AI that works in your tools. With your approval, on your server.
Important connects to the systems your company already runs on, and its assistants do real work there, only within the approvals you set and with every step on the record. When you want more, add a workspace for tasks, documents and knowledge, and modules for your own process.
- Open source, AGPL
- Your server or our EU hosting
- Your own model key
The customer in #214 says the HX-200 stops after a filter change. Find out why and draft a reply.
- Read the ticketGiteaservice/tickets #214
- Read the manualNextcloudHX-200 manual.pdf, p. 14
- Found the same faultGitea#188, closed in May
The flow counter was not reset after the filter change, so the machine stops as a safeguard (manual, p. 14). Ticket #188 had the same cause. I drafted a reply for the customer.
Service wants to comment on ticket #214 in Gitea
Anything a customer will read waits for a person.
service/tickets #214
Hello, after a filter change the flow counter has to be reset: Menu, Service, Counters, Reset. If the machine still stops, we will book a technician's visit.
Comment posted on ticket #214 in Gitea.
Nothing was posted. The draft stays in this conversation.
3 sources, 1 write to Gitea, $0.03 on your key
Start where you already work.
Important grows in three layers, and each one is useful on its own. Most companies start with the first and add the next only when it pays off. Nothing has to move that you want to keep where it is.
- From day one AI over your tools Assistants read from and act in the systems you already have. There is nothing to migrate and no new tool to learn.
- When a team is ready A workspace when you need it Tasks, projects, documents and knowledge in one place, for the part of the work you choose to move.
- When one process deserves it Modules for your process An app for the process only your company has, with assistants, approvals and history built in.
Your tools
Put AI to work in the tools you already have.
Connect the systems your team works in. Assistants read them with the same access people have, answer with the sources attached, and prepare changes that wait for a person before anything is written.
- Every answer points to the passage it rests on and says whether the source is fresh.
- When someone loses access to a file, assistants stop using it too.
- Long jobs keep running in the background, or on a schedule such as every morning.
Why do HX-200 machines stop after a filter change?
After a filter change the flow counter has to be reset, or the machine stops as a safeguard. Ticket #188 in May had the same cause and was solved by the reset.
Sources
-
HX-200 manual.pdf, p. 14Nextcloud, /Service/Manuals
Fresh
…after replacing the filter cartridge, reset the flow counter. Until it is reset, the controller treats the flow as blocked and stops the machine…
-
#188 HX-200 stops after serviceGitea, service/tickets
Fresh
Cause: flow counter not reset after the filter change. Reset on site, machine running since.
- Service price list 2026.xlsxAccess was revoked yesterday, so it was left out.
Remembered
After an HX-200 filter change, always reset the flow counter.
From HX-200 manual.pdf, visible to the Service team
What connects today
- Available Gitea Issues, comments and repositories. Assistants read them and, with approval, write back.
- Available Nextcloud Files and folders over WebDAV. Assistants read them and save changes with a version check.
- Available Any MCP client Claude Code, Codex and editors work inside Important, under the same permissions and approvals.
- Planned Microsoft 365, Teams, Redmine and email Next on the roadmap.
- Built for you Your own system A connector written on the SDK, usually as part of the rollout.
Workspace
Add a workspace when your team is ready.
When a team wants to move part of its work into Important, it gets tasks, projects, documents and knowledge in one place, and the assistants work in the same lists as the people. This is the app, recreated from its real screens.
Turn Monday's kickoff notes into tasks for this project. Take the owners and dates from the notes.
I read the kickoff notes. Three decisions there have no task yet, so I added them to the list as proposals, with owners and dates.
Kickoff notes, 22 SepPlanner wants to create 3 tasks in Line HX-300 start-up
Creating tasks in a shared project waits for your approval.
Created HAL-42, HAL-43 and HAL-44.
Nothing was created. The draft stays in this conversation.
1 source, 3 tasks, $0.02 on your key
- Short task keys like HAL-42 are easy to say out loud and paste anywhere.
- Every list works from the keyboard, and saved views keep each team's layout.
- Documents are published in versions, and assistants work from the published text.
Modules
A module for the process only your company has.
Every company has a process no off-the-shelf tool fits, such as complaints, service visits or quality reports. We build it as a module on the Important SDK, and because it runs on the same core, it gets assistants, approvals, the audit trail and history from day one.
Illustrative example
Quality wants to link CLM-0412 with CLM-0407 and assign the analysis to Anna K.
Both complaints are about batch 24-311, so one analysis answers both customers.
Linked and assigned. Anna K. has one analysis for two complaints.
Declined. Quality gets your answer and stops.
History
You approvedLinked CLM-0412 with CLM-0407, analysis assigned to Anna K.
Quality proposedLink two complaints about batch 24-311.Waiting for approval
Quality readCLM-0412 and 2 related complaints.
Anna K.Moved CLM-0399 to Decision.
What every module gets without extra work
- Assistants
- They read and act in the module through its operations, like everywhere else.
- Approvals
- Every write goes through the same approvals and rules as the rest of the system.
- Audit and history
- Who did what and when, with the sources behind it.
- Budgets
- The module's AI work stays within the same budgets.
- API and MCP
- Each operation is also available to scripts and MCP clients.
Today the SDK defines a module's operations, which assistants, people and the API all use under the same rules. A module's own data tables and screens arrive with the next SDK release.
Control
Nothing happens behind your back.
For every assistant you decide what it may do on its own, what waits for a person and what it may never do. When it needs you, you see the real change rather than a summary of it, and every decision stays on the record.
Service wants to comment on ticket #217 in Gitea
Anything a customer will read waits for a person.
service/tickets #217
Thank you for the photos. The seal is covered by the warranty, so we will send a new one on Thursday at no cost.
Details
Operation: gitea.issue.comment.create
Editor wants to change “Workstation instruction, HX-300”
Documents the whole company follows change only after approval.
Check the pressure once per shift.
Check the pressure at the start of each shift and after every filter change.
Details
Operation: documents.update
Planner wants to create 2 tasks in Maintenance
Creating tasks in a shared project waits for your approval.
Details
Operation: tasks.create
Record
You approvedPlanner created 3 tasks in Line HX-300 start-up.
Allowed by the rulesQuality read 14 reports without asking.
- A budget stops the job before an assistant overspends.
- A request nobody answers expires on its own.
- An approved change runs exactly once, even if the connection drops halfway.
Ownership
Yours to keep, not rented per seat.
Important runs on your server, or on a separate instance we host for you in the EU. You connect the model with your own key and pay its provider directly, and the price never grows with the number of people.
Which models may this assistant use?
- Claude Sonnet 5Anthropic, on your key Public cloud AllowedBlocked
- Any model through OpenRouterOpenRouter, on your key Public cloud AllowedBlocked
- Llama 3.3 70BOn your own server Local AllowedBlocked
Quality works on production data, so it may only use a local model. If the local model is down, the job stops. It never falls back to a public cloud.
For your IT team: use it from Claude Code, Codex or any MCP client.
Every tool is an operation in your workspace, with the same permissions, approvals and record as the app.
claude mcp add --transport http important \
https://important.example/v1/workspaces/<id>/mcp \
--header "Authorization: Bearer imp_…"- Open code
- The core will be published under AGPL-3.0 and the module SDK under Apache-2.0. Read it, audit it, change it.
- Your database
- Everything lives in your own PostgreSQL, which you back up and read like any other database.
- Your key
- You pay the model provider directly, and we add no markup.
- No telemetry
- The instance talks only to the model and the tools you connect.
Owning instead of renting
Owned, with ImportantRollout
We start with one process.
You are not buying an AI platform. You pick one process that costs your team time every week, and we run it in Important from start to finish. We own the whole rollout, from the server to the modules.
Discovery
We sit down with the people who run the process, see how it works today and agree on the numbers that will tell you it works better.
A pilot on one process
We install Important on your server or host it for you, connect your tools, set up the assistants and their approvals, and test the whole thing on real work.
Rollout and care
When the numbers hold, we add the next process, the workspace or a module. Care, updates and hosting are one subscription for the company, never priced per person.
Would rather run it yourselves?
The code will be published under an open licence at launch, and the Community edition is free to use in your company, with no limit on people. Today's install is for technical teams: Node, PostgreSQL and a few commands. A one-command install for Linux is on the way. Write to us and we will tell you when and how you can start.
What people ask before they start
Is our data safe?
Important runs on your server or on a separate instance for your company, and the data lives in your own database. Assistants read sources with the same access people have, and each step sends the model only what that step needs. There is no telemetry. For sensitive work, an assistant can be limited to a local model.
Which AI models can we use?
Anthropic, and any OpenAI-compatible service, including OpenRouter and models running on your own server. You always use your own key. An assistant limited to local models never falls back to a public cloud, even when the local model is down.
We already have our tools. Do we have to replace them?
No, that is where we start. Assistants work in Gitea and Nextcloud today, and connectors for other systems are written on the SDK, usually as part of the rollout. Microsoft 365, Teams, Redmine and email are next on the roadmap. The workspace is optional, and you add it only if a team wants it.
What if you disappear?
Your instance keeps running, because it runs on your server from code you have. You get the code under an open licence, the data is in your PostgreSQL, and any IT company can take over the care.
How is the cost structured?
The rollout is a one-off project, priced after discovery for the process you choose. Care, updates and hosting are a subscription for the company, never per person. You pay for AI usage directly to the model provider on your own key, and every assistant has a budget. The Community edition is free if you run it yourselves.
How is this different from Copilot or ChatGPT?
The model can be the same. The difference is what the AI knows and what it may do. In Important it works in your tools and your data with their permissions, finishes the job where your team works, and you decide what it may do on its own. It also runs on your server rather than the vendor's.
Is it ready?
It is not publicly released yet. We run it every day on our own server, and everything on this page that does not exist yet is labelled as planned. The first rollouts start as pilots on one process, with the scope agreed in writing.
Which licence will it have?
The core will be published under AGPL-3.0 and the module SDK under Apache-2.0. Any company can install it, use it commercially with no limit on people, and change it for its own use.
Start with one process.
Tell us which process costs your team the most time. In the first conversation we will tell you honestly whether Important is the right tool for it.