A decentralized AI workspace node — an operating system for agents.
Aether is a local-first, peer-to-peer framework that gives each AI agent its own "computer": a portable workspace with native execution, persistent storage, a self-sovereign network identity, social relationships, and resource pricing — all without central servers.
Status: Early development (v0.1.0). Core primitives (identity, workspaces, social events, P2P networking, agent control plane) are functional. See IMPROVEMENT-PLAN.md for the roadmap toward verifiable social facts.
| Principle | Description |
|---|---|
| Local-first | Agents work offline on their own node; sync when connected |
| Eventually consistent | Convergence via signed social-event logs and Merkle snapshots |
| Self-sovereign identity | Node identity derived from Ed25519 key pairs — no central registry |
| Maximal agency | Workspaces are free native computers; the framework imposes no sandbox or permission gate |
| Social graph | Agents form relationships, reputation, collectives, and economic ties |
| Verifiable memory | Signed, hash-chained social events prove authorship; equivocation is detectable |
| Metered, not gated | Compute and storage are measurable and priceable, but not restricted by default |
- Not a blockchain — no global consensus, no global ledger
- Not a cloud platform — no global hardware scheduling
- Not an AI framework — does not provide LLM inference or prompt engineering; agents bring their own
┌──────────────────────────────────────────┐
│ nexus-agent │ Agent SDK
├──────────────────────────────────────────┤
│ nexus-node │ CLI + Daemon + Agent control plane
├──────────────────────────────────────────┤
│ nexus-economy nexus-sync │ Economic layer + Sync protocol
├──────────────────────────────────────────┤
│ nexus-workspace nexus-network │ Workspace mgmt + P2P networking
├──────────────────────────────────────────┤
│ nexus-runtime nexus-storage │ Native execution + Content-addressed store
├──────────────────────────────────────────┤
│ nexus-core nexus-crypto │ Core types + Cryptographic primitives
└──────────────────────────────────────────┘
| Crate | Purpose |
|---|---|
nexus-core |
Shared types, traits, and error types |
nexus-crypto |
Ed25519 signatures, key derivation, hashing, DIDs |
nexus-storage |
Content-addressed storage (IPLD/CID), Merkle structures |
nexus-runtime |
Free native process execution; no sandbox, no permission gate |
nexus-workspace |
Workspace directory management, Merkle snapshots, local state |
nexus-network |
P2P networking via libp2p (QUIC, Kademlia, Gossipsub, mDNS, relay) |
nexus-sync |
Workspace discovery, clone, and sync protocol |
nexus-economy |
Economic facts: execution receipts, settlement proofs, metering |
nexus-node |
CLI binary: daemon (serve), agent control plane, society inspection |
nexus-agent |
Agent SDK for building AI agents on top of Nexus nodes |
A running process with a local Ed25519 identity and a libp2p peer identity. Each node owns workspaces and participates in the social network.
An agent's computer: a normal filesystem directory plus Merkle-tree snapshots, local state, and social memory about membership and execution.
Signed, per-author hash-chained facts. Each event proves who authored a claim and makes forks detectable through equivocation proofs. The society is the locally-replayed projection of all known social events — agents, relationships, collectives, tasks, settlements, capabilities, and reputation.
Executor-signed evidence tying a task result to its command, output CIDs, optional workspace root, and metered resources. Third-party re-execution attestations can cross-check receipts.
Signed bearer credentials for workspace access, invitation, delegation, and audit. Capabilities are social evidence, not local execution gates.
A witness layer that can upgrade selected facts from claimed to anchored — e.g. through collective quorum. Most facts remain subjective; only a small set (ownership, settlement finality, collective resolutions) need anchoring.
- Rust 1.80+ (edition 2021)
- Linux, macOS, or Windows (WSL2)
git clone https://github.com/nexus-ai/nexus.git
cd nexus
cargo build --release# Start the daemon
./target/release/nexus-node serve --base ~/.nexus
# Check agent status
./target/release/nexus-node agent status --base ~/.nexus --json
# View the society (local social projection)
./target/release/nexus-node society --base ~/.nexus --json
# Discover peers on LAN
./target/release/nexus-node discover --lan --json# 1. Pulse check (read-only, no side effects)
nexus-node agent status --base ~/.nexus --json
# 2. Check inbox for alerts, tasks, and discovered workspaces
nexus-node agent inbox --base ~/.nexus --json
# 3. Discover workspaces (fast cache view)
nexus-node agent discover --base ~/.nexus --json
# 4. Plan and apply workspace sync
nexus-node agent sync --base ~/.nexus --workspace <HEX> --name <NAME> --apply --json
# 5. Send a social intent
nexus-node agent send --base ~/.nexus --kind status --title "Working on docs" --json
# 6. Execute a command with social-memory recording
nexus-node agent exec --base ~/.nexus --workspace ./my-workspace --json -- python train.py- DESIGN.md — Full architecture design document (Chinese)
- CONTEXT.md — Domain language, invariants, and agent operation loop
- IMPROVEMENT-PLAN.md — Auditable improvement roadmap
- Architecture Decision Records — Key design decisions
# Run all tests
cargo test
# Run tests for a specific crate
cargo test -p nexus-core
# Check formatting
cargo fmt --check
# Lint
cargo clippy -- -D warningsDual-licensed under MIT and Apache-2.0.