Build an agent harness that runs every AI model in one app

I'm building my own ADE — an agentic development environment — for how I work. I run agents for everything: writing, content, and coding. I want one app where all of them live. This is a fresh repo; my hand-drawn mockup is at mock/IMG_7232.jpeg — use it as the layout reference.

Start from the Superset codebase (github.com/superset-sh/superset) — it's the closest to this. Pull the memory system from Hermes (github.com/NousResearch/hermes-agent). I like how Conductor (conductor.build) does its tabs, but it's not open source — so I'm giving you screenshots of it. Match the tab layout and feel from those.

Here's how Superset and Conductor work — they're nearly the same — and what I want to change:

They each have a left rail of workspaces/projects, each one a repo in its own git worktree. Inside a workspace you get tabs, and each tab is a terminal pane where you talk to a CLI coding agent (Claude Code, Codex), plus a changes/diff panel on the right. Tabs can auto-launch an agent on open.

Keep that terminal-tab core, but restructure it:
- Make the left rail two levels: categories at the top (each with a profile photo) — like YouTube, or a coding project — and under each category, named agents (each with a name + profile photo). An agent is basically Superset's "workspace" — its own repo/worktree and its own skills.
- When I select an agent, the tabs across the top are sessions of THAT agent — each tab a terminal running the same agent, so I can have three going at once. NOT split by model the way Superset does.
- Keep the terminal panes (chat with the agent right in the terminal) and the changes panel.
- Agents run Claude Code and Codex by default; let me connect OpenRouter for open-source models.
- Give each agent a Hermes-style memory system out of the box (MEMORY.md / USER.md + providers).

Then onboarding (reuse the new-project flow): a new user creates a category and picks a profile photo, then adds agents under it (name + photo), and that builds the rails on the left and tabs on top. Wrap it in a single one-command install.

For the build itself, use the fable-orchestration skill in this repo (.claude/skills/fable-orchestration) — you're the architect: plan and direct, and hand the actual coding to Opus 4.8 agents. But start with the UI only — before you build anything real, give me a clickable mockup of this layout (the two-level rail, the agent tabs, a terminal pane) so I can react to it. Don't wire up the backend yet.

Okay, this isn't bad. However, there's some things that are just superfluous, like setup, run, terminal — we don't need. Of course, get rid of those emojis. Absolutely no emojis, ban emojis from everything. What I want instead of emojis is an ability to literally upload a profile photo for each agent.

Agent files, that's a good touch — all files is good. You haven't actually connected this, I'm assuming just because it's a mock-up, but it's not actually connected to the CLI, right? It needs to be connected to the CLI so we can access all of this, whether it's Claude Code or Codex or OpenCode or any of those.

We don't need the path at the bottom — that seems unnecessary. We don't need choosing the model picker. It's literally just terminal windows that you can pull open.

And when I go light mode, it doesn't change the actual chat window color — because it's not a true terminal. You're going to have to factor that in with light mode as well, and figure out how to create the terminal wrapper here and start Claude Code (claude --dangerously-skip-permissions), Codex (bypass approvals), etc.

All approved. Now build this out into a proper Mac application (Electron). Migrate the mockup over and run the install yourself when it’s ready.

All right, this is looking good. Everything is approved.

Now what I want to do is have an ability to add my ChatGPT subscription as well, and log into — I guess it would technically be Codex — and use GPT-5.5 within this, as well as any open source model.

I don't know exactly how we can set it up with these different open source models. I'm assuming the easiest way is through an OpenRouter API key.

But I want an ability to, when I open up the application and press that new tab button, below the tabs there should be an ability to select: Claude, of course, then the OpenAI logo, then some popular open source models like Kimi K2, MiniMax M3, and GLM 5.2, with their logos as well. So I can just click those and it will automatically open a session with that model.

To get those open source models, the easiest way is to use an OpenRouter API key and we can just select all of these. I'll let you be the judge of all of this. Go ahead and spawn agents and do this next.

Perfect — all approved.

Now what I need is for us to build out agents files: agents.md, a memory.md, and anything else that Hermes does.

I need you to go through the Hermes GitHub repo, determine how they have their agents self-improving, and then borrow Hermes and bring it right over.

Make this a public GitHub repo people can download. Strip ALL secret keys, organize it, write a walkthrough guide, and package it as a signed Electron DMG. Keep the same simple onboarding: create categories, create agents, add profile photos.