Lunara Intelligence — MCP server
Regulatory obligations for AI systems, with the article that sets each date and a
link to primary law.
Your model very likely knows some of these dates wrong. The EU AI Act was amended
on 27 July 2026 by Regulation (EU) 2026/1744, which deferred the Annex III
high-risk obligations by sixteen months — five days before they would have
applied — while leaving Article 50 exactly where it was. Any training corpus
assembled before that carries the superseded date, and the two are quoted together
often enough that getting one right is no evidence of getting the other right.
This server exists so a model can look the answer up instead of recalling it.
Connect
The hosted endpoint is live. The npm package is not published yet — the
install command below resolves to nothing until the @lunara scope is
claimed, and this repository refuses to advertise it in the meantime.
Point any MCP client that speaks Streamable HTTP at:
Once the package is published, this is the local alternative:
Node 18 or later. No dependencies, no API key, no account. Reading is free and
always will be: a claim nobody can check is worth nothing.
If you cannot run a local process, the same server answers over HTTPS:
Streamable HTTP, one POST per JSON-RPC message, stateless, no key. Both
transports import the same module, so they cannot drift into giving
different dates.
Integrity
Every document this server reads is published with a detached Ed25519 assertion
over its SHA-256, and the server verifies it on each fetch. A corpus that fails
verification produces an error, not an answer — a client that reports a bad
signature and then answers anyway has told you something is wrong and acted as
though it were not.
Verification being unavailable is different from verification failing, and
they are reported differently: a runtime without Ed25519 in WebCrypto, or an
authority serving no assertion, reads as unverified rather than being dressed up
as verified. Keys are at /.well-known/keys.json, the method and a worked
example at https://lunarasociety.com/signing.html.
Tools
What it will not do
It does not answer when it cannot reach the authority. There is no bundled
copy of the corpus. A cached table would let this server keep answering with a
date that was amended months ago, which is exactly the failure the institution
publishing it has already had to correct. If lunarasociety.com is unreachable,
you get an error explaining that, not a remembered date. An oracle that guesses is
worse than an oracle that is down, because you cannot tell which answer you got.
It will tell you nothing applies. Three outcomes of the applicability model
conclude that no obligation in the corpus reaches the asker, including the one
most people hit: the California AI Transparency Act reaches generative systems
above one million monthly users, and below that line it imposes nothing. A
qualifier that only ever escalates is a sales funnel.
It is not legal advice. Obligations are marked verified — checked against
primary law, with that source linked so you can reach the same conclusion without
trusting us. The applicability model is marked interpretation, because which
obligation catches you is our reading. You are free to take the facts and leave
the reading.
Where the answers come from
https://lunarasociety.com/corpus/obligations.json, generated from a single
obligation table and gated in CI so the published corpus and the site can never
disagree. Tense is never stored anywhere — a static document does not know when it
is being read, so it states application dates, and this server computes "in force"
against the moment of your call.
Corrections, including two we published against ourselves and the policy that
requires it: https://lunarasociety.com/evidence.html
If an entry does not follow from the source it cites, that is a defect.
[email protected] — we would rather be told than found out.
