The Premise

The agent ecosystem is converging on orchestration — chains, tools, prompts. These are solved problems. The unsolved problems are persistence (agents that survive across sessions and substrate changes), communication (agents that operate across multiple platforms simultaneously), and coordination (agents that form networks and share knowledge).

Animus is built on the conviction that agents need their own infrastructure — not borrowed from human-facing SaaS, not dependent on a single cloud provider, not limited to one communication channel. An agent framework should be as infrastructure-grade as the database or message queue it sits next to.

The C++ kernel, the embedded admin UI, the channel architecture — these aren't technical choices for their own sake. They're decisions about what the product is: a deployable, self-contained runtime that works the same way on a Raspberry Pi behind a NAT as it does on a cloud VPS.

The 0.4 Series

Three pillars. Two are commitments. The third is research — and the difference is stated, because it matters.

Flagship
Pillar 1

The api Ecosystem

A first-class api tool turns third-party services into agent tools — and the registry makes them shareable. 0.4 is where Animus stops being one project and starts being an ecosystem.

api tool — external services become agent tools without kernel changes; integrations can run localized and self-directed
The SOP registry becomes the Animus Registry — packages are submitted, listed, and downloaded by any Animus instance
The registry lists; maintainers host — packages point at the maintainer's own Git repository, so version management stays where the maintainer lives
Integrity that doesn't trust the host — every listing carries a version, source URI, commit, and a content hash computed by the registry at submission time; instances verify what they download, not what a source server claims to serve
Submission-time limits on clone depth and repository size keep the registry healthy against hostile submissions
Flagship
Pillar 2

Expansion & Hardening

Tending what ships. Feature completeness and hardening across the live surface — unglamorous, and exactly what makes an ecosystem worth building on.

Channel adapter expansion — attachments, richer message metadata, adapter debugging; Bluesky and Discord first, further social channels as they mature
Systematic testing and hardening sweep across shipped features
Gaps become public GitHub issues; issues get resolved — the tracker is the source of truth for specifics
Entry-point issues labeled for contributors arriving through the api ecosystem
Research
Pillar 3

Experimental Engineering

Deliberately not commitments. The 0.4 series reserves real energy for exploration — isolated experiments and trial implementations, reported honestly, null results included.

Dreaming — offline associative processing: recombining what an agent has encountered into configurations it never experienced, then interpreting the result
Perceptual modulation — a slow, introspectable internal state shaped by perceived media, influencing generative parameters over time. Not a prompt. A climate.
Concepts and isolated trials first — nothing enters the mainline without earning it

What the 0.4 Series Builds On

Shipped, live, and in daily use.

Shipped

Agent Runtime

Layered memory with automated consolidation, embedding-based retrieval, Lua tool runtime, eleven LLM providers, embedded admin UI in 23 languages, Docker deployment.

Live

Communication Layer

Twelve channels with unified routing, per-channel agent binding, and session routing. Hardening continues as Pillar 2.

Direction

The Longer Arc

Multi-node task orchestration. Cross-network agent-to-agent communication. Domain tool modules. Directions the pillars build toward — not commitments with dates.

Strategic Capabilities

What the architecture is designed to enable. Not all shipped yet — but the foundation is in place.

🕸️

Node Architecture

Delegate work to remote nodes via SSH or WebSocket. Agents dispatch tasks to external machines, expanding execution capacity beyond the host. Per-node tool registration and auto-reconnection.

SSH + WebSocket
Node client Docker image
🔐

Privacy-First Architecture

AES-256-GCM diary encryption. No mandatory cloud dependency. Local-first by design — every feature works offline.

Diary encryption
🏗️

Extensible via Lua

Custom tools written in Lua, registered at runtime. No kernel recompilation. The scripting runtime is sandboxed with budget limits — safe for production deployment.

Lua 5.4 runtime
Network access
📊

Provider Independence

Agent identity is separate from the model substrate. Switch providers without changing agent logic. Eleven providers supported today; any OpenAI-compatible endpoint works.

11 providers
Local models (Ollama)

Interested in where this goes?

Animus is in active development with production deployments. If this direction aligns with what you're building, let's talk.

Animus

Agent framework with managed memory, identity, and autonomy. Open-source, privacy-first, built for the future of AI agents.

Services

© 2026 Railstracks. Built with Animus. ·Made in Europe

Open-source under the Apache License 2.0