Claude and MCP consulting in Sweden: AI agents that work inside your own systems
Altu is a Gothenburg-based firm that builds on Anthropic's Claude and the Model Context Protocol (MCP): MCP servers that give Claude access to your systems, agents that read and work in your data, and the guardrails that make it safe to let them. We run the same setup ourselves – our own business system exposes more than 50 tools to Claude, and every change to existing data goes through a human approval queue. We work with companies across Sweden and with international teams that need a Claude-native partner in the Nordics.
What we do
- MCP servers for your systems. A layer of well-described tools on top of your CRM, accounting, project tool or data platform, so Claude Desktop, Claude Code or Claude in the browser can look up a customer, create a ticket, search documents or compile a report. We build against existing databases and APIs as readily as against systems we built.
- Governed writes. Reads are free. Writes pass through one shared layer: an empty field is filled at once, a change to an existing value is queued for approval, deletions are always queued. A person approves or rejects, and everything is logged.
- Identity per user. The MCP endpoint is protected with OAuth 2.1. Every call runs as a real person with that person's permissions. No shared API keys, and one connection can be revoked without touching anyone else's.
- Agent workflows in the business. Prospecting that fills the CRM, receipt handling that deduplicates and books to accounting after approval, follow-ups created from the sales pipe. Reasoning happens in Claude on the user's side; our server provides tools and data.
- Advice and training. Which tools should exist, how to describe them, where the guardrails go, and how a team works with Claude day to day. Workshops in Gothenburg or on site.
Why Claude
MCP is an open standard, so what you build on it is not locked to one vendor. Claude is good at tool use – picking the right tool, filling the right parameters, and knowing when not to act – which is the whole difference between an agent that helps and one that causes damage. And Anthropic's stance on safety matters to the CFOs, CEOs and boards we work with.
How it is built at Altu
We run Altu on a system we built ourselves, Altu Central Command: CRM, prospecting, receipts to accounting, projects, time reporting, calendar and mail. On top sits an MCP layer with more than 50 tools. Everyone on the team connects Claude to it under their own login and works in it in plain language.
A salesperson can write "which companies in Gothenburg have open quotes older than three weeks – create a follow-up for each" and get both the list and the follow-ups; those are new records and are created at once. Had the agent proposed changing a quote amount, the proposal would have landed in the approval queue instead.
That is the architecture we build for clients. The full description is in the Central Command case study (in Swedish).
Who it is for
Companies of roughly 10–200 employees that already have their data somewhere – a CRM, an accounting system, a few databases – and want an AI to work in it rather than chat about it. Swedish companies, and international teams looking for a Claude-native partner in the Nordics who can build in TypeScript against European data residency requirements.
We are a poor fit for a chatbot on a website, or for businesses without data of their own to work in. We say so in the first meeting.
How a project runs
- A conversation about which tasks should disappear. Not "what can AI do", but which three things in the working week take time and should run themselves. That gives the list of tools.
- A first step in production within weeks. A small number of tools against one of your systems, OAuth, the approval queue and the log from day one. Fixed price for the step, agreed before we start.
- Extension in bounded steps. More systems, more tools, agent workflows. Each step is quoted on its own. You own everything we build – code, database and tool definitions – and can take over or bring in someone else whenever you like.
You need a Claude subscription for the people who will use the agent. No model infrastructure of your own, and no customer data is sent to a model unless a user asked for it.
Stack
TypeScript throughout. MCP servers in Node and Next.js, Postgres via Supabase with row-level security in an EU region, hosting on Vercel, Anthropic's Claude via API for what runs server-side and Claude Desktop or Claude Code on the user's side for reasoning. Integrations with Fortnox, Gmail, Google Calendar and the databases you already have.
Questions we get
Do we have to replace our CRM or ERP?
No. We build MCP layers against existing systems whenever they expose an API or a database we can read. If nothing does, the first step is usually a data platform that gathers what you have – and you own it.
Can an agent damage our data?
Not in what we build. Reads are free, but every change to existing data and every deletion goes to an approval queue. It is not a setting that can be switched off by mistake; it is how the write layer is built.
Where is data stored?
In the EU, in your own accounts. Reasoning happens in Claude on the user's side; our server only returns what the tool returns.
What does it cost?
A fixed price per bounded step, agreed before we start. A first step in production takes weeks, not months. Get in touch and we will price your situation – and tell you if AI is not the right answer.
Do you work outside Gothenburg?
Yes. We are based at Kungsgatan in Gothenburg and work with clients across Sweden and abroad, on site or remotely, in Swedish and English.
Want to see it working? Get in touch – we will show our own system live.