Drop in a PRD. Skaftor turns it into approved requirements, a design and work orders — then coding agents build them, with every decision kept on the record.
Plugs into the tools you already run
Payment retries, from a PRD to a canary deploy: the requirement, the design that cites it, the work order an agent builds, the gate that checks the diff against the criteria, and the pull request your lead merges.
Drop a PRD. Skaftor asks what it leaves open, then drafts requirements with acceptance criteria you include or exclude line by line.
An agent reads your repository on a workstation and writes the technical design section by section, each tied to the requirements it covers.
One pull request's worth of work each, mapped to the repository it changes and ordered by dependency. You confirm where each one lands.
Claude Code, Codex, Gemini or DeepSeek on a fresh Skaftor Cloud workstation — or your own editor with the context wired in over MCP. Priced before you start.
An AI review, then a clause check that reads the diff against the acceptance criteria. CI has to be green. Nothing counts as done until the gate passes.
Test cases drafted from the criteria and run by QA, with a browser QA agent gathering evidence. Repo-wide deployment gates, then a canary deploy to your cloud.
Requirements live in a doc, decisions in a chat thread, approvals in someone's head. When agents ship ten times faster, the thing that breaks is the connection between them. Skaftor keeps it.
Ask why a line exists and the record answers: the requirement, the design decision that settled it, the pull request that landed it and who approved it. New engineers and new agents start from the same place.
A work order carries its repository, branch, service and acceptance criteria. Claude Code, Codex, Gemini or DeepSeek build it on a Skaftor Cloud workstation — or your editor does, over MCP — and you see the cost as it runs.
CI, an AI review and a clause check that reads the diff against the criteria. A named role approves each stage, and the approval is recorded — so an auditor can see how a change got to production.
Every pull request is read against its acceptance criteria. When one isn't met, the gate says which, shows what it looked at, and tells the author what it takes to pass.
Your pipeline runs first. Red CI stops the gate before anything else is read.
Reads the diff against each acceptance criterion and names the evidence — never the agent's summary.
A scored review with findings for the lead. It informs the merge; it doesn't make it.
Rules proposed from your own code. Each one blocks the deploy or warns only — set per gate.
Per work order, per QA run and per month. A run that hits its cap is held, not finished.
When shipped code stops matching its spec, someone is told — fix the code or update the spec.
Three companies run Skaftor in production and shape the roadmap with us. The logos in the band under the hero are integrations, not customers.
We run stored value across every channel — balances have to reconcile everywhere, every time. Skaftor keeps the reasoning and the blast radius attached to each change, so we ship faster without gambling on what it touches.
Our platform is a web of third-party travel APIs. The hard part was never writing the code — it was remembering why each integration works the way it does. Skaftor turns that into something the whole team can query.
On a statutory billing surface, a wrong invoice is a compliance event, not a bug. Skaftor's gates mean every change clears its checks before it ships — and we can show an auditor exactly how it got there.
Each work order runs on a fresh Skaftor Cloud workstation with the repo, branch and agent ready — or in your own editor, with the context wired in over MCP. The agent writes the code, opens the PR, and reports back.
Add your own AI key. Every requirement, design, build, review and QA run is metered on it — with caps per work order, per run and per month that the gateway enforces.
Anthropic, OpenAI, Google Gemini or DeepSeek — add the key once. Skaftor mirrors the spend; nothing is marked up.
The Build sheet shows the estimate before you start, and each work order carries what its agent actually cost.
Per work order, per QA run and per month. Cloud agents go through Skaftor's gateway, so the next request past a cap is refused.
Team is a flat seat price on your own AI key. Growth adds managed workstations and is priced on usage. Enterprise runs in your own cloud.
The full platform for a team that ships.
For organizations that want agents building in the cloud.
Unlimited seats · managed workstations included.
For large and regulated teams that keep data in their own cloud.
Volume pricing, security review, and a hands-on rollout.
Cancel anytime · AI usage billed by your provider, on your own key.
Skaftor is the system of record for AI-native software delivery. It captures business and technical intent, compiles it into agent-executable work orders, orchestrates AI coding agents (like Cursor and Claude Code) over MCP, governs every change with gates and blast-radius checks, and remembers how and why each change shipped in an engineering memory graph.
Those tools help an individual developer write code faster in the editor. Skaftor is the organization-wide system around them: it captures intent, orchestrates the agents from structured specs, governs the output with approval gates, and remembers the full history. Skaftor orchestrates coding agents rather than competing with them.
It's a way of building software in which humans express intent and AI agents do most of the implementation, while a system of record captures the intent, orchestrates the agents, governs each change, and remembers how and why every line was shipped.
Yes. Build in Cursor, Claude Code, Codex or VS Code with the work order's context wired in over MCP, or let Claude Code, Codex, Gemini or DeepSeek build it on a Skaftor Cloud workstation. Either way the work is dispatched, governed, and observed centrally.
Every workflow step names the role that approves it. Each pull request passes CI, an AI review, and a clause check that reads the diff against its acceptance criteria; deployment gates and QA sign-off guard the release; the Drift Inbox catches shipped code that stops matching its spec. Every artifact is linked in a knowledge graph, with roles and an audit trail.
You do, directly. Skaftor runs on your organization's own Anthropic, OpenAI, Gemini or DeepSeek key, meters every call, and stops at the budgets you set — per work order, per QA run, and per month. Nothing is marked up.
Team is $99 per seat per month for up to 30 seats. Growth adds unlimited seats and Skaftor Cloud workstations where agents build. Enterprise, which runs Skaftor's content, code and AI runs in your own cloud, is coming soon. AI usage is billed by your provider on your own key.
More in the glossary and What is Skaftor?
A zero-dependency terminal client — pull work orders, open pull requests and launch workstations over MCP without leaving the shell.
npm install -g skaftorRequires Node.js 18+ · macOS, Linux, or WSL. The skaftor package is published on npm.
skaftor login <your-login-line>Copy it from a work order's Build sheet or Settings › Workstations.
skaftor wo nextPull your next work order, with its plan and execution manifest.
skaftor wo start · skaftor wo doneWork it in your editor, then attach the pull request.
skaftor up --agent claudeOr launch a Skaftor Cloud workstation and let the agent build it.
contextusewhoamiwo nextwo minewo listwo showwo planwo manifestwo startwo askwo syncwo donewsupsshstartstoprmtargetsverifyprenvtoolscall--jsonTurn intent into shipped code — on one traceable thread.