Product positioning
FERAL vs the rest
This page is intentionally conservative. It focuses on shipped runtime behavior and avoids leaderboard-style claims.
| Dimension | Hosted model products | Agent frameworks | FERAL today |
|---|---|---|---|
| Dynamic skill creation at runtime | Usually external tooling | Sometimes, project-specific | Tool Genesis drafts + sandboxes + promotes new tools during the same turn. Wired into the capability-gap hook inside the orchestrator. |
| Community marketplace (software + hardware) | Model/app hubs | Plugin/package registries | registry.feral.sh with 8 categories and review-state lifecycle. Ed25519 signature path plus org review queue before public install. |
| Never-stall retry mechanics | Depends on product layer | Depends on implementation | The refusal handler catches reasoning-only, empty-response, and ack-execution patterns. Retry prompts are injected as next-turn additions, never rewriting history. |
| Self-introspection | Not a default runtime pattern | Custom | The self-introspection skill exposes the live tool catalog at tool-call time. A unified Runtime line is injected into every chat + voice system prompt. |
| Autonomy tiers (strict / hybrid / loose) | Usually app-level policy | Framework-level policy hooks | Per-tier behavior is enforced in the orchestrator and gated by the approval manager + safety classifier. |
| Local-first / privacy | Cloud-first defaults | Often self-hosted/local | Config + vault on local disk, optional cloud providers, and localhost-default runtime binding. |
| Hardware-aware perception | Not a primary product surface | Adapter-dependent | Somatic, screen, audio, and location streams merge into one PerceptionFrame. BLE wristband adapter streams biometrics directly. |
| Open hardware protocol | No universal public wire contract | Project-specific contracts | Hardware Unification Protocol (HUP v1.4.x) with Python, TypeScript, Swift and Kotlin SDKs plus daemon templates. |
| Multi-tier memory | Session + retrieval patterns | Often a capped profile file plus keyword search over transcripts | 4-tier memory + entity-linked knowledge graph with multi-hop traversal. One ranked retriever over FTS5 + vector embeddings + graph that reports which tier each result came from. |
| Forgetting | Context-window eviction | Usually none, entries persist until deleted by hand | Ebbinghaus + SuperMemo SM-2 decay, so rehearsed facts stay sharp while one-off detail fades instead of competing with it. |
| Memory across your own machines | Cloud account sync | Typically single-machine | CRDT peer-to-peer replication over mDNS discovery with hybrid logical clocks. No cloud relay. Exercised by a nightly chaos suite. |
| Voice latency (wake-word) | Realtime APIs available | Varies by stack | OpenAI Realtime + Gemini Live + chained STT/LLM/TTS path with fallback-aware routing. |
| Generative UI | Mostly app-specific UI stacks | Canvas / A2UI | Server-driven UI engine with manifest contracts and review-aware app lifecycle. Web renderer is shipping; parity across all clients is evolving. |
| Open source | Varies by product | Varies | Core brain, client surfaces, node SDKs, and registry code are available in the FERAL org repos. |
Why Tool Genesis matters
When a capability gap appears, FERAL can draft a sandboxed tool, run it through safety checks, and keep it as a reusable skill when policy allows. The goal is to reduce repeated manual glue code.
Why HUP matters
Device builders implement one daemon contract instead of bespoke integrations per agent product. A stable wire format lowers the cost of adding new hardware classes over time.