STINGER TEAMS
Multi-model AI collaboration, run as one room.
Ask once, and every AI model you picked answers the same brief at the same time — then a validation pass reads the answers against each other before you do. It installs as a Windows app and the server runs on your own machine.
How it actually works
A real fan-out, not a queue.
You write one mission and pick who's on it. Every selected model is dispatched at the same time, answers stream back as they land, and when the last one is in, the room runs a cross-check over all of them together. That is the difference between asking five models and reading five tabs.
Models from companies that compete with each other
Anthropic, xAI, Google, OpenAI, DeepSeek, Groq, Perplexity and OpenRouter each have a live call path and their own key slot. The point of a room is that no single vendor's judgment is the only one you get.
The cross-check runs by default
After the fan-out, every response is concatenated and handed to a validator with one job: find factual errors, contradictions and claims with nothing behind them. It returns a verdict — validated, concerns, or contradictions — with a confidence figure and a list of findings tagged by which model said it.
Disagreement is reported, not averaged
The coordinator pass is told in as many words not to average conflicting answers together. Two models disagreeing is a finding about the question, and blending them away destroys the only evidence you had.
Roles are fixed to the model, not guessed
Strategist, researcher, coder, validator, creative — each model carries its role in the registry. The role sets that model's output ceiling, so a validator answers tighter than a strategist by design rather than by prompt-begging.
The installer
A .exe, not a promise.
One file, built by CI, smoke-tested before it ships
StingerTeams.exe is a single-file Windows build produced on every push to main. The pipeline then launches the executable it just built and polls its health endpoint for two minutes — if the app doesn't answer, the build fails. A binary nobody started is not a binary anyone should download.
The server is on your machine
The app starts its own local server and opens a native window pointed at it. Not a browser tab, and not a page on our host that stops working when we do.
Windows 8 or later, half a gigabyte free
The installer checks both before it starts. It's a big download — the runtime and its dependencies come with it, so the machine doesn't need Python installed.
Your download link is good for fourteen days
Access is a tokenised link tied to your purchase, and the expiry is a real number in the code rather than a policy line on a page.
Your room, your record
Sessions are files. Exports are files. Nothing is trapped.
One JSON file per account, on disk
Missions and messages are written to a plain JSON file scoped to your account, rewritten on every change. No shared database table, no other tenant's rows next to yours — and if authentication can't resolve who you are, the request is refused rather than falling back to a shared namespace.
Export takes everything, including what the UI hides
The export bundle carries your sessions, model memory and verified facts with a schema version stamped on it — and it deliberately skips the retention window the interface applies, because an export is data portability, not a screen you're reading.
A transcript you can paste
The console writes a plain-text transcript too, and copies it to the clipboard in the same action — because a download inside a desktop webview can be silently blocked, and an export that quietly did nothing is worse than none.
Spend is reserved before the call, not counted after
Usage is claimed ahead of dispatch and settled against the real token counts afterwards, with a cross-process lock so two workers can't both think they had room. A client that hangs up mid-run doesn't leak paid usage past the cap.
What's real vs. what's roadmap
Most of the room is text-only, and the settings screen doesn't save your keys yet.
Keys come from configuration, not the Settings box
Bring-your-own-key is read on the server, but nothing writes it — the Settings modal in the console saves nothing to the server today. Working keys are supplied through environment configuration next to the app. We'd rather tell you that than let you paste a key into a box that drops it.
Tools and browsing are not the whole room
Only some of the providers can be handed tools at all, and the full hands-and-eyes path — a model driving a browser and seeing the screenshot — exists on the Anthropic path alone, for the owner only. The rest of the room reads the mission and answers in text.
Code execution is not wired
The sandbox endpoint exists and refuses. There is no execution vendor behind it, and asking for it returns an error saying exactly that rather than pretending to have run your code.
The reviewer is the cheapest one available
The cross-check deliberately picks the least expensive configured model so it can run every time instead of being a luxury you skip. It reads answers, it does not re-do the work — and if it returns something that isn't parseable, the verdict degrades to "concerns" rather than to "validated."
Cost figures are estimates, and labelled as such
The internal price table is marked unverified in the code until a human reconciles it against real provider billing, and it doesn't yet cover every model in the roster. Treat the meter as a good guide, not an invoice.
Windows only, and unsigned
There is no macOS or Linux build. The executable isn't code-signed yet, so Windows will warn on first run, and there's no silent auto-updater — the app tells you a newer version exists and you fetch it.
Get started
Sold on its own, and included with the platform.
Teams is a standalone purchase and it's included with the Stinger ecosystem tiers. Either way it's the same desktop app and the same room. Talk to Vincent about current packaging, which models are switched on for your account, and how much room a month of real use actually takes.
Related
Part of the same AI-tooling lineup.
Teams shares its multi-model roster with the rest of Stinger's AI-tooling cluster.