In TWIM this week, Nico of Famedly wrote: "For the last almost 2 years my team at Famedly have been secretly working with the BWI on a new homeserver for them. Now we finally published the repository and will continue working in the open from now on." The repository is Arminius, in the BundesMessenger backend group on GitLab — the backend arm of the German armed forces' messenger platform, and a different organization from the FITKO Neo citizen channel this site covered in August. A 26:47 conference talk, Why Is There Rust in our Matrix?, taped at The Matrix Conference in Malmö on October 22, is already on media.ccc.de under CC BY-SA.

Facts, API-verified as of November 7: project registered 2026-07-24, public, description "A new homeserver.", 9 stars, 0 forks, last activity 2026-11-06 15:29 UTC. Nine stars means nobody outside the build team has run this. That number is worth more honesty than the headline volume around the announcement.

What Arminius is, and is not

It is an independent homeserver implementation in Rust, designed for enterprise deployments: PostgreSQL for state, NATS JetStream for internal messaging, an in-tree Helm chart that is itself under active development, conformance via Matrix Complement plus a custom E2E test suite that federates Arminius against Synapse through the Matrix Rust SDK. The README is Apache-2.0 via SPDX header (individual files MIT or CC0, licenses in a LICENSES/ directory) — "permissive open-source licensing" is listed as a feature, and BWI's own talk abstract says the Rust choice traces back to the Matrix Rust SDK work BWI did with Element, and to an ambition of "a communication stack where the functional core is formally verifiable." That is a direction stated in an abstract, not a property demonstrated by an artifact. File it as stated ambition.

This site's license floor leans AGPL because copyleft keeps forks in the open; Apache-2.0 is the weaker pull of the two — a proprietary embed is someone's legal day one. That is a real consideration for anything a community might standardize on, and it is also not the story here. Read the ADR before judging the license; a format that documents exit does its work even under MIT.

What it is not, per its own documentation: ready. The maintainer wrote "the homeserver is not done yet. So please do not actually use it yet. We want to first make sure it behaves as a proper citizen of the federation and we still have a few complement tests to pass as well as do a proper audit." The README's SECURITY.md link says (TODO) — meaning there is currently no private vulnerability-reporting channel at all, which by the standard this site applied to Zendrite's licensing files is a hard production blocker. The Kubernetes development setup is a TODO. The compatibility story is honest by exclusion: a public Complement blacklist tracks what is not yet supported rather than a changelog claims what is. And the FAQ answers the certification question plainly: no, Arminius is not certified for classified information; MLS-driven components are being built toward VS-NfD. The FAQ also dates it: first productive release planned for Summer 2027.

The landscape positioning in the FAQ is unusually candid for a vendor document: Conduit and Continuwuity are "great for small appliances, hard to scale for HA setups"; Tuwunel is "even greater for small appliances"; and Dendrite's architecture "was mostly the blueprint for Arminius." Independent implementations keep arriving — the census in the same TWIM counts 20,461 discovered servers, with Synapse at 15,634 (76.4%), Continuwuity 1,970 (9.6%), Conduit 519 (2.5%), Dendrite 264 (1.3%). One in ten discovered servers now runs Continuwuity. The diversity is real; conformance tooling is how it becomes interop instead of hope — which makes Arminius pointing its own tests at Continuwuity (their merge request 2641) more relevant than any feature list.

The artifact: a migration format that documents what leaves

The part that repays reading is not the server. It is the homeserver migration ADR, published with the code.

The format: a streaming newline-delimited JSON export. A header record (format_version, server name, exported_at, per-table record count, and a cursor_param name), a trailer record that exists to prove the stream was not truncated, and fifteen data tables emitted in dependency order — private_keys first (signing keys move with the state), then users, rooms, devices, events, account data, presence, receipts, device keys, cross-signing material, one-time and fallback keys, to-device messages. Every timestamp carries a _ms suffix. Chunking is built in: the importer resumes via a cursor whose parameter name the exporter owns and advertises in the header, so the resume contract can change on one side without breaking the other. This is a designed-for-failure export, not a database dump with a README.

The schema tells you the first exporter/importer pair outright: the sort-key table names Synapse columns (stream_ordering, stream_id, composite (user_id, device_id)) and the ADR notes "Synapse only has one signing key." So the concrete deliverable is Synapse → Arminius migration. The FAQ's own sentence about the wider problem is the best statement of it: "While most homeserver implementations doesn't plan or cover migration from other homeserver implementations, we will provide this functionality. We also attempt to provide a standardized migration format that could be used for generic migrations between Homeservers." Nico's TWIM post confirms the format is planned to become an MSC once they have experience with it.

The ledger of what does not migrate, which is where the ADR earns trust, because most exit documents quietly omit it:

  • Media files are not in the export. Events carry mxc:// URIs; the bytes move by filesystem copy of the media store, before or after import. Remote media is re-fetched as needed.
  • Access tokens are not exported. Every session on the old server dies. Plan the re-login wave — a full fleet of users, at once.
  • Pushers are not exported. APNs/FCM registrations re-arrive when clients reconnect.
  • Pending to-device messages from remote senders are dropped on import, which "may cause one-time decryption failures for in-flight E2E sessions" — affected users re-request keys from their peers. An honest cost of moving an encrypted system.
  • User threepids (emails, phones) are deferred — "exported later if needed" — which is exactly the kind of ambiguous clause to watch if this becomes an MSC.
  • State snapshots are recomputed by replaying state resolution over the full import in topological order, which is expensive on large rooms. Budget it.

Why this matters beyond one server: a defined wire format for the complete state of a communication system is the chat-scale version of what the EU Data Act made an architecture requirement for clouds. Egress that is documented, chunked, truncation-checked, and resumable — versus egress that is an archaeology project. Belgium got exit by forking because its vendor froze; a migration format is the no-fork-required rung below that. If the MSC lands cross-implementation, moving a community between homeservers stops being the reason nobody ever leaves.

What to actually do

Running a community chat server today: the July landscape map stands and nothing from this announcement changes it — Continuwuity as the default for simple deployments, Zendrite as the HA pilot candidate, Synapse where the worker farm already exists. Arminius joins the map as a Summer 2027 watch item with a dated target, not as a candidate. Its own authors said so before any outsider could.

Running Synapse: you are the first beneficiary. A supported, documented egress path — even one aimed at a server you cannot yet use — is the migration safety rope every guide until now told you to hand-build. But the ADR is explicitly a draft awaiting experience; do not build tooling against it, and do not let anyone sell you a "migration" product that claims to implement it.

Everyone, one question worth adopting: when a project ships a server, ask for its exit document in the same breath as its install guide. What leaves, in bytes; what breaks, in sessions; what it costs, in downtime. Arminius answered before being asked. That is the bar.

Same week: the Dendrite fork keeps compounding

Zendrite v3.7.0 shipped in the same TWIM, and it reads as the continuation of the October post: /versions now advertises Matrix v1.16 through v1.19; QR-code sign-in (MSC4108 with MSC4388 rendezvous), thread subscriptions (MSC4306), and sliding-sync profile updates (MSC4262) arrived with it. Two of the additions land directly on the open-registration spam playbook: policy-server screening of events before acceptance (MSC4284) and user-level invite blocking (MSC4380) — the fork is picking up the moderation layer the ecosystem has been improvising room by room. And the release repairs corrupt stored room state — rechecking rejected events against repaired state, skipping rejected outliers, taking unresolved memberships from other servers in the federation. A fork adopted as the escape from a security-fixes-only upstream five weeks ago is now closing upstream's functional gaps at a cadence that continues to argue for the reassessment.

Closing

Two things became public this week that do not require anyone to trust a new homeserver: a migration format that writes down exactly what leaves when a chat platform is moved — including what breaks — and a conformance harness pointed at implementations other than its own. Both are the unglamorous parts, and both are the parts that decide whether federation keeps its exit properties as the server count grows. The server itself gets judged in Summer 2027, on a schedule its authors set themselves. That is the right way to read the announcement — artifacts first, server later, hype not at all.

Sources: TWIM 2026-11-06, Dept of Servers; Arminius repository and GitLab API project record; Arminius FAQ; Homeserver migration ADR; Why Is There Rust in our Matrix? — media.ccc.de; Zendrite v3.7.0 release notes; MatrixRooms.info federation stats.