What's new

What's new in Eldric 5.0.

Eldric 5.0 is generally available — released 2026-05-24 as v5.0.0, patched through v5.0.38. A summary of what shipped, grouped by area, with links into the detail pages. Gaps or questions: office@eldric.ai.

Operating your cluster

One console for every node and model.

Run your whole cluster from the admin console — and let Eldric learn from how you work. Assign node roles, manage every model across the cluster, and turn your own sessions into reusable skills — all from the GUI, nothing leaving your hardware.

Operator work that used to mean editing files on each node lives in the admin console, and an opt-in learning loop turns the way you work into skills you can review and reuse.

Manage the whole cluster from one console.

Day-to-day cluster operations live in the admin console — no per-node file editing:

What you can do

Node roles from the console — choose what each node does (inference, data, edge, and the rest) with checkboxes instead of a startup flag.
Cluster-wide model storage — see every model on every node from one place, with size, date and where each one lives, and load, unload or fetch them without leaving the page.

Eldric can learn reusable skills from your sessions.

An opt-in learning loop reviews completed conversations on a schedule you set and distils the repeatable procedures in them into named skills. You stay in control: turn it on, pick the model it uses, set the cadence, and review what it proposes before anything is kept. Nothing leaves your cluster.

Honest scope. Off by default and admin opt-in. It proposes skills for your review — it does not act on them autonomously.

Latest patches — 5.0.20 → 5.0.38

What landed in 5.0.20 → 5.0.38.

A long run of hardening — and a preview worth watching. A tiny specialist model that renders an earthquake in your chat, plus native inference, a cluster-wide model store, and one-operation updates.

Most of this run was platform hardening you don't see. One thing you can: a new kind of tiny, on-device specialist model — shown here on seismic waves.

A specialist model that renders an earthquake — in your chat.

Preview · visualization, not prediction

Ask, in plain chat, to see a real earthquake, and a roughly one-megabyte self-contained model renders the seismic wavefield spreading out from the source — frame by frame, on the CPU. No GPU, no cloud; it's small enough to run on a Raspberry Pi. It's a specialist model — one tiny expert for one physical domain — built on the xLSTM architecture.

Honest scope. It visualizes the wavefield of a known earthquake through a known physical model — it is not earthquake prediction and it does not image the subsurface (a scoped simulation of a known model, not forecasting). A preview of a new family of compact, on-device specialist models.

The platform got harder to break.

Available today

A long run of reliability work landed across these patches:

What changed for you

Native inference — run models directly, with no external backend to install.
Cluster-wide model storage — upload a model once and it appears in every node's menu, pulled on demand.
One-operation cluster updates — roll an update across every node from a single action.

Latest patches — 5.0.17 → 5.0.19

What landed in 5.0.17 → 5.0.19.

Answers that just work, and an install that just runs. Tool answers fire server-side regardless of the model; a plain install serves the whole platform; structured-ML runs natively.

Two patches focused on reliability: the platform now drives the right tool for a question itself, a vanilla install serves everything out of the box, and the structured-ML workloads moved into the main service. What's live today, with honest scope.

Tool answers in chat — fired server-side, model-independent.

Available today

Ask a forecasting question, a natural-hazard question, or a drug-interaction question and the platform now recognises the intent and runs the right tool itself — returning a real, sourced answer card. It works the same whether you're on a local model or a cloud one, because the dispatch happens server-side rather than relying on the model to call the tool.

What changed for you

Forecasting / hazard / FDA drug-interaction questions return a real result card, not a guess.
Consistent across every model — local or cloud — because the platform drives the tool.

Honest scope. These are data lookups and series forecasts — the platform extends a series or fetches a sourced record. They are not predictions of real-world events.

Install just works — the whole platform, out of the box.

Available today

A plain package install now serves the full platform — chat, agents, retrieval, the structured-ML workloads — with no manual role configuration. Install, browse to the chat shell, go.

What changed for you

dnf install eldric-aios and the platform serves everything — no role flags to set.
Updates stay atomic; configuration lives in the admin GUI.

Structured-ML runs natively — no separate daemon.

Available today

The structured-ML workloads — time-series forecasting, control policies, and associative retrieval — now run inside the main service rather than a separate process. Fewer moving parts to deploy and operate.

What changed for you

Forecasting, policy execution and retrieval are served by the main platform directly.
One service to install and update for these workloads.

Latest patch — 5.0.16

What landed in 5.0.16.

The chat shell became the place you do the work. Interactive forms in chat, a guided knowledge-base wizard, and multi-agent swarms you opt into per tenant.

5.0.16 builds on the 5.0 platform with the form-hub keystone — tools that need structured input now render a form right in the chat, no admin-panel detour. Below: what's live today, with honest scope on what's still in flight.

Chat-first forms — tools render a form in the chat.

Available today

When a tool needs structured input — a knowledge-base name, a connection string, a set of parameters — the chat now renders an interactive form inline instead of asking you to hand-craft a request or jump to an admin screen. You fill it in, submit, and the platform invokes the tool server-side and renders the result back in the conversation. The chat shell becomes the single place to drive the platform.

What changed for you

Invoke tools from chat by filling a form — no JSON, no admin-panel detour.
The result comes back in the conversation as a card.
One consistent form experience across the chat shell.

Knowledge-base wizard — guided retrieval setup.

Available today

Creating a knowledge base for retrieval no longer means knowing the API. A guided wizard walks you through it — pick the source, set chunking and embeddings, and the platform stands up the namespace and ingests your documents. Semantic search over your own content, set up in a few steps from the shell.

What changed for you

Stand up a knowledge base from a guided flow, not raw API calls.
Source, chunking and embedding choices come with sensible defaults you can override.
Your documents become searchable context for the agents.

Multi-agent swarm — opt in per tenant.

Available today · admin opt-in

Multi-agent orchestration — several specialised agents collaborating on one task across a chosen topology — is available and gated behind a per-tenant switch. An admin turns it on for the tenants that should have it; it stays off everywhere else. Paired with per-tenant feature toggles, you decide exactly which capabilities each tenant sees.

Honest scope. Swarm is admin-opt-in, not on by default — a deliberate choice so the capability is enabled where it's wanted and invisible where it isn't.

What changed for you

Turn multi-agent swarms on per tenant from the admin shell.
Per-tenant feature toggles gate this and other capabilities cleanly.

Manage your device & network fleet from one panel.

Available today · live control on the way

A device-fleet panel plus a networking-plugin catalogue: bring the switches, routers, firewalls and IoT devices you already run into one surface, with platform-aware plugins (Cisco, Fortinet, Aruba, Palo Alto, HPE and more) over SNMP / Redfish / RESTCONF. The AI reads your fleet and reasons across it — triage, drift, capacity, postmortems.

What changed for you

One panel for your network estate + a plugin catalogue, discovered and managed from the admin GUI.
The AI reasons over the fleet you already operate — it reads your management plane, it doesn't replace it.

Honest scope. Today: the catalogue, the fleet panel, discovery, and AI that reads and reasons. Live device control / change dispatch through the platform is on the way — approval-gated by design.

See how Eldric compares to the alternatives →

Cumulative since GA

What landed across 5.0.7 → 5.0.11.

The first patch wave after GA was substantial. Multi-controller HA, multi-site federation, an extension marketplace, and a networking plugin catalog.

Below: the nine areas that grew most. Each carries an honest status — what's live today, what's in flight for an upcoming 5.0.x patch, and where the gates still are. We name what's pending; nothing surprises you in production.

Multi-controller HA — production-grade Raft consensus.

Consensus today · Full HA bootstrap in flight for an upcoming 5.0.x patch

A multi-controller Eldric cluster now forms, replicates state, and fails over under live test. The consensus story has lifted from "designed and in progress" to "validated end-to-end under production-like conditions" — controllers form a quorum, replicate the durable log, and survive a primary failure with no manual hand-holding.

Honest scope. Production-grade HA bootstrap on the primary controller is gated on two pieces still in flight for an upcoming 5.0.x patch: cross-controller identity replication (so any controller can authenticate any user), and a leader-aware client endpoint (so clients always reach the current leader without operator intervention). The consensus layer is real; the full operator story is not yet. See clustering & HA for the live status.

What changed for you

Run three controllers, lose any one, the cluster keeps serving.
Crash-recovery from the durable log is automatic — no replay scripts.
Add a controller to a running cluster without restarting workers.

Multi-site federation — seven layers shipped.

Available today

You can now run more than one Eldric cluster and have them act as one platform. Choose a hierarchical topology (one regional aggregator on top of branch clusters) or a peer mesh (every site talks to every other site directly). Cross-site routing, regional aggregation, lateral sub-cluster traffic, four discovery mechanisms, an install-time wizard that asks the right questions, and customer-facing federation dashboards — all live.

What changed for you

One sign-in, one knowledge base view, queries answer across all sites.
Pick hierarchical or peer-mesh on first install — the wizard walks you through it.
A branch office that loses WAN keeps serving its own users; reconciles on reconnect.

Extension marketplace — everything configurable in the admin GUI.

Available today

Every extension now exposes its configuration in the admin GUI. No editing files on the server, no command-line for the routine work. Multi-device plugins (network gear, server fleets) get a dedicated device-fleet UI. SNMP plugins get a MIBs-import dialog. The marketplace itself filters by tier, category, protocol and auth — find the right plugin without reading the catalog cover to cover.

What changed for you

Configure plugins from the admin shell — schemas drive the form, validation is server-side.
Manage a fleet of devices (switches, firewalls, servers) from one panel per plugin.
Import vendor MIBs through the GUI for SNMP-based monitoring packs.

Networking plugin catalog — sixteen plugins shipped, more in flight.

Sixteen today · more in flight

The networking plugin program shipped its first wave: integrations for Cisco Meraki, Fortinet FortiManager, Aruba Central, Palo Alto Panorama, F5 BIG-IQ, MikroTik RouterOS, Arista eAPI, Dell iDRAC Redfish, Dell OpenManage Enterprise, A10 ACOS, Barracuda CGF, Citrix NetScaler, Check Point SMS, Extreme Cloud IQ, H3C iMC, and Huawei iMaster NCE — plus two SNMP category drivers covering UPS power gear (APC, Eaton, Vertiv, Socomec, Tripp Lite, CyberPower) and printer fleets across vendors. Eldric supports your existing hardware fleet; no rip-and-replace.

What changed for you

Monitor and configure your existing network from the same chat shell you use for everything else.
UPS power gear and printer fleets get plugged in without a per-vendor plugin per device.
HPE and Eviden flagship sub-tracks are in flight — strategic for datacenter and HPC customers.

Full catalog: networking plugins.

Three deployment postures — pick the shape that fits.

Available today

Eldric now ships with three explicit customer postures, each with its own page and positioning. Enterprise — you own a building, you run your own AI. Datacenter — you operate the iron for everyone else, AI becomes the next tenant. Service provider — you sell platform to other companies, multi-tenant isolation is the product. Same binary, three honest framings.

What changed for you

For enterprises — Eldric as your private AI platform.
For datacenters — Eldric as your next datacenter workload.
For service providers — Eldric as your customer platform.

Eldric AI Datacenter — the complete operating system framing.

Page in flight

Sitting above the three postures, we now describe Eldric as the complete AI datacenter operating system. Hardware-agnostic: HPE, Eviden BullSequana, NVIDIA, Dell, or a mix. One platform for inference, retrieval, agents, training, telemetry and tenancy — not a stack you have to assemble. The flagship page captures the OS-as-a-whole claim that the posture pages segment.

What changed for you

One operating-system framing for the whole datacenter, vendor-neutral.
Procurement, deployment, day-2 operations all described as one product, not a stack.
Flagship page lands alongside the other catch-up items this wave — see Eldric AI Datacenter.

Better search across your knowledge bases.

Available today

RAG search now combines lexical and semantic scoring in a single ranked result set. Proper names, product names, IDs, and entity disambiguation work the way you'd expect — search for "Hochreiter 1997" and you get the right paper, not whatever was semantically nearest. Search for a specific customer or invoice number and the exact match wins.

What changed for you

Proper-name and ID searches return what you asked for.
No tuning required — the fusion runs by default on every knowledge-base query.
Semantic recall is preserved; exact-match disambiguation is now layered on top.

Eldric in your inbox — per-protocol identity.

Phase 1 available today

Eldric can now have its own identity inside the messaging channels you already use. Give it an email address on your domain, an SMS number, a chat handle — and customers, colleagues or systems can Cc it into a thread. Eldric responds in-thread under that identity, with the same authentication and tenant scoping as the chat shell. Phase 1 covers email and SMS; more protocols follow.

What changed for you

Cc: eldric@yourdomain.com on any thread — Eldric picks it up and answers.
One per-protocol identity per tenant; isolation enforced server-side.
Useful for support triage, internal Q&A, and any workflow that already lives in email or SMS. See Eldric identity.

Native tool calling in the webchat.

Available today

Chat-side tools — file operations, search, web fetch, the rest of the catalog — now route through each model's native tool-calling API where the backend supports it. The XML-format fallback stays in place for models that don't, so nothing breaks. The result for customers is fewer parsing edge cases, fewer half-finished tool runs, and noticeably more reliable agent behaviour on long tool chains.

What changed for you

Tool calls land more reliably, especially on multi-step agent runs.
Same chat shell, same tool catalog — the protocol underneath is just cleaner now.
XML fallback covers older or smaller models without a behaviour change.

Latest shipping items

Closing the gap to 5.0.

v5.0.1 shipped — 2026-05-25. macOS Sparkle auto-update lands existing v5.0.0 installs (and any prior alpha builds) on v5.0.1 on the next poll cycle. Linux customers run sudo dnf update eldric-aios. See release notes.
Sixteen plugins now load. Eight Bosch sensor plugins plus seven networking integrations (Cisco, Extreme, FortiGate, FRITZ!Box, Redfish, SNMP, UniFi) were rejected by an over-strict manifest check in 5.0.0; the check now matches what the catalog ships.
Seventeen more science sources queryable. Semantic Scholar, Crossref open-access papers, OpenAlex and fourteen others come online in the Source Registry. See Admin → Science in the cluster dashboard for the live list.
🎉 Eldric 5.0 is generally available. Released 2026-05-24 as v5.0.0 — multi-tenant, multi-worker, multi-distro. See the release notes for the full feature list and what's next for the roadmap from here.
Smart memory inference (preview). The inference daemon now consults the matrix-memory layer directly at prompt boundaries, grounding model answers in your own data without a separate retrieval call. Sub-2 ms per-token overhead on CPU. Pro+ opt-in. See smart memory inference.
Eldric for iPad — Phase C live. Floating composer, NavigationSplitView, Apple Pencil + Scribble, drag-drop ingest from Files / Photos / Mail. Multi-window in preview. TestFlight today; App Store next. See iPad page.
Theme cluster sync. Per-user theme settings sync across all clients (Mac GUI, iOS, iPad, web) via the controller. Settings follow the user, not the device.
Notification center. Long-running agent tasks now surface as in-app notifications with progress + one-tap return. Apple Push wired through to lock-screen alerts on iOS / iPad.
Mobile rating dashboard. Admin-gated view of customer-feedback ratings on mobile. Reads from the platform feedback endpoint.
Agent Builder UI. The "AI that builds AI" flow now has a chat-shell admin interface — describe the agent you want, the platform generates code, tool definitions, prompts and test cases. Sandboxed test runs before deployment.
Smoke pass rate at GA quality. A registry fix on the data worker collapsed six interrelated red checks into one ship. The 106-check end-to-end suite now sits comfortably above the GA threshold (≥ 99/106 green) on the demo cluster. See road to 5.0 GA.
Industrial transports — MQTT publish GA. The publish half of MQTT-Sparkplug-B flipped from preview to GA alongside the existing Modbus + WebSocket transports. OPC-UA ships at preview in 5.0 GA — the source is in the release, the cluster RPM bundles the library at the next cut. MQTT-subscribe stays preview for the 5.0 line. See xLSTM & IoT.
ARM64 minimal RPM live. The minimal edge runtime is now signed-published as aarch64 alongside x86-64. The one-line installer detects the host. See edge install.
macOS auto-update live. The macOS GUI carries a cryptographically-signed updater (Sparkle-based) pulling from the package server. Stable and beta channels.
ENRN v18 distilled router (preview). Single-pass classifier trained on cluster routing data; lower-latency routing decisions; LLM stays in the loop for ambiguous queries. Pro+ opt-in.
EMM compressed retrieval (preview). EMM Compressed variant of the matrix-memory layer — smaller .emm files, faster at high concurrency. Per-knowledge-base opt-in.
Cross-distro DEB packaging. Ubuntu 24.04 + Debian 12 DEBs follow shortly after GA; today the WSL2 / container path is the recommended route on those distros.
Model badges in chat. Every model in the picker carries a small coloured backend badge — Ollama, OpenAI, Eldric Inferenced, vLLM, llama.cpp and the others — and the same badge appears under every assistant message. Green for models on your own cluster, brand-coloured for external APIs. No silent re-routing between local and external. See what the badges mean.
Tool query info in chat. Toggle Show tool query details in chat settings. Every RAG, science, web and comm tool call now shows a one-line summary of what was actually queried before the result. Default on. See help index → Chat shell.
Agent Builder Share, Publish + Marketplace. Agents you build in chat now have one-click share inside your tenant, publish-to-marketplace for the cluster admin to review, and an in-place editor with versioning. Sandboxed runs validate every change before deploy.
iPad Phase D — multi-window, keyboard shortcuts, cross-window drag. Two Eldric windows side-by-side via Stage Manager and Split View. Keyboard shortcuts for the common chat actions. Drag a message from one window to another to fork a conversation. See iPad page.
macOS notification center. Long-running tasks now surface as native UNUserNotification entries on the Mac, with APNs wired through for remote pushes from cluster events.
xLSTM workload control plane. Admin UI for the four xLSTM workload classes (policy · forecast · encode · retrieve) with per-tenant enable, per-model selector, and live load chart.
EMM Compressed migration CLI + per-namespace toggle. Existing matrix-memory files can be migrated to the EMM Compressed variant on a per-namespace basis, with a single command. The chat shell shows migration progress.

The two preview features are written up customer-facing on advanced retrieval. The release-notes finalisation is on release notes.

Edge runtime & packaging

The minimal node.

A lighter-weight build of the platform now ships alongside the full distribution. eldric-aios-minimal includes only the kernel, controller, data, edge and knowledge modules — about a quarter the disk and memory footprint of the full release. Intended for branch offices, kiosk boxes, Raspberry Pi-class hardware, and small VPS instances that proxy chat traffic to a larger cluster.

What changes for you: on small edge hosts, install eldric-aios-minimal instead of eldric-aios. Science, training, NOVA, swarm, comm, media and IIoT modules stay on the data-centre nodes; the edge proxies up to them. Documentation: features § Edge.

xLSTM Edge gateway.

The Edge server now proxies xLSTM workload requests (policy / forecast / encode / retrieve) to the appropriate eldric-xlstmd worker. Path translation handles /v1/eldric/<workload>/... → /api/v1/<workload>/..., capability headers thread through from the cluster secret, and the worker registry refreshes every 30 seconds from the controller. Least-connections selection across multiple xLSTM workers. WebSocket upgrades pass through unmodified.

ARM cross-compile.

Three new Dockerfiles for ARM targets shipped this wave: Raspberry Pi 4, generic ARM64 server, and NVIDIA Jetson. deploy-rpm-fanout.sh classifies aarch64 from the filename so heterogeneous clusters install the right architecture automatically.

Capability-gated APIs

The new authorisation surface.

Every kernel endpoint is now gated by an explicit capability rather than a role string. The capability set is fixed in the binary; tenant admins compose customer-named roles from those capabilities (Pro+ feature, ships in an upcoming 5.0.x patch). On 5.0, the existing four roles (Viewer / Developer / Admin / SuperAdmin) map onto the capability set under the hood. No flow changes for end users.

What this means for integrations: 403 responses now carry a required_capability field in the JSON body that tells you what was missing. Your client should surface that field; the existing chat-shell does. Webhook-subscriber and plugin authors should respect the same field on inbound responses. Reference: API reference — Authentication.

Chat & webchat

First-run admin signup.

On a fresh install, the very first signup becomes the system administrator automatically. No shipped default password. The chat shell shows a first-run card with email + password + an "I will be the administrator" confirmation. Subsequent signups land as regular users in the default tenant. Public-host warning preserved: on internet-reachable clusters, disable open signup from the Admin Console immediately after claiming the admin role.

PKI & ACME endpoints surfaced.

Internal-CA management and Let's-Encrypt ACME flows are now wired through the chat shell. Generate certificates, deploy them, rotate them, request and renew Let's-Encrypt certificates — all from the Admin Console. Documentation: the features documentation.

Federated learning & dormancy.

The federated-learning round orchestration ships in the controller-broadcast + worker-train + shard-aggregate pattern. Cluster dormancy (the "idle for N minutes" tier downgrade that frees worker capacity for batch jobs) is wired through the controller. Both are off by default; admins opt in per tenant. Tracking notes on the features catalogue.

Clients

iOS slash-autocomplete + Experiment Designer.

The iOS app now exposes a bottom-sheet slash-command autocomplete that mirrors the macOS GUI's slash-menu behaviour. Type / and the available commands appear; tap to insert. The Experiment Designer renders an inline status card for science experiments — phases, sub-tasks, output links — and accepts both stub {type,id} arrivals and full payloads without an async fetch round-trip.

Project switcher + KB scope.

Sessions now carry a projectId; the sidebar filters by the current UserContext project. Knowledge bases gained a projectId selector filter and scope-parse logic — the chat header KB selector lists only KBs scoped to the active project.

Sparkle auto-update.

The macOS GUI uses Sparkle for in-app updates. The appcast lives at https://repo.eldric.ai/appcasts/macos-gui.xml. Update channels are stable (default) and beta (opt-in via the GUI's settings).

Operations

Model registry.

Pre-trained ENRN classifiers and EMM matrix-memory seeds now ship through the package server alongside the platform RPM, DEB and PKG. Pin to a specific version (enrn-intent-v1), or follow the latest within a family. The installer handles the download and checksum verification.

Stress scaffold.

A stress-test harness ships with 5.0 — parallel-user × request-count load runs with pass/fail thresholds against p99 latency and error budget. Verified on our demo cluster as part of the GA cut.

Smoke audit baseline.

The 5.0 GA smoke suite runs 106 checks against a clean cluster. The current baseline holds at 93 green — the open items pair with the WIP markers in the features catalogue (xLSTM daemon cold-start, parallel knowledge-base search at concurrency 4); none block GA.

Coming in 5.0.x

Portable cluster bundles.

A new .nexus bundle format ships in 5.0 as admin tooling. Pack the matrix memory, vector documents, knowledge-base sources and tenant configuration of one Eldric installation into a single signed file; ship it; unpack on the other side with a clean merge. Useful for tenant portability, project handoff between teams, seeding an edge install with a tested baseline, lightweight federation between regional clusters, and disaster recovery. The bundle is scope-filtered, so a project-level export never carries other tenants' data with it.

Customer-named roles.

Today on 5.0, four built-in roles (Viewer / Developer / Admin / SuperAdmin) gate the platform. In an upcoming 5.0.x patch, tenant admins compose their own roles from the underlying permission set — name them anything ("Auditor", "Lab tech", "Finance"), scope them to the tenant, grant them to users. On migration, every existing superadmin keeps full access via a default per-tenant "Admin" role; everyone else starts with the explicit grants you give them.

5.0.x scope clarity.

What's shipped in 5.0 GA, what's queued for upcoming 5.0.x patches, and what's explicitly out of scope are catalogued together. The features catalogue carries the same picture in long form, with WIP markers on the items that ship in an upcoming 5.0.x patch.

Next.

For the long-form catalogue of everything that's in 5.0, read features. To install today, start at get started. To apply the platform to your own domain, read apply Eldric to your domain.