Scrum4Us builds and maintains your software with AI agents — on a scrum board you control. Every change arrives as a verified pull request, every agent turn is metered in the ledger, and your agents can run on your organization's own AI account.
Now in private pilot — room for five organizations to build along for free.
How it works
No prompt roulette: Scrum4Us works the way a good team works. Work lives on a board, nothing ships without your approval, and everything is traceable afterwards.
Write your idea as a story — plain language is enough. Pick it for the sprint and dispatch an agent; it works the story task by task, updating the board as it goes.
Every story becomes code in an isolated, throwaway sandbox that can reach your repository, the model gateway and the platform's own job API — nothing else. The result arrives as a pull request with an explanation, never straight into your product.
You review the pull request and merge it — nothing merges without you. The board rolls along automatically, and the ledger shows what the work cost — per turn, per user, per model.
Built-in skepticism
Three mechanisms keep you from having to trust an agent's output.
Every build job runs in its own disposable container with deny-by-default network access: it reaches your repository, the model gateway and the platform's own job API — nothing else. Discarded after the run.
The platform checks every delivered pull request server-side — right repository, right branch, a real diff at the reported commit. An agent cannot claim work it didn't do.
Plans and documents that agents write can be sent through adversarial review rounds — independent AI reviewers, deliberately drawn from two different model families, leaving a durable review record.
Beyond building
Next to the build copilot, you define agents for your daily work: a stated purpose, instructions, a model policy and the knowledge they may use. Every employee can make them personal — within the boundaries you set.
Purpose: answer questions with citations from your protocols.
Answers always come with the source attached, from documents you manage yourselves. If the protocols don't cover it, it says so — no improvised policy.
Purpose: get new colleagues productive from your own handbook.
Grounded in the documents you publish — house rules, how-tos, architecture notes. New hires ask; the agent answers with the source attached, in your organization's own vocabulary.
Knowledge base
Protocols, manuals, decisions — uploaded once, shared with the whole organization, and quotable by every agent you allow.
Add text or markdown documents to a collection. Indexing happens inside the upload itself: the moment it completes, the document is searchable. No queue, no processing spinner.
Questions search your documents by exact wording and by meaning, and both rankings are fused — so the right paragraph surfaces whether you use the document's words or your own.
Agents answer with the documents they used attached as citations. When nothing relevant exists, they say exactly that instead of improvising.
You set the dials
What an agent may do, what it may know, what it may cost and whose account it runs on are settings your admins manage — not a consultancy project.
Register your organization's own Anthropic or OpenAI key. Turns served on it are billed by your provider — the ledger records them with real token counts and zero platform cost. Keys are encrypted at rest; a test button proves the connection.
Per agent, you decide whether the platform may step in when your provider fails. Off means off: a clean error in the chat, never a silent switch to our models — verified in production.
Dispatch any job with an explicit model or capability tier, plus an effort level from LOW to MAX. The choice is recorded in the ledger next to what it cost.
A monthly cap for the organization, optional caps per user. A turn that would breach the cap is refused with a clear error before the money is spent.
Name, purpose, instructions, model policy, knowledge bindings, on or off — creating an agent is form-filling, not a development project.
Every employee can add personal instructions to any agent. Platform guardrails and your organization's rules always win — the merge order is enforced by the platform, not by convention.
Privacy without asterisks
Built in the EU for organizations that must account for their data. Below is what is engineered in today — the roadmap says what comes next, and we don't blur the line.
Honest footnote: chat content is processed by the AI provider an agent is configured for — the platform's models or your organization's own Anthropic or OpenAI account. Admin screens for the processing register and self-serve export are on the roadmap; the APIs beneath them are already live.
Where we're going
Built in this order, each step behind a live verification gate before we call it done. No dates — we'd rather be honest than optimistic.
Point Scrum4Us at your own model endpoint — vLLM, Ollama, TGI — behind a fail-closed egress contract. The design has been adversarially reviewed; the build is not yet scheduled.
Mail a document to an agent, or photograph a receipt and have it staged as an expense entry — input beyond the web interface.
Your agents inside your own application, with delegated identity for your users, plus a system-to-system API.
The full platform as an installable bundle on your premises or in your cloud — web layer included, with a proper update story.
A processing-register view for your admins and self-serve export & erasure screens — on top of the APIs that exist today.
Pricing
Introductory pricing during the pilot phase. AI usage on platform models is billed at cost plus a fixed, visible margin — with a budget cap you set yourself.
Become a design partner? The first five organizations use Scrum4Us free for six months, in exchange for weekly feedback and a case study.
Apply with your organizationPrices excl. VAT. AI usage on platform models: cost + 25% platform margin, visible in the ledger in real time. Turns on your own AI account are billed by your provider, not by us.
Frequently asked questions
No. Agents work in an isolated sandbox and deliver every change as a pull request that the platform verifies server-side. Merging is a human act — there is no auto-merge path in the product. That is a property of the architecture, not a setting.
Conversations are scoped to the individual user under database row-level security, expire automatically under your organization's retention term, and can be exported or erased per user through the platform API. Which AI provider processes them is your configuration: platform models, or your organization's own Anthropic or OpenAI account — with fallback off, there is never a silent switch.
Not yet. The foundations are in place (purpose binding, retention, private-by-construction conversations), but we take that step together with a customer, after a proper DPIA. We'd rather promise less than too much.
Then platform-billed usage stops — new agent turns are refused with a clear error until you raise the cap. Turns on your organization's own AI account with platform fallback off are billed by your provider and don't touch the platform cap; enable fallback and they count against it, since they may land on a platform model. No surprises on the invoice either way.
No. Your code lives in your own Git repository and is always yours. Knowledge documents are plain text and markdown you keep the originals of, and per-user conversation export is built into the API. Cancel monthly.
Five seats. Six months free. Your way of working shapes the product.
Apply with your organization