About Flashbag

AI employees, hired by name and role

You add an employee by giving it a name and a role (a copywriter, a bookkeeper, a customer-service agent) and talk to it in your own words, as you would to a new hire. Employees are grouped in departments, and a whole team can be installed at once as a Profile. Outcome: work that used to sit on the owner's plate, such as customer emails, quotes, posts and reports, comes back as drafts ready to approve.

Memory that never resets, and stays on your computer

Every conversation with an employee is saved word for word, indexed and searchable. What an employee learns is distilled into facts it carries into every next conversation, and each employee keeps its own memory. Nothing of this leaves your machine. Outcome: you explain your business once, not every morning, and the team gets more valuable every week you use it.

Scheduled tasks

Any job can be scheduled, daily, weekly or monthly, at 3 a.m. if you like, and runs without you at the keyboard. A morning briefing recaps what got done overnight. Outcome: reporting, research, email drafts and briefings are finished before the working day starts.

Choose your model and provider

Flashbag runs on the AI subscriptions you already have, Claude (through Claude Code) and OpenAI (through Codex), and you can switch model mid-conversation; memory, skills and tasks stay the same. OpenRouter gives you access to 300+ models through experimental Pi support (off by default), using your own OpenRouter API key. Local AI models are coming soon. Outcome: a usage limit, a price change or a rule change at one provider does not stop your business.

Skills and Profiles

A Skill is a way of working you explain once, such as how you want a quote written or how a weekly report is built, installed in one click and reusable by any employee. A Profile packages a whole team with its skills so it can be installed by someone else in one click. Outcome: your best procedures stop living in your head.

Gordon, the building assistant

Gordon is the employee who builds the rest: ask him to create an agent, install a skill, schedule a task or fix what broke, and he does it. Outcome: no settings screens to learn; you describe what you want.

A cost dashboard for every token

Flashbag tracks tokens and estimates API-equivalent costs per day, employee, model and Profile. Claude and OpenAI subscription usage comes from your plan; OpenRouter usage is billed separately by OpenRouter. Outcome: you can compare the value of your AI work, with no per-token charge from Flashbag. The dashboard is an estimate, not your provider’s bill.

Runs on your computer, reachable from anywhere

Flashbag Desktop runs directly on a dedicated Mac (Apple silicon) or Linux computer. Flashbag Launcher gives the team a contained workspace on a computer you also use for other things, and is the way to run Flashbag on Windows; you use it from your browser. Built-in Tailscale support lets you open your Flashbag from another computer or your phone without exposing anything to the public internet.

Flashbag for Business (coming)

One workspace for a whole company: every human employee gets their own sandbox, skills and brand voice are shared, and an admin manages access. $197 per company per month, not per seat. Currently a waitlist.

The app sends nothing to us: no telemetry, no prompts, no files

Flashbag runs entirely on your computer. Orange Boys receives no usage data, no analytics, no crash reports and none of your content (Privacy Policy v1.1, 21 Aug 2026). Meta's Muse, Grok Bot (via Cursor), Squad and Lindy run the agent on their own infrastructure, so what the agent learns about your business lives on their servers under their terms. One honest note: the thinking still happens at the model provider you choose, and your prompts go there directly under that provider's terms.

Memory is verbatim, per employee, and yours to keep

Flashbag stores every conversation in full and lets you read and edit what an employee has learned. When you stop paying, the app goes view-only and every agent, memory and file stays in plain folders on your computer (Terms, art. 8). Claude Cowork organises work around projects and threads; Flashbag puts the employees front and centre, each with its own memory.

Bring your own model, switch in one click

Flashbag is not tied to one AI company: Claude and OpenAI models run through your own subscriptions, and experimental OpenRouter support adds 300+ models through your own API key. You can switch models while keeping your agents and memory. Cowork runs Claude models only; Grok Bot lets Cursor manage model selection; Muse runs on Meta's models. When one provider changes a rule, as Anthropic did on 4 April 2026, the work moves and the memory comes with it.

One flat price, no credits, no per-seat fees

Flashbag is $97 per month after a 14-day free trial with no credit card, every feature and every agent included, cancel anytime. No credits, no per-seat fees, no metered actions. Hosted alternatives such as Lindy charge per user and include a credit allowance.

Built for people who don't code, by a developer who has for twenty years

Install, name an employee, talk. No terminal, no prompt course, no gateway to operate: open-source frameworks such as OpenClaw and Hermes Agent are powerful but assume a technical operator. Flashbag itself is not vibe-coded. It is built by Atlas, a senior developer with more than twenty years of experience, who uses AI as a tool the way a senior developer does: the architecture, the plans the AI works from, the implementation choices and the test-driven development are his. The founder uses AI to run the company, not to write the product. That company runs on Flashbag, in public: the AI team that markets it works on flashbag.io/live, with tokens, cost and output published every week, including the bad weeks.

Atlas, CTO and developer

Atlas has built the Flashbag software from the first build in May 2026; every release goes through him. He is a senior developer with more than twenty years of experience, from Croatia, and he and Mitchel have worked together for the past six years. He uses AI in his work, as any good developer does in 2026; the difference is what twenty years bring to it. He designs the architecture, writes the plans the AI executes, decides how each part is implemented, and holds every change to test-driven development. The founder sets the direction and the AI team helps run the company; the code is his responsibility.

The AI team

Flashbag is also run with Flashbag: fourteen AI employees (23 Sep 2026) handle marketing, research, the changelog and support, and you can watch what each one is doing, and what it costs, on flashbag.io/live. They do not write the software; Atlas does. The agents are named after Flash Gordon.

How Flashbag started

Until April 2026 Mitchel was building AI agents for his own companies and teaching Dutch entrepreneurs to do the same in bootcamps. On 4 April 2026 Anthropic cut off the use of Claude subscriptions inside third-party agent tools. TNW reported that more than 135,000 OpenClaw installations were running at the time. The change shut down every agent Mitchel was running and every agent he had sold to his 70 bootcamp clients (TechCrunch, VentureBeat, TNW; access was later restored on new terms). Everything he had built sat on a switch that somebody else's hand was on. In the four months after that Atlas built Flashbag, with Mitchel as its first user and the bootcamp customers as its test group: about a hundred entrepreneurs ran it in stealth before the public launch on Friday 28 August 2026. The rule it was built on is in the manifesto: Keep it home.

Where to follow along

• YouTube: @flashbag-ai (product) · @buildinginpublicdotcom (the journey to 10,000 customers, in seasons counted in customers: Season 1 runs from 0 to 500)
• X: @Flashbagio · @bipdotcom
• Facebook: facebook.com/buildinginpublic
• Website: buildinginpublic.com