What Is Panda Code and How Does It Run Multiple Coding Agents?

What is Panda Code?

Panda Code is a free, open-source desktop app that puts multiple AI coding agents in one window. It runs Claude Code, Codex, Kimi Code, Open Code, GLM, Mimo, Grok Build and Antigravity from a single interface. It does not sell a model or a plan. It drives the CLI tools you already installed, using your own logins, and adds projects, Git, a terminal, a browser, a task board, agent teams and chat bots around them.

Why does a GUI for coding agents matter?

Running several agents means several terminals. One waits on an approval you did not see. Another finished ten minutes ago. Afterward you cannot tell which agent touched which file. Panda Code was built by a YouTube creator who ran into exactly this while using Claude Code, Codex, Kimi Code and others side by side.

The app puts the sidebar (projects and threads) on the left, the chat in the middle and the prompt box at the bottom. Approvals show up in a thin bar above the prompt box, one line per request, with deny and approve buttons. If several pile up, you get a single "approve all" line.

How do you install Panda Code?

Download the right file from the project's GitHub releases page, then install Git and at least one coding agent and log into it.

The builds are not signed yet, which is why the OS warns you. After setup the app updates itself. It checks in the background and installs a new version only when no agent is working.

How does Panda Code work with different agents and models?

A model picker in the prompt box lets you choose the agent and model: Claude Code with Opus or Sonnet, Codex with GPT models, or Kimi, Open Code, GLM, Mimo, Grok Build and Antigravity. The list is searchable, shows each model's context size, and lets you set reasoning effort. Each task remembers its own model, and a new task starts with the one you used last.

GLM and Mimo work by pasting an API key into settings. Panda Code then runs Claude Code with that key. Keys sit in your system keychain and each one goes only to its own provider.

Other prompt-box details:

• Type @ to search and attach project files.
• Type / at the start to get the agent's slash commands and skills, so existing Claude Code skills still work.
• Agent steps (reading, editing, running commands) collapse into a small row. Click it to see every step.
• Press Enter while an agent works to queue a message. Press Command Enter to steer the run immediately. Steering works with Claude Code and Codex. Other agents queue instead.

What are the permission modes?

Panda Code uses each agent's built-in modes rather than adding its own layer.

How do worktrees let several agents share one project?

A worktree is a second copy of your project on its own branch. A bar above the prompt box shows the project, folder and branch. Click the branch name and move the task into a new worktree, and that agent works in isolation from the others.

When it finishes, choose "merge into main" from the same menu. Panda Code merges the saved commits into main on your machine. If you have uncommitted changes or there is a conflict, it stops and changes nothing. Afterward it offers to delete the worktree.

Each turn's changes are saved, so you can open a diff per turn (side by side or in one column) instead of only seeing the branch total. A changes panel handles Git status and commits, and Command J opens a real terminal in the task's folder.

An optional GitHub integration uses the gh tool. It shows repos, pull requests and issues, and each action (review a PR, fix failing checks, merge, solve an issue) opens as a side thread you can keep chatting in.

How does the task board work?

Click "board" and you get four columns: pending, in progress, review and completed. Add a task with a title and description, pick an agent, and press start. When the agent finishes, the task moves to review. Completed tasks are archived after 7 days. The board spans all your projects, and any card opens in a side panel for chatting.

In the sidebar, each thread takes one line under its project, with status for running, needs-you and finished. You can pin and drag threads, and Command F searches within a long thread.

What are agent teams?

Click the team button in the prompt box and you get an architect plus up to three workers. The architect splits the job into parts and hands them out. Workers run at the same time. When they finish, the architect checks the work and gives you a summary and a list of changed files.

You can mix vendors. The architect could be Claude Opus, one worker Codex, another Kimi or GLM. Named teams can be saved in settings, such as a cheap team with GLM workers or a strong team with Opus and Codex, and selected from the model picker. Codex also supports a /goal command, where it keeps working turn after turn toward a stated goal, optionally with a token budget.

What can the bots do?

Bots connect an agent to Telegram, Slack, Discord, WhatsApp or a local web chat. You message from your phone and the work happens on your computer. Each bot has its own profile, memory, schedules, channels and rules. One example is a bot that posts a repo summary every morning. Every bot conversation is also a normal task in the app, so you can open it on desktop and continue.

What does the built-in browser and device support add?

The browser opens beside the chat. Its start page lists every local server and port. You can pick an element on a page and attach it to your message, or screenshot the page for the agent. Click a file name in an answer to open it, and HTML, SVG, image and PDF files open in the browser pane.

A pandacode run.sh file in your project adds a run button that starts the app in its own terminal and opens it in the browser. The agents can write and maintain that file. For mobile work, the iOS simulator and Android emulator appear next to the chat, and agents can control the phone to test what they built.

Beyond that, the app can run agents in macOS and Linux virtual machines on Apple-silicon Macs (using the free Tart tool), or in Daytona, Cloudflare and AWS sandboxes. A deploy menu starts tasks for Vercel, Cloudflare or AWS using their CLIs and your login, with each command needing your approval.

How does it handle usage limits and tokens?

An activity view shows limits for each installed agent, such as Claude's 5-hour and weekly limits and Codex limits, plus reset times. It also estimates what your usage would cost at API prices over a week, month or year. That figure is only an estimate. If you hit a limit mid-task, "resume at reset" continues automatically.

Per-task "token saver" styles (caveman, lean code, quiet tools) shorten agent answers. The instruction is sent only when you change it. The creator also reports that follow-up responses from Claude Code start in about 4 seconds instead of about 8 in his tests, because the next process is started in the background.

Is Panda Code really lighter than other options?

According to the creator's tests, the window opened in about a third of a second and used 130 to 160 MB when idle. He measured T3 Code at about 3 seconds and about 700 MB. The app uses the web view already on your computer, ships no copy of Chrome and runs no background Node server. These are single-developer measurements on his hardware, so check on your own machine.