Three ways to get started with SubKit.
Two ways in: request a beta invitation or host SubKit yourself. Then, depending on your situation: you want to set it up yourself. Someone already runs SubKit for you and you want to describe your first workflow. Or you work with a coding agent or AI client and want to use SubKit from there.
The public beta is being prepared. Installation packages and instructions will follow once it is released. This page already explains what to expect.
Request a beta invitation →Start self-hosting
Start self-hosting
Guidance from prerequisites through the package, startup and setup to your first harmless workflow.
Path 2Describe your first workflow
Connect a system, describe what you want in a sentence, review and approve.
Path 3From an AI agent
Connect SubKit to Claude Code, Codex, Cursor or a browser-based AI client.
Start self-hosting
SubKit runs as a complete stack: the Engine, Cockpit and required services, including the database, run together using the Compose file for that release. A computer with Docker is enough to try it; later, the same stack runs on your server.
Check prerequisites
Check off what you already have. If anything is missing, sort it out before downloading—it will save trouble later.
Next: your own server.
The same stack and workflows. Approvals apply per environment — approve production separately and deliberately. Multiple teams run as separate tenants on one Engine. You can run SubKit so that it is reachable only within your company network or through your company VPN. Your team can then work in the browser on the go as well – on approved private computers, iPhones and iPads. Your IT sets up the VPN and approves the devices; SubKit does not include a VPN of its own. Sign-in, permissions and workflow run logs still apply.
Operations are part of it.
Always update using the new version's Compose file, back up the database regularly and test restoring a backup. Infrastructure and model providers can incur costs.
Describe your first workflow
Does someone already run SubKit for you? Then you do not need to install anything. You need access, a connected system and a sentence describing what should happen.
Sign in
An invitation gets you straight in; otherwise, sign in through your browser. You get the command-line tool from your team's engine.
Connect a system
A connector connects to one of your team's tools, based on the description of its documented API. Before saving it, SubKit actually tests the connection. You attach credentials separately; they remain encrypted and are never exposed. Stored credentials alone do not authorize any changes.
Describe
In the design chat—in the cockpit or on the command line—you describe to the AI assistant the task you want to automate. It examines the systems involved, proposes the steps and shows them to you. An empty chat starts with cards showing your own workflows.
“When a contact inquiry comes in through the website, email it to the sales inbox.”This is what a first sentence looks like
Review
A trial run shows what would happen with real inputs: it reads for real, holds back everything that would be sent, and shows it to you. This runs automatically when you create a workflow; you can repeat it at any time. Runs against real systems have real effects—point a connector at the Test environment for your first attempts.
Approve and let it run
An authorized person approves what the workflow may access: which environment, systems and addresses, with read or write access—until revoked or for a limited time. Approval carries over to each new version that does not expand its scope; a broader scope requires new approval. From then on, it runs on request, on a schedule or when another system sends an event.
See everything that needs attention in one place.
The “Needs you” list in the cockpit shows everything waiting for you: proposed fixes, suggested rules and recent failures. The same overview is available on the command line.
Changes take a readable form.
Team chat and an AI agent run workflows; they do not rewrite them. You make changes step by step in the cockpit, as a file on the command line or in the design chat—always as a new version that a person reads and that can be restored.
From an AI agent
For AI clients, SubKit is an MCP server. Once connected, every approved workflow appears as a callable tool, alongside connectors, knowledge and skills—with the permissions of the connected access credentials and no more. Every call from an agent is recorded in the history along with its usage.
Issue access credentials
An API key provides separate access with its own permissions. Access limited to reading and running cannot design or change anything from an agent. The key is displayed once—store it securely.
Configure it in your client
Claude Code, Codex, Hermes, Cursor and similar tools launch the command-line tool locally as an MCP server; it forwards calls to your team's engine. A browser-based AI client connects directly to your team's engine and signs in using SSO or a key.
What happens here: You add SubKit to your client's MCP configuration. The verified commands and downloads will follow when the public beta is released.
Use it
The agent finds workflows, starts them, reads results, queries connectors and searches the knowledge it is allowed to access. Actions with external effects ask for confirmation first. Every call appears in the same history as everything else.
- Find and start a workflow“What workflows are there for invoices?” — “Start the daily report.”
- Look up knowledge“How do we handle refunds over €1,000?” — with a source citation
- Track a runIf something takes longer, the agent receives the result as soon as it is available
For coding agents: start the session with your team's knowledge
The intended direction: in a repository, SubKit opens every session with a short summary of the confirmed rules in your domain. What the session has learned can be submitted—reviewed and prepared in chat, then explicitly confirmed by a person. Only then does a fact become knowledge or a practice become a skill. This path has not yet been fully released to the public.
Team chat instead of an agent?
Connect Slack and Google Chat through the command line, and Teams through settings. The SubKit bot then answers questions and starts workflows there—with the same permissions and the same history.
Talk about your workflow
A short conversation about where to start. Not a sales meeting. Contact →
Five steps, five workflows.
This is how a team gets started without turning it into a major project. Each step has a goal; how quickly you progress depends on your systems and access credentials.
Something harmless that demonstrates the four steps
Connect Gmail, describe a weekly digest of all runs, read the preview, approve it, let it run and open the history entry. Afterwards, everyone in the room knows the four steps: design, review, approve and start.
The weekly report
One sentence, three systems, one email. This is where most people understand the preview: the email is ready before anyone receives it.
The first write operation, in test first
Send inquiries to the CRM. The connector points to the Test environment, and the approval dialog names both write targets. Only once the run in Test is correct do you use the same workflow with the production connector and a separate approval.
House rules
The price limit as a skill, the house style as a skill, the price list as a source. From now on, every new draft knows these rules without anyone having to repeat them.
The team
Invite colleagues, connect the bot to Slack or Teams, define roles: who starts workflows, who approves them and who only views them. Set a budget. From now on, the “Needs you” overview is where the day begins.