STINGER FORGE
An AI coding workspace with five roles at one desk.
Forge seats five named roles — Lead, Prompt, Forge, Critic, Studio — at one desk, each bound to a model from a different company. Lead can hand part of a job to a specialist mid-turn, and the running cost sits on the same screen as the code.
How it actually works
Not a race between models. A lead that can call for help.
Forge has five named roles — Lead, Prompt, Forge, Critic, Studio — and each one is bound to a provider and a model you choose. They are jobs at one desk, not licences you buy by the head. You talk to Lead. When the job needs a second head, Lead dispatches another role with a brief, waits for the answer, and carries on with it. That is one conversation that reached across vendors, not five answers to compare.
Only Lead can delegate
The dispatch tool is restricted to the Lead role, so a sub-agent can't spawn its own crew. Nesting stops at four deep and the tool loop stops after eight rounds — then it says it stopped and asks you to continue, rather than spinning.
Nine tools, and the risky ones are narrowed
List files, read, write, generate an image, read a skill pack, dispatch another role, deploy. The two deploy tools are available to Lead and Forge only, and both need a token you paste yourself — there is no ambient credential to borrow.
A missing key falls back instead of failing
If a role's provider key isn't there and an xAI key is, the run switches to Grok and keeps going. That's a key you haven't added — not a claim that Forge routes around a vendor outage.
Twelve skill packs, not five hundred
Roles can list and read a small library of skill packs. It's twelve, deliberately — the app says so itself rather than advertising a number nobody reads.
The roster
Five chat models across four companies. Named, because a vague roster is a roster you can't check.
xAI · Grok 4.5
The default for the Lead, Forge and Studio roles. Metered rather than flat-rate, and the fallback everything else lands on.
Anthropic · Claude Opus 5 and Claude Sonnet
Two of the roster's five slots are Anthropic, called with prompt caching on so a long file you keep re-sending isn't billed as new every turn.
OpenAI · GPT-4.1, Google · Gemini Flash
Gemini Flash is the default for the Prompt and Critic roles. Image generation reaches three more models — Grok Imagine, DALL·E 3, Imagen 3 — trying them in order until one answers.
Voice on Groq
The mic records in the browser, auto-stops on real silence detection rather than a timer, and transcribes on Whisper Large v3 Turbo. The text lands in your draft box — it does not send itself.
Your work, your keys
The transcript never reaches our database.
Rooms live in your browser
Files, messages, open path and project label are stored locally per room. The server writes two things per run: an encrypted usage ledger row, and nothing else. Your conversation is not in a table we can read.
Keys are encrypted, and never in local storage
API keys are held encrypted at rest with AES-256 and are deliberately excluded from the browser store. An earlier version kept them in the clear; that store is actively deleted on upgrade rather than left to rot.
A budget with a stop
Every room carries a dollar budget — twenty by default — and a pause switch. Per-model input and output rates are in the catalog, so the meter is arithmetic you can check, not an opaque credit balance.
Export is a plain JSON file
One button writes the room's messages to a JSON file, another writes the usage ledger to one. There's a published ingest endpoint for that ledger, so the cost record can leave for your own system.
What's real vs. what's roadmap
The desktop app is a real window. It is not yet a real download.
Desktop: a native window, and honestly a wrapper
The desktop app is a native Windows window around the hosted app, and it earns its keep with something a browser can't do: pick a folder and read it into the room, then write changes back with path-traversal guards. It reads up to forty files, 80 KB each, 2 MB total — bounded on purpose.
The installer is built, the binary isn't shipped
The one-file executable spec and the Windows installer script are both real and in the repo. What downloads today is the source, and it needs Python on the machine to start. It isn't code-signed yet either, so Windows will warn. We'd rather say that than let you find out at the download.
Phone: a responsive layout, not an app
The same page collapses to a four-tab bottom bar — Files, Code, Design, Crew — and works on a phone. The iOS and Android shells are configured but not assembled, and there is no store build. Checkout is web-only by design.
No shared room
Forge is one person at the desk. There's no live collaboration, no second cursor, no shared session — a room lives in one browser, and a room you export is a file you hand over, not a link you invite someone to.
A plain editor, on purpose for now
The code pane is a text area that recognises six file types. No syntax highlighting, no language server, no inline diff. It's where you read and nudge what a seat wrote, not a replacement for your IDE.
Bounded context, stated
A prompt is capped at four thousand characters, a request carries at most twenty-four files, and each one is truncated before it goes on the wire. A deploy pushes at most twenty files. These are real ceilings, and you should know them before you plan around them.
Get started
$157 a month, metered like Teams, and a straight answer about what's finished.
Forge is sold on its own rather than as a line on a tier sheet. Ask Vincent what's switched on today, what a room costs to run at your usage, and which of the limits above would actually get in your way.
Related
Part of the same AI-tooling lineup.
Forge sits alongside Stinger's other multi-model AI tools.