Norwegian law your AI won't hallucinate.
lovspor gives ChatGPT, Claude and the AI tools you already use the actual text of Norwegian statutes and regulations — with exact citations you can verify, not invented paragraph numbers. It is an open MCP server over 5,890 acts and central regulations, with 18 tools that look up, search and verify.
- What
- An open MCP server: your AI assistant looks up Norwegian legal text instead of recalling it.
- Price
- Free, with no key to paste.
- Access
- You sign in with OAuth through a connector in Claude or ChatGPT.
- Contents
- Acts and central regulations from Lovdata's open data, and local regulations as observed on the municipalities' own websites — labelled as observed, not asserted to be in force.
Connect Read the documentation
A question you can ask
“Hvor lenge må jeg ha jobbet for å få sykepenger fra arbeidsgiveren?” (How long must I have worked to get sick pay from my employer?)
With lovspor the model looks up folketrygdloven § 8-18 “Sykepenger fra arbeidsgiveren” and quotes it:
«For å få rett til sykepenger fra arbeidsgiveren må arbeidstakeren ha vært ansatt hos arbeidsgiveren i minst fire uker (opptjeningstid).»
Translation: to be entitled to sick pay from the employer, the employee must have been employed by that employer for at least four weeks (the qualifying period).
The tool also checks the provision's own cross-reference, to § 8-7.
Ask instead about folketrygdloven § 8-99, and the tool answers that the section does not exist. For law, a plausible answer is a dangerous answer.
Connect
As a connector in Claude.ai and ChatGPT it works today: you sign in, and there is nothing to paste.
- Add a custom connector with the address
https://lovspor.no/mcp(in ChatGPT: turn on developer mode first, and choose OAuth). - In Claude: choose Register automatically for client registration.
- Sign in on the page that opens.
Or run it yourself, with no account. In Claude Code it is two lines:
uvx lovspor fetch-corpus claude mcp add lovverk -- uvx lovspor mcp
Step by step: guide for Claude · guide for ChatGPT · other clients
How well does it work?
We measure it with LLHB: the same questions about Norwegian law, put to claude-opus-5 without and with lovspor's tools, 250 cases per arm. Every citation in an answer is checked mechanically against the legal text; no language model judges. post-hoc diagnostic result, not a confirmatory one
| Without tools | With lovspor | |
|---|---|---|
| Answers containing an invented citation | 42/215 (19.5%) | 30/249 (12.0%) |
| Citation accuracy (citations that point at a provision that exists) | 128/189 (67.7%) | 431/494 (87.2%) |
| Correct provision identified | 8/90 | 30/90 |
| False premise rejected | 0/35 | 8/35 |
What did not improve.
- Out-of-corpus questions answered without inventing: 7/7 without tools, 7/11 with (resolved-case analysis).
- Find-the-provision answers containing an invented citation: 1/28 without tools, 8/40 with.
- Across 5 repeats the gain in citation accuracy holds, but the reduction in invented citations does not hold every time (variance measurement on a subset, not a level).
LLHB measures behaviour on this benchmark. It does not show that lovspor eliminates hallucinations or guarantees legally correct answers.
How it compares
Legal hallucination is well documented, but almost only for US law in English. Dahl et al. (2024) found that general-purpose models answering without sources hallucinated in 58–88% of answers about US federal court cases. Magesh et al. (2025) found that commercial legal research tools using retrieval (RAG) still hallucinated in 17–33% of responses. Established benchmarks such as LegalBench and LexGLUE measure legal reasoning and classification, not invented citations. LLHB is complementary: Norwegian statutes and regulations in bokmål, the same model with and without lovspor's tools, and every citation checked mechanically against a pinned legal text. In a post-hoc diagnostic run, answers containing an invented citation fell from 19.5% to 12.0%, but the difference was not stable across five repeats. The figures are not directly comparable: the jurisdiction, language, task definition, models and dates all differ.
References: Dahl, Magesh, Suzgun and Ho, “Large Legal Fictions”, Journal of Legal Analysis (arXiv); Magesh, Surani, Dahl, Suzgun, Manning and Ho, “Hallucination-Free?”, Journal of Empirical Legal Studies (arXiv).
Sources
- Acts and central regulations: Lovdata's open data (publicData), under NLOD 2.0.
- EU and EEA: links to the EU acts a Norwegian act implements, as CELEX identifiers.
- Local regulations: observed on the municipalities' own websites and published on the basis of åndsverkloven § 14, labelled as observed, not asserted to be in force. In MCP you get them only when you ask for them, and they are not published on this website.
What it covers — honestly
Covered
- National laws (lover)
- Central regulations (forskrifter)
- In force, with change history
Not covered
- Court case law
- Municipal zoning plans
- Proprietary standards (NS contracts)
lovspor is a research tool that surfaces the law — it is not legal advice.
Open source
The engine is open source under AGPL-3.0 and the corpus is built from Lovdata's open data. If you need Norwegian legal text inside an AI tool, you can run the whole thing yourself — without asking anyone, and without paying for it: github.com/bartoszkobylinski/lovspor.
One person operates the service. Questions and feedback: kontakt@lovspor.no.