# Show HN: Let AI order DoorDash for me – under my watch – no playwright > Source: > Published: 2026-10-17 02:18:56+00:00 No code. No glue scripts. # Connect your AI agent to everything. Your browser. Your containers. Your databases. Configure the tools, teach it a skill, then watch it work — every step on screen, nothing running where you can't see it. macOS 12 or later · Apple Silicon or Intel · 251 MB · Free, no account · Runs entirely on your machine Also available for Windows — [download Wysebee-0.3.0-setup.exe](https://github.com/wysebee/desktop/releases/download/v0.3.0/Wysebee-0.3.0-setup.exe) (187 MB) What it connects to ## An agent that can only talk isn't much use. Wysebee hands yours the three things it actually needs to finish a job — and nothing beyond what you hand over. ### Browser Let it use the web like you do. Point the agent at a page and it clicks, reads, scrolls and fills in forms — the same actions you would take, driven by the skill you gave it. The sites that have no API, the internal dashboard nobody will ever expose: still reachable. - Pull a report nobody exposes an API for - Fill the same form a hundred times - Check a page and tell you what changed ### Containers Give it a machine, not your machine. Commands run inside the node's own Docker workspace — never on your host. The agent gets a real shell, a real filesystem and whatever runtime you pick, and it is sealed off from everything else you own. - Run a script against real data - Build, test, read the output - Keep the work in one folder you can delete ### Databases Answer questions in the data itself. Connect the database you already run and the agent queries it directly — no export, no copy, no pipeline to maintain. What it can reach is exactly what you configured, and nothing else. - Ask in plain English, get real rows - Reconcile two systems that disagree - Draft the migration, then show you it A real run ## “Find me something sweet nearby.” One sentence, no code, and no API key for a food-delivery site that never offered one. Watch the left half of the window: that is your own browser being driven, live, while the agent narrates every call on the right. What it came back with Turtle Brownie Ice Cream Sundae — Bocados Bake Shop - **CA$11.50** before taxes and fees - **~27 minutes** to the selected address - **100% liked** over 12 ratings · shop 4.6★ Plus a fallback it offered unprompted: the Holy Rolly Classic Cinnamon Roll at CA$7.16, in case you'd rather skip ice cream. ### Your browser, not a datacenter's The page on the left is the real tab, on screen, on your machine. No headless browser running somewhere you can't see, no scraping farm — the agent works the window you already have open. ### No API needed DoorDash publishes nothing for this. The agent did not need it to: it read the page, clicked through and filtered the same way you would. ### Eight calls, all on the transcript browser_open_tab, snapshot, click, snapshot, click, snapshot, tabs, snapshot — each one listed as it happened, each one openable to see what went in and what came back. How it works ## Three steps. None of them are writing code. You decide what it can reach and how the job is done. It does the typing. 1. 01### Configure the toolsPick what the agent is allowed to touch — a browser, a container, a database, an API you already run — and fill in the fields. It is a form, not a file. No SDK, no server to write, no deploy step. A form, not a file 2. 02### Set up the skillWrite down how the job gets done, once, in plain language. The agent loads that skill when it needs it, so the thing you explained on Monday is still being done your way in March. Plain language, not prompts you re-paste 3. 03### Watch it workThen let it go. Every step is on screen as it happens — what it opened, what it ran, what it changed. You are not handing over the keys and hoping; you are watching the work and stopping it if it goes wrong. On screen, step by step Under your eye ## An agent with real hands needs real limits. Letting something act on your behalf is the part that should make you nervous. Two things in Wysebee are built for exactly that worry — and both are on before the first step runs. ### You watch the work happen. The agent does not disappear and come back with a result. Every action it takes is on screen as it takes it — the page it opened, the command it ran, the row it changed — so a wrong turn is something you catch in the moment, not something you discover afterwards. - Each step surfaced as it happens, not summarised at the end - Stop it mid-run; nothing is a black box you cannot interrupt - The skill you wrote is the instruction set — no hidden prompt ### It gets a container, not your filesystem. Shell commands never run on your host — they run inside the node's own Docker workspace, and the file tools can only reach that one folder. A poisoned web page can still try to steer the agent. It just lands somewhere that holds no secrets and can see one directory. - cap_drop=ALL, no-new-privileges, an unprivileged uid, and memory/CPU/process ceilings - An allow-listed environment built from scratch — never the one your shell has - No credential key, no database route, and none of your source inside the container - Paths resolve through symlinks before the containment check, so a link pointing out is caught like a ../ escape **One deliberate exception:** the admin-only Local Command tool runs a fixed program on your host, because sometimes that is the point. Even then the arguments never reach a shell, and the environment is allow-listed. [Read the security posture](https://www.wysebee.com/desktop#security) Any provider ## Bring your own model. Or don't bring one at all. Paste a key from whichever provider you already pay for — or point Wysebee at a model running on your own machine and skip the key entirely. Same agent, same tools, same screen; only the brain changes. ### Hosted, with a key - OpenAI - Anthropic - Google Gemini - Alibaba Cloud (Qwen) - DeepSeek - Mistral AI - Sakana AI Keys are encrypted on this machine. They can be replaced or removed, but are never shown again once saved — not in the UI, and not to the agent. ### Local, with no key Ollama · LM Studio Or any OpenAI-compatible server on your machine or your network. Point Wysebee at the URL, leave the key blank, and **nothing leaves the machine** — not the prompt, not the page the agent read, not the rows it queried. Pick a model that supports tool calling — in Ollama, qwen3 or llama3.1. ## It runs on your machine. Wysebee Desktop turns your computer into a set of **nodes** — named bundles of tools you build by filling in a form — and gives your agent exactly those, and nothing else. The work happens where you can see it, on hardware you own. macOS 12 or later · Apple Silicon or Intel · 251 MB · Free forever · No telemetry Also available for Windows — [download Wysebee-0.3.0-setup.exe](https://github.com/wysebee/desktop/releases/download/v0.3.0/Wysebee-0.3.0-setup.exe) (187 MB) ### No sign-in, no account Open the app and start building. Nothing dials out, and there is nothing to create an account for. ### No code, anywhere Every tool is a template you fill in. No SDK to learn, no server to write, no deploy step between you and a working agent. ### Skills it keeps Write down how a job is done once. The agent loads that skill when it needs it, instead of you re-explaining every session. ### Everything stays local A SQLite file in your user data directory and a folder per workspace. Delete those and nothing is left. FAQ ## Questions we get asked ## What does Wysebee actually do? It connects an AI agent to the things you already run — your browser, a Docker container, your databases, an API you own — and then runs that agent against them while you watch. You configure which tools it may touch and write down the skill that describes how a job is done. The agent does the work; you keep your eyes on it. ## Do I need to know how to code? No. Every tool is one of a handful of templates you fill in — a request to an API you already have, a command inside a container, a query against a database you connect. There is no SDK to learn, no server to write and no deploy step. If you can fill in a form and paste an API key, you can give an agent something real to do. ## Can I see what the agent is doing? That is the point. Every action is surfaced as it happens — the page it opened, the command it ran, the change it made — rather than summarised after the fact. You can stop a run in the middle. The skill you wrote is the instruction set, so there is no hidden prompt deciding things on your behalf. ## Can the agent read or write files on my computer? Only inside a workspace you gave it. Shell commands never run on your host — they run in that node's Docker container, and the read_file and write_file tools are confined to that one workspace folder, with symlinks resolved before the containment check so a link pointing out is caught like a ../ escape. The container drops all capabilities, sets no-new-privileges, runs as an unprivileged user under memory, CPU and process ceilings, and gets an environment built from an allow-list rather than inherited from yours. The one deliberate exception is the admin-only Local Command tool, which runs a fixed program on your host on purpose. ## What is a skill? A written description of how a job gets done, in plain language, saved once. The agent loads it when the work calls for it, which is what stops you re-explaining the same process every session — and what makes the way you do things on Monday still the way it is done in March. ## Which AI model does it use? Whichever you point it at. Paste a key for OpenAI, Anthropic, Google Gemini, Alibaba Cloud (Qwen), DeepSeek, Mistral AI or Sakana AI — keys are encrypted on your machine and never shown again once saved. Or run the model yourself through Ollama, LM Studio or any OpenAI-compatible server, leave the key blank, and nothing leaves the machine at all. Pick a model that supports tool calling; in Ollama, qwen3 and llama3.1 both do. ## Is it really free? Yes. Wysebee Desktop is free to download and use, with no account, no sign-in, and no telemetry. Everything runs on your own machine. ## Which platforms does it support? macOS 12 or later on Apple Silicon or Intel, and Windows 10 or later on 64-bit. The Mac build is Developer ID signed and notarized by Apple, so it opens with no Gatekeeper warning. The Windows installer, app and uninstaller are Authenticode-signed with an EV certificate and timestamped, so Windows names Wysebee Inc. as the publisher instead of warning about an unknown one — and it installs for your user only, with no administrator prompt. A Linux build is not available yet. ## Do I need Docker? Not to run the app. Wysebee starts and works without it. Docker Desktop becomes necessary only the moment the agent needs to run a shell command — that is, the moment you create a workspace and connect it to a node. ## Where does my data go? Nowhere. Everything durable is a SQLite file in your user data directory and a folder per workspace, both on your machine. There is no sign-in and nothing dials out. ## How do I get help? Join the Discord. That is where questions get answered, rough edges get reported, and what ships next gets decided. Bugs and feature requests for the desktop app can also go to the GitHub issue tracker. ## Come build with us. No forms, no pitch deck, no sales call. The Discord is where the questions get answered, the rough edges get reported, and what ships next gets decided. Free, no account required. Download and run it in minutes.