I built the fastest & lightest AI coding agent GUI - 12MB, 4X FASTER, Fully Opensource!

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: I built the fastest & lightest AI coding agent GUI - 12MB, 4X FASTER, Fully Opensource! · Channel: AICodeKing · 15:27 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI developer tools, Privacy and data protection

Summary

Panda Code is a free, open-source desktop app that brings multiple coding agents, project files, Git tools, a terminal, a browser, and task management into one window. The video tours features including agent permissions and approvals, file diffs, work trees, GitHub integration, task boards, agent teams, bots, virtual machines, and deployment tools. The app uses coding agents and logins users already have, stores data locally, and is presented as lightweight, with no account required. The creator also explains installation, privacy, token-saving features, limitations, and how to build the app from source.

Key points

Panda Code runs multiple coding agents in one desktop app and is free and open source.
It uses coding tools and subscriptions the user already has, without selling its own model or plan.
The app includes project and Git views, a terminal, a browser, task boards, and agent teams.
Users can set agent permission modes, approve requests, queue prompts, and inspect changes by turn.
Work trees let agents work on separate copies and branches before changes are merged.
The task board, GitHub view, and agent teams support managing and coordinating multiple tasks.
Bots can connect to messaging services, while virtual machines and cloud sandboxes provide separate environments for agents.
The app stores data locally, has no accounts or tracking, and keeps keys in the computer’s keychain.
The apps are unsigned, and simulator and virtual-machine features are mostly for Macs.

Chapters

Start	Chapter	Summary
00:00	Introducing Panda Code	The creator introduces Panda Code as a lightweight, open-source app for managing coding agents and projects in one window.
01:38	Installation and getting started	The video explains where to download the app, how to handle first-launch warnings, and how to set up Git and an agent.
02:08	Projects, models, permissions, and chat	The creator tours the interface and shows model selection, permission modes, prompts, approvals, and turn-by-turn file changes.
05:53	Files, browser, attachments, and work trees	This section covers Git changes, the built-in browser and app previews, attachments, mobile testing, and isolated work trees.
08:32	GitHub, task boards, teams, and bots	The creator demonstrates GitHub workflows, task management, collaborative agent teams, Codex goals, and messaging bots.
11:12	Sandboxes, usage tools, and shortcuts	The video covers virtual machines, cloud sandboxes, deployment, usage estimates, token-saving options, updates, and the quick overlay.
13:56	Privacy, open source, and limitations	The creator describes local data storage and the MIT license, explains how to build the app, and notes its current limitations.

Quotes

It runs all your coding agents in one window. — 00:31

It uses the web view that your computer already has. — 01:05

The agent changes direction without stopping. — 05:22

Pandacode stores everything on your computer in a local database. — 13:56

Concepts and entities

Panda Code, Claude Code, Codex, T3 Code, GitHub, Tart, Daytona, Cloudflare, AWS, Jev

Full transcript (original language)

00:00 [music] >> Hi. Welcome to another video. So, today I want to show you something that I made. It’s called Panda Code. It’s a desktop app for coding agents. If you watch my channel, you know that I use a lot of coding agents. Claude Code, Codex, Kimmy Code, Open Code, GLM, Mimo, Grok Build, Anti-Gravity, all of them. But, each one runs in its own terminal. So, I always have five terminals open. One agent waits for an approval that I didn’t see. Another one finished 10

00:31 minutes ago and I don’t know which agent changed which file. Panda Code solves this. It runs all your coding agents in one window. You also get your projects, Git, a terminal, a browser, a task board, agent teams, and bots in the same window. And, it’s free and open source. You don’t need an account or a license key. It doesn’t collect any data about you, which is just amazing. Now, one important thing. Panda Code doesn’t sell you a model or a plan. It uses the coding tools that you already installed with your own logins. So, if you pay for Claude, ChatGPT, Kimmy, or the GLM coding plan, Panda Code uses that plan.

01:05 You don’t pay anything extra. If you ask me, that’s how it should be. It’s also very light. It uses the web view that your computer already has. So, it doesn’t ship its own copy of Chrome. It doesn’t run a node server in the background, either. In my tests, the window opened in about a third of a second. It used about 130 to 160 megabytes of memory when idle. I tested T3 Code, too. Its window took about 3 seconds to open and it used about 700 megabytes. So, this is kind of great. So, let’s get right into it. First, let’s install it. Go to the Panda Code GitHub page. The

01:38 link is in the description. Open the releases page. There, you get three files. A DMG for Macs with Apple chips, an installer for Windows, and an app image for Linux. The apps are not signed yet. So, your computer shows a warning the first time. On a Mac, open system settings. Go to privacy and security. Then, click open anyway. On Windows, you may see a blue SmartScreen box. Click more info, then click run anyway. On Linux, run chmod +x on the app image, then run the app image. Next, install Git, then install

02:08 at least one coding agent and log into it. For example, install Claude code and log in like you normally do. And that’s it. You’re good to go. Panda code also updates itself. It checks for new versions in the background. It installs a new version only when no agent is working. So, you only download it once. Now, let’s look at the app. On the left, you have the sidebar. It shows your projects and your threads. In the middle, you have the chat. At the bottom, you have the prompt box. You can also see these anime wallpapers. There are 21 of them. Each day of the week gets a morning scene, an afternoon scene, and a night scene. And the app’s main color changes to match the

02:41 wallpaper. I think that’s super cool. You can turn the wallpapers off in settings. You can also import your VS Code color theme. So, let’s open a project. Click open project and pick a folder. Then click new task or press command N. A new page opens and asks, “What should we build?” The prompt box is ready for you to type. In the prompt box, you see the model picker. Here, you pick the agent and the model. You can pick Claude code with Opus or Sonnet. You can pick Codex with the GPT models. You can also pick Kimiko, Open Code, GLM, Mimo, Grok Build, and Antigravity. You can search the list.

03:14 It shows how much context each model has. You can also set how hard the model thinks. Each task remembers its own model. A new task starts with the model you used last. For GLM and Mimo, you just paste your API key in settings. Panda code then runs Claude code with that key. Your keys stay in your computer’s keychain. Panda code sends each key only to its own provider. So, if you have the cheap GLM coding plan, this is a really good option, for sure. Next to the model picker, you can set the permission mode. There are four modes. Standard means the agent asks you before it does risky things. Accept

03:48 edits means the agent can edit files on its own, but it still asks before it runs commands. Plan mode means the agent can only read. It looks at your code and makes a plan. Full access means the agent can do everything without asking. Panda Code uses each agent’s own built-in modes for this. It doesn’t add a fake layer on top. Now, let me send a prompt. I’ll type “Add a dark mode switch to the settings page.” If you want to point at a file, type the at sign. Panda Code searches your project files. Use the arrow keys to pick a file and

04:20 press enter to add it. If you type a slash at the start, you get the agent’s slash commands and skills. So, your Claude Code skills work here, too. Let’s send it and see. Okay. The answer comes in live with clean formatting. The agent’s steps, like reading files, editing files, and running commands, stay inside this small work for row. So, the chat stays clean. You only see the parts that matter. To see every step, just click the row. Sometimes the agent needs your approval. The request appears in a thin bar right above the prompt box.

04:51 Each request is one line with deny and approve. If there are many requests, you get one line with approve all, so you never miss an approval. That was my biggest problem with terminals. Now, here’s something really useful. You can type while the agent work. If you press enter, Panda Code puts your message in a queue. It sends the message when the agent finishes. You can see the queue above the prompt box. You can edit or remove any message in it. If you press command enter, you steer the agent instead. Your message goes into the current run right away.

05:22 The agent changes direction without stopping. This works with Claude Code and Code X. Other agents put the message in the queue. This is pretty amazing, to be honest. Sometimes the agent asks you a question. For example, which option do you want? You see a small card with the options. Just click one. Now, let’s check the agent’s work. Panda Code saves the changes from every turn. Below each answer, you see the files that turn changed. Click a file to open the diff. You can see the changes side by side or in one column with colors for the code. So, you see exactly what each turn changed, not

05:53 only the total for the whole branch. The changes panel shows your Git status. You can look at diffs, pick files to commit, and make the commit right here. Press command J to open a real terminal in the tasks folder. When the agent names a file in its answer, you can click the file name to open it. In the files panel, you can click an HTML file, an SVG, an image, or a PDF. It opens in the built-in browser. Yes, there’s a browser built in. It opens next to the chat. The first page lists every local server on your computer with its port.

06:24 So, if your app runs on port 3000, you see it right there. You can pick an element on the page and add it to your message. You can also take a screenshot of the page and send it to the agent. So, if a button looks wrong, select it and say, “Fix this.” Chef’s kiss. There’s also a run button. Your project can have a file called pandacode run.sh. When it does, pandacode shows a run button at the top. Click it, and pandacode starts your app in its own terminal. Then it opens your app in the browser. And you don’t need to write this file yourself. Pandacode asks the agents to write it and keep it up to

06:57 date. If you make mobile apps, pandacode supports the iOS simulator and the Android emulator. The phone appears on the right side next to the chat. You can tap through your app right there. The agents can control the phone, too, so they can test what they just built. This is super cool, as well. Now, let’s talk about attachments. Click the plus button in the prompt box. You can attach images, PDFs, documents, whole folders, or project files. There’s also a sketch tool. Draw a rough layout of what you want and send the drawing to the agent. You can make mood boards, too. A mood board is a set of reference

07:28 images for one project. Add some images and captions, then attach the board to a message. The agents can also open the boards on their own. Now, let’s talk about work trees. This matters when you run many agents on the same project. A work tree is a second copy of your project on its own branch. Above the prompt box, there’s a bar. It shows the project, the folder, and the branch. Click the branch name, then you can move the task into a new work tree. Now the agent works in its own copy on its own branch. So, it doesn’t break what the

08:01 other agents do. When the agent finishes, click the branch name again, choose merge into main. Panda Code merges the saved commits into your main branch on your computer. If you have changes that you didn’t commit, it stops and changes nothing. It does the same if there’s a conflict. After the merge, it offers to remove the work tree. You can also switch branches from the same menu. For example, go straight back to main. So, the whole work tree flow takes about two clicks, which is great for sure. You can also turn on GitHub. This uses

08:32 the GH tool from GitHub. You get a GitHub view with your repos, their threads, pull requests, and issues. From there, you can review a pull request, fix failing checks, merge, or solve an issue. Each action is one click. Each one opens as a thread on the side, so you can keep chatting with the agent. When an agent opens a pull request in a task, Panda Code links it to that task. You can see it and merge it from the task’s PR tab. Now, let’s talk about many tasks at once. Click board at the top. You get a task board with four columns: pending, in progress, review, and completed. Add a

09:05 task with a title and a description. It waits in pending. Click start and the agent you picked starts working. When the agent finishes, the task moves to review. When you’re happy with it, mark it completed. Panda Code archives completed tasks after 7 days. The board shows tasks from all your projects. You can open any task in a side panel and chat with it there. The sidebar helps too. In all projects, each thread is one line under its project icon. You can see which tasks are running, which ones need you, and

09:36 which ones finished. You can also pin threads and drag them into the order you want. If a thread is long, press command F. You can search your messages and the agent’s messages. Panda Code highlights every match. Now, let me show you agent teams. This is one of my favorite features. Click the team button in the prompt box. You get an architect and up to three workers. The architect makes a plan. It splits the work into parts. Then it gives each part to a worker. The workers run at the same time. And you can mix agents here, for example. The architect can be Claude Opus. One worker can be Codex. Another

10:08 worker can be Kimmy or GLM. Each worker appears as one line in the chat. Click the line to see that worker’s details at the top right. When the workers finish, the architect checks their work. Then it gives you a summary and a list of all the changed files. You can also save named teams in settings. For example, a cheap team with GLM workers or a strong team with Opus and Codex. Then you pick the team from the model picker. So, this is kind of great. Codex also supports goals. Type {slash} goal. Write what you want and if you want a token budget. Codex keeps working on the goal turn after turn. A banner

10:41 shows its progress. Now, let’s talk about bots. This is something that I really like. You can connect an agent to Telegram, Slack, Discord, WhatsApp, or a local web chat. So, you can message the bot from your phone. Then it works on your project on your computer. Each bot has its own profile, memory, schedules, channels, and rules. For example, you can make a bot that sends you a summary of your repo every morning. Every bot chat is also a normal task in Panda Code. So, you can open it on your

11:12 computer, see what the bot did, and keep talking to it. The bots list in the sidebar shows every bot with its latest message. Panda Code can also run agents in virtual machines. A virtual machine is a separate computer that runs inside your computer. On Macs with Apple chips, Panda Code can make macOS and Linux virtual machines. It uses is free tool called Tart for this. The agents can see the screen, click, type, and run commands inside them. So, an agent can test your app on a clean machine without touching your real one. If you want to use the cloud instead, Pandacode supports Daytona,

11:45 Cloudflare, and AWS sandboxes, too. There’s also a deploy menu at the top. You can start a deploy task for Vercel, Cloudflare, or AWS. The agent uses their command-line tools with your login. You approve every command first. Pandacode also hides your keys in the output. Now, let’s talk about the parts that save you money and time. First, agent usage. Click the activity icon at the top. You see the limits for every agent you installed. For example, your Claude 5-hour and weekly limits, and your Codex limits. You also see when each limit reset. It also estimates what your usage would cost if you paid the API price.

12:19 You can see this for a week, a month, or a year with a chart and a split by model. Take it with a grain of salt because it’s only an estimate, but it’s fun to see how much value you get from your plan. If you hit a limit in the middle of a task, choose resume at reset. Pandacode continues the task by itself when your limit resets, so you don’t have to sit and wait. Second, the token saver. You can turn on short answer styles for each task. There’s caveman, lean code, and quiet tools. They make the agent write shorter answers, so you use fewer tokens, and Pandacode sends this instruction only

12:50 when you change it, not with every message. So, the token saver doesn’t waste tokens on itself. Pandacode saves tokens in other ways, too. Claude code gets Pandacode’s own instructions once and then reuses them from its cache, so they cost almost nothing per message. It also makes follow-ups faster. When a Claude code turn ends, Pandacode starts the next process in the background. In my tests, the first word of a follow-up came in about 4 seconds instead of about 8, which is pretty good, for sure. Third, agent updates.

13:22 Pandacode checks for new versions of your coding agents. You can update each one with one click. Then there’s the quick overlay. Press command shift space anywhere on your computer. A small launcher opens. You can start a new task or check your running tasks without opening the app. So if you’re in your editor and you get an idea, you just send it. There are more shortcuts, too. Command K opens the command palette. Command B shows or hides the sidebar. Command comma open settings. Command plus and command minus make the whole app bigger or smaller, not just the text. There’s also an optional add-on called Jev.

13:56 It can review the agent’s work, check your project rules, and test your app in a separate Chrome window. It’s off by default. Pandacode works fully without it, so you can ignore it if you want. Now, let’s talk about privacy, because I think this matters a lot. Pandacode stores everything on your computer in a local database. Your prompts go only to the agent or service that you pick. There are no accounts and no tracking. Pandacode makes only three automatic requests. It checks for new versions of your coding agents. It downloads a

14:27 public list of model prices for the cost estimates. And it checks for its own updates, which you can turn off. Your keys stay in your computer’s keychain. Pandacode is open source under the MIT license. There’s one small extra rule. If you share a copy of it, you must say that it’s a copy of Pandacode. So, you can read the code, build it yourself, and help make it better. To build it yourself, clone the repo, run npm ci, then run npm run desktop, and that’s it. Now, I want to be honest about a few things. First, the apps are not signed

14:57 yet, so you see that warning the first time. Second, the simulator and virtual machine features are mostly for Macs. Third, Pandacode depends on the coding agents themselves. If an agent changes something, Pandacode may need an update to catch up, but Pandacode updates itself, so you get those fixes automatically. Overall, it’s pretty cool. Anyway, let me know your thoughts in the comments. If you like this video, consider donating through [music] the Super Thanks option or becoming a member

15:27 by clicking the join button. Also, give this [music] video a thumbs up and subscribe to my channel. I’ll see you in the next one. Until then, bye.