BigAgentCreator
Design an AI agent as a workflow on a canvas, save it as a folder, and run that folder in Docker on your machine.
You stay in VS Code or Cursor. The drawing is not a picture of the agent. Each save writes the Agent SDK files. A trigger, a tool, a guide, and a Telegram chat are files you can open. Secret values stay in .env. The image never contains one.
What you can build
- A Telegram bot that answers from rules you wrote, refuses what you marked as out of bounds, and replies in the same chat.
- Several ways in. Telegram, a webhook, a schedule. Each arrow is a path. Two triggers can reach the agent. One can skip it.
- A path with no model. If a trigger never reaches the Agent card, that path answers in code. A message can come in, a function can multiply it by two, and nothing calls an LLM.
- One agent definition per folder. The Agent card is
bot/agent.ts and instructions.md: name, model, and instructions. Remove it when the flow does not need it. Add a subagent when the main agent should delegate.
- Tools and guides. An action is something the run can do. A guide loads only when it is needed. A service is a system the model may call. A host is a system only your code calls.
- A local run you can try. Ship builds the image and starts it. The playground is http://127.0.0.1:3000/playground. Rebuild applies the current
.env without copying secret values into the image.
How a flow behaves
Dropping a card does not connect it. You draw the arrow. The number on a card is the step that arrow enters. The same number can run together. A higher number waits.
The model is the Cursor API (CURSOR_API_KEY). This Docker image cannot run the SDK sandbox, so the generator leaves that switch off.
This release, 0.2.1
|
|
| Design |
Canvas, Agent, limits, triggers, actions, guides, services, hosts, subagents, hooks. |
| Prove |
Fixed cases you can add and open as TypeScript. |
| Secrets |
Names in the project. Values in .env. Logs hide them. |
| Connect |
Open a Cloudflare tunnel and register the Telegram webhook. The bot token is not shown. |
| Ship |
Start, stop, rebuild, logs, and the playground link. |
Kubernetes, AWS, GCP, and Azure are the same image later: each cloud only starts it and injects the same secret names. 0.2.1 does not deploy there yet. The next step after that is a versioned image: build once, keep the digest, deploy that digest instead of an unnamed rebuild. The line is ROADMAP.md in the repository root. There is no second model runner.
Start
- Install the extension and open a folder. Use Examples and load Telegram desk, or open
examples/telegram-desk yourself. The first load keeps your current files in .keep/original. Restore my folder puts them back.
- Open BigAgentCreator on the activity bar.
- Initialize the workspace if the folder has no agent files yet. Agent and Limits appear. Either can be removed.
- In Secrets, set
CURSOR_API_KEY. For Telegram also set TELEGRAM_BOT_TOKEN (the BotFather value, with the colon) and TELEGRAM_WEBHOOK_SECRET (letters, numbers, _, and - only).
- Ship → Start. When the status says the playground is ready, open it and ask something the instructions allow.
- For the real Telegram chat, point an HTTPS tunnel at port 3000 and use Connect → Register webhook. Then message the bot. Leave the tunnel open while you test.
Develop this extension
From apps/extension:
npm install
npm test
npm run compile
npx @vscode/vsce@2.32.0 package --allow-missing-repository --skip-license
The Marketplace icon is media/icon.png. The activity bar uses media/icon.svg.
| |
