For you. For your team. Shared context, defined workflows.
Connect your systems, bring your knowledge together, and set up recurring tasks once. Then use the reviewed workflows together – with clear permissions and traceable results.
Open source · MIT · Public beta in preparation – packages will follow once approved
You create your workflow in conversation with the AI assistant. Once you have reviewed and approved it, it runs in defined steps – on your own or as a team, with each person's permissions.
Describe your task – the AI assistant helps you turn it into a reusable workflow.
An example: A new customer pays. Someone notices the payment, copies the customer into the CRM, and lets the team know in chat. Three systems, the same steps, every time.
And when that someone is in a meeting, on vacation, or simply tired, the CRM is out of date and nobody has heard about the sale. Whether you are a solo business, a small team, or a department in a large organization, the knowledge of how to do things properly often lives in one person's head.
“When a new customer pays in Stripe, create the contact in HubSpot and post it in our sales channel in Slack.”A description you might give SubKit
“Every Tuesday, compare the supplier's price list with the shop, update prices, hide sold-out items, and report discrepancies over 15 percent.”A second example, from retail
The assistant suggests the steps and shows you the draft. You review it. If it is right, you approve it. From then on, it runs – on request, on a schedule, or when another system sends a notification.
Both descriptions are examples of workflows you set up yourselves – not ready-made templates included with the product. The same structure supports supplier reconciliation in your shop, a weekly report from three systems, and a knowledge base that answers questions with sources.
You define which systems a workflow may access and what it may do there. SubKit enforces these permissions and logs every run.
Stay in control of your data. Give AI only what it really needs for the task.
You choose which sources and connectors a workflow uses – and with which permissions. If a workflow includes an AI step, its inputs may be sent to the model provider you selected. More on this: Connectors, skills and sources.
EU Data Act · EU AI Act
The last point is the crucial one. Many tools can automate something. SubKit sets up workflows so you can trace and explain what happened afterward.
Use AI to organize knowledge, condense content and draft connectors or workflows. An authorized person reviews the results. After that, your team can reuse confirmed knowledge and run approved routines without having an AI plan the workflow again every time.
an assistant helps. It understands your sentence, looks at the system you point to and writes down the workflow. It organizes and condenses knowledge as a proposal. Everything stays a draft that a person reads; nothing is adopted automatically.
the approved workflow defines which steps happen and what they are allowed to do. Routines without an AI step run without a model call. If an AI step is explicitly configured, the workflow calls a model through a connector – at exactly that point, for exactly that task; this causes usage. Model responses and data from external systems can vary, so fixed steps do not always mean the same result.
That keeps what a run did traceable.
Every run follows the defined steps and is logged. Four steps, always in this order. Approval is the point where a draft becomes binding.
Describe the task, the tools involved, and the desired result – in the terminal or in chat. SubKit examines the systems involved and writes down the workflow.
Review the draft: Are the data, steps, and proposed changes correct? You see every step and which system it addresses. A trial run shows what would be sent – for a reconciliation, for example, a list of prices that would change. Runs against real systems have real effects; use the Test environment for them.
An authorized person defines which systems the workflow can reach and what it may change: which environment, which systems and addresses, and whether it may read or write. Approval can remain valid until revoked and carries over to each new version that does not expand its scope. If a workflow needs more, it needs new approval.
Start the approved workflow and see in the history which version ran and how it ended. On a schedule, on request, or when another system sends a notification – nobody needs to be present.
The underlying rule: A reviewed workflow defines the steps and permitted effects. AI can be involved at specific points; model responses and results from external systems may vary. Team chat and an AI agent run workflows; they do not rewrite them. Changes can be made step by step in the cockpit, as a file on the command line, or in the design chat – always as a new version for a person to read.
Nobody needs to learn a new application. SubKit shows up in the tools you already have – and those places are not interchangeable. You choose according to the task.
For an AI agent, every approved workflow is a tool.
SubKit is an MCP server for Claude Code, Codex, Cursor, and other AI clients. Every saved workflow appears there as a callable tool, alongside connectors, knowledge, and skills. The agent can start the workflow, read the result, query connectors, and look up knowledge it is allowed to see.
And the other way around: An MCP server your team already uses becomes a connector in SubKit. The admin decides, tool by tool, which of its tools may run without asking first.
What it cannot do: more than the person using it is allowed to do. Access that permits reading and running does not permit designing. Actions with external effects still require confirmation. And every call appears in the same history as everything else – an agent is not a back door.
That is the real benefit: Your team gives its AI tools access to real data sources and real workflows – with exactly the permissions each person has, and without anyone copying credentials into a prompt.
Tools such as Claude Code or Codex are great for experienced developers: They can do almost anything, and you have to watch them work. For colleagues in marketing, sales, editorial work, or administration, that is precisely what makes them cumbersome – and sometimes risky. Too many permissions, too many tools, too much freedom at each step.
SubKit turns that around. A workflow does one clearly described thing, following your organization's rules, on systems approved for that purpose. You use the same maintained workflows, sources, and rules with your individual permissions – not everyone has to set up their own coding agent on their computer.
Which systems: Shopify, HubSpot, Stripe, Google Workspace, Slack, Mailchimp, Shopware, Lexware Office, sevDesk, Mollie, PayPal, CleverReach, or Microsoft 365 – available connections depend on documented interfaces, configured access, and permissions. House rules relating to legislation are your team's requirements, not legal advice.
Ultimately, any authorized person in the company can carry out a colleague's tasks – through SubKit, under the same rules, along the same path. The knowledge stays in the organization, even when that person is on vacation or leaves.
Your domain knowledge, house rules and playbooks are available to the whole team. SubKit connects this context with approved data access and reusable workflows. Everyone sees and uses what they are authorized to.
The connection to one of your team's tools – with its own credentials, kept secure, and an environment: test or production. The assistant also helps you create connectors – without having to work through API documentation. With defined permissions, you can reuse them in other workflows.
“Anything over €1,000 needs two approvals.” “This provider reports errors with status 200.” Provider quirks, your house rules and playbooks as shared work instructions. That way the whole team reuses confirmed domain and process knowledge from one place. Skills help with drafting; limits are enforced by technical permissions and approvals.
The expense policy as a PDF, the internal API description, a website. Answers state what they read – or say that it isn't there. Sources have versions with an author and a schedule on which SubKit rereads an address. Who sees which knowledge and files follows their permissions.
A workflow can read, pass on, or generate a file you attach in chat – a report, an invoice, an export. It is stored in SubKit, not in a prompt. Permissions determine who can see which knowledge and files.
The direction: collect material in the inbox, review and prepare it in chat, then explicitly confirm it – only then does it become accepted knowledge or a skill available to the next draft. Raw excerpts are not confirmed knowledge. This complete process has not yet been publicly released. If a provider changes its interface, SubKit suggests a fix – it is applied only after approval.
SubKit is a framework for AI agents that design and fix workflows – producing a reviewed, versioned workflow artifact, plus the defined runtime for that exact artifact.
General-purpose agent tools such as Claude Code are designed to let a model work freely: with a shell, file system, subagents, and long runs. SubKit deliberately takes the opposite approach. The agent helps during design; during execution, it participates only where the workflow specifies an AI step. Between the two is an artifact that a person can read, review, and approve – and that then follows the same path every time.
That makes SubKit the central place for defined business workflows: one place where AI assistance, permissions, credentials, and history come together, rather than being handled separately in each tool.
A server, schedules, and updates. Credentials, sign-in flows, token renewal. Retries and error categories. Permissions, approvals, history. An interface and chat for colleagues. Trial runs, versions, test separated from production. And all of it again for the next workflow, because a script in a repository is rarely reusable for the next task.
Every workflow, connector, and skill can be reused for the next one, with a description and category your agent can read too. Your coding agent remains your tool for software. For your company's routine work, it designs workflows in SubKit and starts them there. See the comparison →
Connector (the system that holds the source of truth), skill (the rules the agent follows), and source (the documents it looks things up in) – linked by a shared category.
Workflows, connectors, sources, skills, and files share the same category, such as “Sales” or “Accounting.” Anyone opening a category sees everything belonging to that area of work that they are allowed to see – and whether it was created by a person or by AI.
Acceptance: You inspect a real run and check the result. Start with a task whose outcome you can judge confidently.
What three people wired together manually becomes a traceable operation. Who may do what, what needs approval, what happened when, what resources it used – all in one place.
If a provider changes its interface, SubKit detects it and suggests a fix instead of silently breaking. The fix is applied only after approval.
Transferring large numbers of records from one system to another and reconciling them field by field is a different task that needs different tools.
If nobody has described it and nobody has approved it, nothing runs.
One person, their own tools. A pricing page without an export: SubKit reads it daily and emails you when the price changes.
“How do we usually do this here?” – answered from your own documents, with a source citation. Or, honestly: “That is not documented anywhere.”
Forward contact requests to the team inbox according to fixed rules. Compile and send a report from five systems every morning. Maintain a product catalog according to defined rules.
Detect duplicate and outdated contacts and clean them up after review. Something needs a yes before it goes out – SubKit waits for a defined period, then proceeds or leaves it alone.
Your team's confirmed knowledge becomes something AI assistants can look up according to their permissions – instead of guessing.
These examples are possible workflows to set up, not ready-made templates included with the product or customer reports. Available connections depend on documented interfaces, configured access, and permissions. All use cases
SubKit is open source under the MIT license and can be self-hosted – if you like, reachable only within your company network or through your company VPN. You decide which services and model providers you connect. Calls to them can send data to these providers. Setup, updates and backups are part of running it.
The self-hosting offering is free. Infrastructure and external providers may incur costs. Multiple teams remain cleanly isolated as tenants.
Or start self-hosting → · Talk about your workflow
The public beta is being prepared. The reviewed installation guide and approved packages will follow.