a solution that's genuinely plug-and-play across different agents.

From what I've seen so far:

• @Drepheus has built an agnostic implementation, though it currently only supports OpenCode and basic VS Code Terminals.
• @Akasha53 seems to be working on something that could cover Claude, Codex, and Gemini, but it looks like it's still in progress.
• @mexicorea and @78i87 have made Codex-specific implementations.
• @sryku2000 has made a GitHub Copilot CLI support

To be honest, I'm a Claude Code user myself, so I haven't been able to test all of these setups thoroughly.

@Drepheus @Akasha53 @mexicorea @78i87 @sryku2000 it'd be great if you could weigh in here. Since you've all been tackling this from different angles, it might be worth comparing approaches and seeing if there's a path toward a unified, truly agent-agnostic solution. Would love to hear your thoughts!

[sryku2000]

It’s great to see so many different approaches to agent integration. To achieve the “truly agent-agnostic” goal @pablodelucca mentioned, I’d like to propose a structured path forward based on my experience with the GitHub Copilot CLI implementation.

I believe we should focus on Decoupling the UI from the execution logic through an abstraction layer. Here’s a 3-step plan:

1. Capability Discovery (Settings UI)
   Instead of hardcoding supported agents, we can provide a Settings toggle for users to enable/disable the Agent CLIs they have installed locally (e.g., Gemini CLI, Claude Code, GitHub Copilot).

This ensures the UI only offers functional options to the user.

Users can manage their own CLI paths or environment variables here.

2. Dynamic Agent Spawning
   When clicking the + Agent button, the UI should present a list of "Enabled Providers" from the settings.

The user selects the platform (e.g., "I want a Claude Agent for this task").

The system then initializes that specific agent instance and runs the corresponding CLI command in the background.

3. The "Provider" Abstraction (Core Logic)
   We can implement an Abstract Class (or Interface) to standardize how we interact with different CLIs. Each new agent type just needs to implement these core methods:

TypeScript
abstract class BaseAgentProvider {
abstract readonly type: AgentType; // e.g., 'gemini', 'claude', 'github-copilot'

// Standardizing the lifecycle
abstract spawn(options: SpawnOptions): Promise;
abstract sendMessage(message: string): Promise;

// Session management: each provider handles its own log/session folder logic
abstract readSessionData(sessionId: string): Promise;
abstract archiveSession(sessionId: string): void;
}
Why this works:

Scalability: Adding a new agent (like a future "OpenAI-Agent") becomes as simple as creating a new class that extends BaseAgentProvider.

Maintainability: We won't need to rewrite the UI every time a CLI update changes its output format; we just update that specific provider class.

[pablodelucca]

I like the abstraction, but just gotta make sure this translates properly to different kinds of agents. We'll be looking into it soon!

[MikaelDDavidd]

Hey! I built a standalone server that runs Pixel Agents outside VS Code — in any browser.

This solves the problem for users who use Warp, iTerm, Hyper, or any terminal that isn't VS Code's integrated terminal.

How it works

• Node.js server (Express + WebSocket) on port 3333
• Scans ~/.claude/projects/ every 5s for active JSONL transcript files
• Serves the same React webview with a WebSocket adapter replacing postMessage
• Detects Claude Code sessions from any terminal automatically
• Full asset loading (characters, furniture, floors, walls)
• Layout shared with VS Code extension

Quick start

cd standalone && npm install && bash start.sh
# Open http://localhost:3333

PR: #166

The approach could be extended to support other agents (Codex, Cursor, etc.) — the scanner just needs to know where they write their logs.