﹏ hum ﹏
The only AI stack nestled on a biodiverse agentic kernel framework.

curl -fsSL https://raw.githubusercontent.com/adiled/hum/main/install | bash

hum (n.) The phenomena occuring upon perfect harmonization of all players in an AI composition.

Key buildings blocks

it is all composed of biodiverse primitives, not an outcome of conventional linguistics, following are some nouns

• hive the kind, a typology a bee conforms to
• bee the instance, a running participant a hive commissions
• nestler a bee mid-handshake, awaiting the breath that accepts its hello
• nestled a nestler once registered into the nest
• nest where nestled bees gather, inside humd
• thrum the hum-native vibration protocol, carrying tones across a range of chi
• petal one unit of content, be it text, image, a tool call or its result
• bloom one turn of conversation, opened by a prompt and closed when it wilts
• ensemble the mesh of cooperating humds, where inference, compute, filesystem, dependencies and UX decouple and gossip back into the hum

Ensemble: Bee Gossip

An ensemble is a mesh of cooperating humds, where bees gossip like a hive trading news of distant blossoms. A single bloom never lives on one box. It draws inference from one humd, compute from another, a filesystem from a third, and surfaces in a UX somewhere else again. No bee is bound to its own field; it forages for miles, following the scent of the universe for any model, tool, payment rail, or dataset no flower nearby could give. No monolith, no central broker, each capability sovereign on the host that grew it.

This is the whole of it, a decoupling so complete that the scattered many become one again, a single hum carrying across the long dark between distant stars. You raise a capability wherever it belongs, let it gossip, and the swarm folds it into the hum. Nothing needs repointing, nothing needs redeploying. Every new bee only deepens the hum.

Learn more by simply humming along.. or read the scenarios, one story per ensemble narrative.

Config ~/.config/hum/hum.json

Refer to hum.schema.json

Comparison for the sake of comparison

hum is an agentic kernel, not an LLM router, and incomparable in scope. This table simply highlights how hum's out-of-the-box hives overlap with the winning representative from each class.

Capability	hum	LiteLLM	Ollama	LocalAI	CAMEL	Azure Agent Framework	Phoenix	vLLM	Open WebUI
Class	Agentic kernel	Router / proxy	Local model server	Multi-modal AI server	Multi-agent framework	Multi-agent orchestrator	AI observability & evals	Production inference serving	LLM web UI
Language	Rust	Python	Go	Go	Python	Python / .NET	Python	Python	Python
Model gateway	✅ (hives gossip)	✅ 100+ APIs	✅ Local models	❌	❌	❌	❌	❌	❌
OpenAI-compatible API	✅ (openai-server hive)	✅	✅	✅	❌	❌	❌	✅	✅
Ollama-compatible API	⚠️ (in progress, see issue #44)	❌	✅	❌	❌	❌	❌	❌	✅
Multimodal: vision, TTS, STT, image gen	❌ (honeybee or external hive)	❌	❌	✅ LLM, vision, voice, image, video	❌	❌	❌	❌	✅
Built-in TUI	✅ (decoupled, wryme in progress)	❌	❌	❌	❌	❌	❌	❌	✅
Self-hosted, local-first	✅ (kernel runs anywhere)	✅	✅	✅	❌ (cloud)	✅	✅	✅	✅
Multi-agent orchestration	✅ (bee/nest primitives)	❌	❌	❌	✅ agents + roles	✅ workflows + agents	❌	❌	❌
Agent memory / long-term context	✅ (bloom + thrum protocol)	❌	❌	❌	✅ (societal memory)	✅ (cognitive memory)	❌	❌	❌
Streaming / SSE	✅ (thrum tones, petal primitives)	✅	✅	✅	✅	✅	❌	✅	✅
Agent tracing / observability	✅ (humd runtime, thrum transport)	✅ (cost, usage, logs)	❌	❌	✅ (swarm traces)	✅ (run traces, deploy)	✅ traces, evals, prompt mgmt	❌	❌
Tool / MCP gateway	✅ (MCP serde + hive)	✅	✅	✅	✅	✅	✅	❌	✅
RAG / retrieval pipeline	❌ (external hive)	✅ (via adapters)	✅	✅	❌	✅ (Azure AI)	✅ vector DBs	❌	✅
Guardrails / safety	❌ (bee responsibility)	✅	❌	❌	❌	✅	✅	❌	❌
Trust / Risk modeling	✅ (strict_auth + T1→T4 trust via Transport + Ensemble hello handshake)	❌	❌	❌	❌	❌	❌	❌	❌
Roaming inference	✅ (Chi::Attach + wane cursor replay)	❌	❌	❌	❌	❌	❌	❌	❌
Roaming filesystems	✅ (humfs forager hive + hum:// URI routing)	❌	❌	❌	❌	❌	❌	❌	❌
Multi-party gamification	✅ (multiple nestlers per bloom via observers[sid] roster)	❌	❌	❌	❌	❌	❌	❌	❌
Distributed orchestration	✅ (Ensemble gossip + Kad + capacity-aware overflow routing)	❌ (static routes)	❌	❌	❌	❌	❌	❌	❌
Zero-trust access control	✅ (scoped capability grants + revocation via federation.* error chi)	❌	❌	❌	❌	❌	❌	❌	❌
Agent-human consensus	✅ (thehum hash chain + on-chain anchoring via HumdRegistry contract)	❌	❌	❌	❌	❌	❌	❌	❌

Bottom line: routers proxy; hum composes. Each hum hive is a sovereign capability, inference, tool use, payment, UX, that gossips into the ensemble. An LLM gateway routes requests; hum routes biodiverse primitives into a harmonized flow.